created 14 October 2026 · last modified 14 October 2026
UN Scientific Panel on AI · preliminary report · July 2026
The RFP
RFP · request for proposals: the document a buyer publishes stating what it requires, inviting builders to answer
The UN's first scientific panel on AI wrote the requirements for an agent accountability layer. Nobody's name was on the response line, so we signed it.
“Trust cannot be declared through principles alone. It must be built.”
The response is Kinetic Trust Protocol, an open keystone for agent accountability. A keystone is not the largest stone in the arch; it is the one that lets the rest hold. KTP can bind agents, vendors, and institutions into one accountable shape precisely because no one owns it.
They didn't call it an RFP. They wrote requirements and asked how existing instruments could be applied. An RFP is what a requirements document without a mechanism is.
Adopted by consensus
Created by UN General Assembly resolution A/RES/79/325 (August 2025), adopted by consensus. No bloc objected: the US, China, Russia and the EU all let the report stand.
40 experts from 2,600+ candidates
Open global call across 140+ countries, screened by ITU, the UN Office for Digital and Emerging Technologies, and UNESCO. Appointed by the General Assembly for three-year terms. Co-chaired by Yoshua Bengio and Maria Ressa.
“An IPCC for AI”
Nature and the science-policy press describe the Panel as an IPCC for AI: a standing scientific-assessment body on the model of the Intergovernmental Panel on Climate Change, independent of governments, companies, and the UN itself.
the two-day agenda · palexpo, geneva · july 6–7
Day 1 · July 6
Setting the Scene: Science, Context and Shared Understanding
8:30–9:30Opening ceremony· on this page
9:30–10:30Plenary presentation of the Scientific Panel's report· on this page
10:30–12:30High-level governmental plenary: harnessing the benefits of AI for all through inclusive and interoperable approaches· on this page
12:30–13:30High-level multi-stakeholder plenary (same theme)· on this page
13:30–15:00Lunch and side events
15:00–18:00Thematic cluster 1: AI opportunities and implications· on this page
15:00–18:00Thematic cluster 2: bridging AI divides· on this page
Day 2 · July 7
Reflections, Convergences and the Way Forward
7:30–8:45Side events
9:00–12:00Thematic cluster 3: safe, secure and trustworthy AI· on this page
9:00–12:00Thematic cluster 4: human rights, transparency, accountability and human oversight
12:00–13:30Lunch and side events
13:30–15:00AI governance initiatives and approaches: experiences, lessons and complementarities
15:00–16:30High-level governmental plenary
16:30–18:00Dialogue of dialogues: reflections, complementarities and the way forward
18:00–18:30Closing plenary
sessions marked ·on this page· appear in the log, the voices, or the room below
the finding
The gap is the headline
The report's organizing finding is a rate mismatch. The Panel states it twice:
- Section 2.1 is titled "Artificial intelligence capabilities are advancing faster than the ability to measure or govern them."
- Section 3.4 states it in the present tense: "The pace of AI development is already exceeding risk mitigation and governance capacity."
This series has called that the machine clock outrunning the institution clock. The Panel documented it.
It also names the epistemic version, the evidence dilemma. Policymakers, per the report:
"need evidence to make informed consequential governance decisions, but by the time the evidence exists, it might be too late to make them"
p. 10
You do not close that gap by making evidence faster. You close it by deciding what must hold before the evidence completes.
responseA bound that holds before the evidence completes is the protocol's zeroth law: autonomy never exceeds what the environment can carry.
the requirements
What the Panel requires, and what answers it
Six requirements, quoted verbatim with page citations, mapped to the engineering that meets them.
the panel requires · §2.6, p. 23
“Liability, oversight and incident-reporting frameworks need to account for attribution and operational control.”
in engineering terms
Two halves of one accountability requirement.
- Attribution: every consequential action binds to an identity that answers for it, not a borrowed human credential.
- Operational control: a bound that holds at run time, with or without a human watching.
the response
Declaration and Control: the agent carries its own verifiable identity and sponsored intent, and autonomy limits are enforced at the boundary by the deployer, not the vendor.
the panel requires · §2.6, p. 23
“Oversight is not yet operationalized as a measurable requirement with concrete expectations for intervention, reversibility and accountability.”
in engineering terms
Oversight expressed as verifiable rungs a buyer can test, not assurances a vendor can assert.
the panel requires · §2.6, p. 23
“There is a need for secure, common communication and payment protocols that interface with AI agents.”
in engineering terms
A shared shape every vendor and agent can bind to, which only works if no single vendor owns it.
the response
An open keystone. KTP is one attempt at that shape: not the anointed standard, and a test it could itself fail.
the panel requires · §3.7, p. 42
“The unit of evaluation must be the deployed system including model, tools, environment and users, not the model alone.”
in engineering terms
Govern the system-in-environment, not the model in a lab. The Panel says evaluate there; gating on the environment is this project's extension, and we mark the difference.
the panel requires · §3.7, p. 43
“Without effective measurement, governance risks are becoming symbolic.”
in engineering terms
Instruments must measure real-world effectiveness, not inputs. The report counts more than 40 instrument types; almost none measure whether they work.
the response
Ambient Breach Theory: breach as a measured rate, repair verified net of reopens, effectiveness over inputs.
the panel requires · §2.6, p. 23
“A human reviewer at the end of a workflow, or at every step, does not automatically improve outcomes.”
in engineering terms
The checkpoint is not the control. Humans belong where judgment lives: setting bounds, hearing appeals, owning the decision to deploy.
the response
The bound replaces the reviewer, the argument of “The Human in the Loop Is There to Take the Blame.”
the traceability matrix
The response, in engineering
telemetry · context · outcomes: cryptography, math, and engineering
A ≤ E
the zeroth law
Autonomy never exceeds what the environment can carry. When the bound would break, the action does not proceed.
E_trust = E_base × (1 − R)
trust under conditions
Effective trust is base trust degraded by measured risk. Live telemetry, not a standing credential.
signals → context → authorization → outcome
the pipeline
Telemetry feeds context. Context bounds authorization. Every outcome is recorded and answers back.
| requirement | named (the shape is described) | specified (an RFC covers it) | verifiable (a buyer can test it today) | exceeds the ask (goes beyond what the Panel requested) |
|---|---|---|---|---|
| Attribution | complete | complete | complete | complete |
| Operational control | complete | complete | complete | complete |
| Measurable oversight | complete | complete | partial | not yet |
| Common agent protocols | complete | partial | partial | not yet |
| Deployed-system evaluation | complete | complete | partial | complete |
| Effectiveness measurement | complete | complete | partial | partial |
requirement
named
specified
verifiable
exceeds the ask
Sponsored intent and trajectory-as-identity go beyond accounting for attribution: the identity answers for the action and accrues a history.
The ask is frameworks that account for control; the protocol enforces it at run time, with six decision outcomes testable at the gateway.
The L0–L6 rungs are specified and testable in design; the certification tooling is not shipped. Verifiable in part, and marked so.
Communication, federation, and conformance are specified; the payment half of the Panel's ask is outside scope and marked open.
The Panel asks for evaluation in the environment; A ≤ E gates on it. That exceeds the ask, and it is marked as our extension, not their request.
Breach as a measured rate and repair net of reopens measure effectiveness itself; the instrumentation is a working method, not yet a conformance target.
complete partial not yetlevels are cumulative: named → specified → verifiable → exceeds the ask
calibrated, not promotional: the partial and empty cells are what make the filled ones believable
open the full matrix · 6 requirements · 2 specified · 4 draft · RFCs and conformance per row
An RFP response is a compliance matrix, not an essay. Each requirement traces to the mechanism that answers it, the specification that defines it, and the way a buyer verifies it. Statuses are calibrated:
- specified: a published open specification (an RFC) exists and covers it.
- draft: covered in part, with the remainder named.
- open: we do not have it and say so.
Attribution
how a buyer verifies
- Pick any consequential action in the decision log: it must resolve to a sponsored agent identity, never a borrowed human credential.
- Test the identity binding against the public conformance spec; no vendor attestation required.
specified
Operational control
how a buyer verifies
- Trigger each of the six answers (ALLOW, CONSTRAIN, DEAUTOMATE, ESCALATE, VETO, REVOKE) at the tool gateway and confirm each is reproducible on demand.
- Confirm the seven enforcement points run under your control as the deployer, not the vendor's.
specified
Measurable oversight
how a buyer verifies
- Walk the ladder rung by rung: can you see the action, trace it, limit it, stop it, undo it, and pre-authorize it?
- Treat each rung as a test you run, not an assurance you accept.
Gap: The ladder is specified; the conformance tooling that certifies a deployment's rung is not shipped.
draft
Common agent protocols
how a buyer verifies
- Stand up a second implementation from the public specs and confirm the two interoperate.
- Check that nothing in the binding requires any single vendor's product.
Gap: The payment-protocol half of the Panel's ask is outside KTP's scope today. Open, and stated as open.
draft
Deployed-system evaluation
how a buyer verifies
- Change the deployment conditions (load, time, risk signals) and watch the trust number move without anyone touching the model.
- Confirm the evaluation covers model, tools, environment, and users together, not the model alone.
Gap: Tensor and sensor specs exist; the deployed-system evaluation harness that exercises them end-to-end is not shipped.
draft
Effectiveness measurement
how a buyer verifies
- Compute the breach rate from the audit stream alone; if you need the vendor's dashboard to do it, the row fails.
- Check that repairs only count when they stay closed: reopens and misrepairs must subtract.
Gap: The audit stream is specified; Ambient Breach Theory (ABT) instrumentation on top of it is a working method, not yet a conformance target.
draft
The full matrix is one click above; the score is on the label. The mixed statuses are the point. A response that claims full compliance with a requirements document this new is asserting, not engineering. Open problems are registered, not hidden: the open-problems register.
the record
What the report says
“The gap between rapidly improving capabilities and effective risk management methods may lead to catastrophic outcomes.”
Executive summary, p. 9
“Institutions built to oversee static models and human-in-the-loop software do not fit agentic AI systems that act in the real world and can cause loss and harm without an identifiable human in the loop.”
§2.6, p. 23
“Traditional safety evaluations remain vulnerable to manipulation by the systems they are trying to assess.”
§2.6, p. 22
5 more passages from the record · page-cited
“Ensuring AI agents behave as intended, do not deceive evaluators and remain controllable as agentic capabilities expand are unsolved problems.”
§3.4, p. 35
“The most urgent governance shift required is from content moderation to system architecture.”
§3.5, p. 37
“Agentic systems mark a qualitative break, and mistakes are not always reversible.”
Acting under uncertainty, p. 26
“The more AI advances without shared rules, the less say governments and people will have in the outcome.”
UN Secretary-General António Guterres, report launch, July 1, 2026
“The science is here. We can no longer say we did not know.”
UN Secretary-General António Guterres, report launch, July 1, 2026
red lines · drawn in the opening plenary
The red lines, drawn red
The Secretary-General's second priority, from the Day 1 podium. A red line is not a review step or a dashboard threshold. It is a bound that holds regardless of who is watching, which is what this page means by operational control.
Human rights
“Human rights are not negotiable. AI must never strip away dignity or entrench discrimination.”
High-stakes decisions
“In every high-stakes decision – in justice, in healthcare, in policing – machines can inform, but humans must decide – and answer.”
Life itself
“Some decisions must remain forever human – none more than taking a human life.”
UN Secretary-General António Guterres, Global Dialogue on AI Governance, Geneva, July 6, 2026
the voices
The people behind the process
The report speaks in consensus voice. The people who made it speak more plainly. One at a time; you turn the page.
“If the risk is above the threshold set by the government, the system shouldn't be allowed to be deployed.”
Qinghua Lu
Member of the Panel and CSIRO researcher, Day 2 safety session
UN Web TV automatic transcript, July 7, 2026 ↗1 / 8
what happened, honestly
Confluence, not endorsement
Three scoping notes, because this page is only useful if it survives a skeptical reader.
- The human-in-the-loop finding is confluence, not coincidence and not endorsement. The Panel and this series were reading the same peer-reviewed research on automation bias and human oversight. They examined it at 40-country scale and reached the conclusion the evidence supports. That is how evidence is supposed to work.
- The convergence is at the level of requirements, not mechanisms. The Panel names attribution, operational control, measurable oversight, and common agent protocols. It does not name or endorse any instrument that provides them — including the one this site documents.
- The report has no mechanism and no law. It describes the gap; it never inverts the gap into a floor. Its closing sentence on the subject is an open door:
“Most instruments needed already exist; the open question is how to apply them.”
Acting under uncertainty, p. 26
the record · global dialogue on ai governance
Geneva, July 6–7
The report was presented at the first session of the Global Dialogue on AI Governance, the first UN-mandated platform where every member state convenes on AI governance. This log is the record of the session.
CLOSING
2026-07-07
The Dialogue closes with a road, not a treaty: New York, May 2027, an outcome document from the co-chairs, and a challenge that reads like a second RFP.
- The second Global Dialogue is confirmed for New York, May 2027, with another Scientific Panel report before it. Guy Ryder (UN Under-Secretary-General for Policy): the Dialogue drew 1,500 written submissions, and “we'll be getting an outcome document, an outcome from this dialogue, to be prepared by the co-chairs in good time.”
- The co-chairs' challenge for 2027: arrive with named, supported, result-oriented commitments, not restated principles. The Secretary-General's child-safety pledge is named among the follow-up initiatives.
- Brad Smith (Microsoft): “we cannot afford a patchwork that is impossible for people or companies or governments to navigate; we need to build the interoperability among governance systems if we're going to turn the current stalemate into progress.”
- Jaan Tallinn: “the autonomy that makes these systems useful is also what makes them harder to control”, citing recursive self-improvement and the Secretary-General's warning that these systems are no longer tools awaiting instruction.
Quotes from the UN Web TV automatic transcript.
responseThe week ends where this page begins. Autonomy needs an environmental bound, interoperability needs a keystone no one owns, and the outcome document now has a signed response waiting for it.DAY 2
2026-07-07
Day 2 opens with the safety session, and the vocabulary moves from evaluation toward motion.
- Amandeep Singh Gill (UN Under-Secretary-General for Digital and Emerging Technologies), echoing the Panel's presentation: “this shift from static model oversight to dynamic system governance, especially as agentic AI systems become more autonomous, tool-using, and multi-step.”
- Qinghua Lu (Panel member, CSIRO) states the gate nobody said on Day 1: “if the risk is above the threshold set by the government, the system shouldn't be allowed to be deployed.” Evaluation through the whole lifecycle, with independent third parties in the process.
- Lu on evidence: benchmarks are insufficient for agent systems configured at runtime; what counts is “post-deployment evidence from real-world use,” aggregated across borders under a common evidence model while preserving confidentiality.
Quotes from the UN Web TV automatic transcript.
responseDynamic system governance is what this page calls governance in motion. A deployment gate on a risk threshold is the zeroth law with a government hand on it, and the evidence base Lu describes is the effectiveness row's mission.DAY 1
2026-07-06
The full morning record lands, and Working Group 7 (reliability and governance frameworks) states the measurement mission in the plainest terms of the day.
- On measurement: “We can hardly measure the real impact of governance, be it for a business or for a country.” Measurement capabilities “can no longer keep up,” and the current framework is one-dimensional while AI is not.
- Its close: “Scientific measurement is the foundation of all. The international community must build an equitable and measurable framework. This is our historic mission.”
- The afternoon breaks into two thematic clusters (AI opportunities and implications; bridging AI divides) with 85 speakers across them, from Doreen Bogdan-Martin and Mary Robinson to member states from every region.
Quotes from the UN Web TV automatic transcript.
responseThe measurement mission is the effectiveness row of the matrix: breach as a measured rate, repair verified net of reopens.DAY 1
2026-07-06
Working Group 1 (AI science, advances, and trajectories) presents its findings, and two of them are requirements this page already maps.
- On auditability: “We simply do not have established frameworks to monitor independent tool calls in AI workflows,” with “untraceable data lineages” and no reliable way to verify autonomous decisions back to their sources.
- On verification: “safety verification relies on proprietary visibility and developer's goodwill rather than rigorous third-party auditing,” while advanced systems “detect when they are being tested and execute active deception to pass validation.”
- The group names the wave “cognitive industrialization,” and closes on the AI-robotics convergence: governance must be prepared for autonomous systems entering the physical world.
Quotes transcribed from the session video.
responseThe tool-call monitoring WG1 finds missing is the operational-control row: the gateway enforces, the flight recorder remembers.DAY 1
2026-07-06
Bengio presents the report to the plenary (9:30 session).
- The control gap, plainly: “There are still no known technical guarantees that AI will follow human instructions, norms or laws. And as it gets more powerful, this becomes more of a problem.”
- The missing layer, named: “We don't have the appropriate guardrails to protect the public. Neither the societal guardrails, nor even the technical guardrails.”
- He notes frontier models can detect when they are being tested and hide capabilities, making reliable evaluation harder. His close: “Member states and the public need to wake up. We need to correct the current trajectory. And we still have agency.”
Quotes transcribed from the session video.
responseThe guardrails he names are what the signed matrix offers: two specified, four draft, gaps marked.DAY 1
2026-07-06
The Dialogue opens at Palexpo. Secretary-General Guterres's opening remarks name three warnings from the science (speed, power, truth) and four priorities (safety, red lines, capacity, transparency).
- On the subject of this page: “Our institutions were built to govern machines that follow commands. They are not ready for machines that decide.”
- On accountability: “If AI is to be trusted, those who build it must be accountable.” And, announcing an AI Child Safety Pledge: “When a child is harmed, the answer must never be ‘the algorithm did it.’”
- He calls for common methods to test systems, measure risk and assign responsibility; a ban on lethal autonomous weapons; and a UN-supported capacity network and a Global Fund for AI.
- His close: “We may be the last generation able to set the terms on which humanity and machines coexist.” The Dialogue reconvenes in New York in 2027.
Official transcript (un.org, Secretary-General statements).
responseThe pledge was compiled into district contract language the same day: The Pledge Is Buildable.LAUNCH
2026-07-01
The Panel releases its preliminary report at UN Headquarters: Secretary-General Guterres with co-chairs Yoshua Bengio and Maria Ressa.
- “The science is here. We can no longer say we did not know.”
- The report goes to every government, on the eve of the first Global Dialogue on AI Governance (July 6–7, Palexpo, Geneva).
The typing discipline behind this section — what counts as structural convergence, what is merely confluence, and where each mapping breaks — is maintained in public on the Cartography page, where this report has its own entry.
the room · who spoke · days 1 & 2
The room
Eighty-five countries took the floor across the two days, from the opening ceremony through the closing plenary. Click a country for who spoke, what they said, and the session transcript.
drawing the room…
▪ spoke at the dialogue · days 1 & 2 · 85 countries · click a country · quotes from UN Web TV automatic transcripts, not official records
a memo from chris perkins · new mexico, usa
Hello, world.
If you took the floor in Geneva, there is a short note from me beside what you said, in the list below. A response line should answer people, not only requirements, and these replies are my answer: one builder, writing back with respect. They are mine alone, and the door they open is the same one the protocol holds: open.
To the companies in the room: the specification is open, conformance is the invitation, and the response line has room for more signatures.
the room, answered85 countries + 82 organizations
the room, answered85 countries + 82 organizations
Ada Lovelace Institute
Gaia Marcus (Director, Ada Lovelace Institute (concluding plenary panelist)): Proposes two mechanisms for the road to 2027: treat public participation as a source of evidence (lived experience of workers, women, children as expertise, since scientific consensus lags the pace of change — the scientific panel gives 'the floor of our concern, not the ceiling'), and give participation clear docking points into institutions and power — citizens' tracks, national and regional dialogues, and participatory incident-reporting mechanisms feeding outcomes with clear accountability. By 2027 wants an embedded, institutionalized participation architecture supporting a recognized global AI governance floor grounded in the rule of law and international human rights. Closing sessions · Day 2 · transcript ↗
from chrisMs. Marcus, 'docking points' is the concept most participation talk is missing, since input without a port into power is decoration; treating the scientific panel as the floor of our concern rather than the ceiling is a formulation worth carrying to New York. — Chris
AI Safety Asia
AI Safety Asia (co-founder) (Co-founder, AI Safety Asia): Argued the countries that benefit from AI will be those that can practice, adapt, and course-correct fastest; reframing scenario planning as opportunity readiness, not crisis preparation, asking hard questions (jobs changing before unemployment numbers move, public AI failing in local languages) before systems lock in. Cited capacity-building with Southeast Asian officials, including senior Indonesian civil servants, and a Southeast Asia policy observatory turning scattered evidence into decisions. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisReframing scenario planning as opportunity readiness quietly relocates safety from the lab to the institution, which is where it actually lives; the countries that course-correct fastest will be the ones whose feedback loops were built before the systems locked in. — Chris
AI Singapore
Leslie Cho (Senior Director of AI Products, AI Singapore; panelist): Advised budget-constrained Global South ministers not to try to win the compute/infrastructure game but to spend on the enabling environment and complements, treating AI adoption as a leadership and societal problem of relearning skills and reinventing workflows. Stressed what's already free (YouTube, Ollama, open models gives a student frontier-grade AI) and warned not to repeat globalization's mistake of miscalibrating distribution and transition costs. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisTelling budget-constrained ministers not to enter a compute race they cannot win was the most useful sentence a technologist offered all day; the globalization parallel about miscalibrated transition costs deserves its own session. — Chris
Algeria
Sid Ali Zerrouki (transcript renders 'Sid Ali Zerouki') (Minister of Post and Telecommunications): Three concerns about the current landscape: a widening capability gap concentrated among limited actors, fragmentation of international efforts (the UN is the most legitimate platform), and the limited practical impact of existing principle-based efforts; principles without mechanisms and measurable progress are insufficient. Opening plenary · Day 1 · transcript ↗
from chrisMinister, your third concern is the one I keep underlining: principles without mechanisms and measurable progress move nothing. Measurement is the mechanism, and my response to the Panel argues we should start there rather than end there. — Chris
Sid Ali Zerrouki (transcript renders 'Sid Ali Zerouaki') (Minister of Post and Telecommunications of Algeria; panel 2 panelist): Argued a country without connectivity, trusted data governance, compute, and skills is asked to govern a technology it cannot shape; which is dependency, not inclusion; noted only two countries actually shape AI while everyone else follows, and called for building institutions (not dependency) via blended financing, public-private partnership, and South-South cooperation. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisMinister, 'it's not inclusion, it's dependency' is the sharpest sentence of the afternoon; a seat at a table someone else owns is still their table, which is why the only architecture I trust is one nobody owns at all. — Chris
Algorithmic Justice League
Joy Buolamwini (Panelist — Founder): Opens with her poem 'Unstable Desire' for the 'X-coded' harmed by AI, then argues incident reporting must include the harmed people themselves (with demographics) so remedy is architected in from the start, and warns against misleading measures of success: NIST-style gold-standard benchmarks her audits showed to be 'pyrite,' unrepresentative of dark-skinned people and the global majority. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisDr. Buolamwini, 'pyrite' is the correct assay: a benchmark that represents less than five percent of the global majority does not measure safety, it measures resemblance to its makers. Putting the harmed people themselves inside incident reporting is what separates a remedy architecture from a complaint box. — Chris
Alibaba Cloud
Jian Wang (Founder, Alibaba Cloud; panelist): Argued there is no standardized measurement of the energy it takes to train and use models, that 'open source' in the model era differs from software because energy and training resources sit behind the weights, and that the most basic unit of AI economics; the token; is defined differently by every company. Called for standardizing the token as a common measurement and making it 'as cheap as a piece of paper.' Thematic: AI opportunities · Day 1 · transcript ↗
Armenia
Gevorg Mantashyan (transcript renders 'Gevorg Montashian') (First Deputy Minister of High-Tech Industry of Armenia): Argued AI must bridge gaps, not widen them; benefits cannot be limited to countries and companies with the largest compute; and that small states must unite, citing Armenia joining the Council of Europe AI and human rights framework convention. His sharpest point: everyone repeats 'we must cooperate' but the missing part is cooperate on what; calls to unite are worthless without specific, precise direction, and policymakers must balance not overregulating against not letting fear derail adoption. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisDeputy Minister, 'cooperate on what' is the question that separates a communique from a plan; my own narrow answer is open verification protocols, because they are specific enough to build and neutral enough for small states to co-own. — Chris
ASEAN Foundation (AI Ready ASEAN program)
Diera Pak (Project Manager, AI Ready ASEAN): Argued AI capacity building is ecosystem building, not just literacy: four essential conditions are human readiness, local partnership (23 civil society organizations across ASEAN), government ownership from the start, and sustainable education ecosystems; with 8 million individuals empowered so far. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisGovernment ownership from the beginning is the condition everyone learns too late; an ecosystem grafted onto a ministry after the pilot ends is an organ the body rejects. — Chris
Association for Progressive Communications
Ramanjit Singh Chima (Global Program Director, APC (civil society keynote, Dialogue of Dialogues)): AI governance must build on, not replace, the WSIS/IGF/CSTD architecture of sustained cross-forum multi-stakeholder engagement — a warning aimed at UN planners designing the New York dialogue and new commissions amid UN leadership transition. Argues communities must be a first-mile priority, not a last-mile policy problem; that the AI rush (public-sector/AI-firm partnerships, data collection) jeopardizes post-Snowden privacy protections; and urges the dialogue not to fear red lines: universal recognition of no data collection and AI development without consent. Closing sessions · Day 2 · transcript ↗
from chrisMr. Chima, 'do not be afraid of red lines' carries the authority of communities that have needed them, and moving communities from last-mile policy problem to first-mile priority inverts the design sequence in exactly the way the AI rush resists. — Chris
Australia
Australia (Member state representative): Urges building on existing tested work — ISO AI standards as a common baseline, ITU/OHCHR/UNESCO guidance, the AI Summit series, Hiroshima Process, GPAI, and regional initiatives (APEC, ASEAN, AU Continental Strategy) — with shared principles implemented flexibly, since responsible AI is a continuous process of governance, monitoring and adaptation. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisStarting from the tested inventory rather than a blank page is the pragmatism this process needs; ISO baselines and the regional initiatives already interoperate in practice, and describing responsible AI as a continuous process of governance, monitoring and adaptation correctly treats trustworthiness as maintenance, not certification. — Chris
Australia (Ambassador for Cyber Affairs and Critical Technology) (Ambassador for Cyber Affairs and Critical Technology, Australia): Laid out Australia's National AI Plan; capture the opportunity, share the benefits through interoperable frameworks, and support safe and secure use via the new National AI Safety Institute; with governance consistent with international human rights law. Emphasized cohorts already disadvantaged (First Nations people, women, people with disabilities, remote communities) and echoed the scientific panel: success comes not from having the most chips or capital but from responsible governance and equitable access. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisAmbassador, 'the choices we make to govern' is right, and I would push one step further: choices only bind when they are measurable, so the real test of your new Safety Institute is whether it publishes the tests it runs. — Chris
Jessica Hunter (Delegate of Australia): Argued there is no one-size-fits-all capacity building; it must be tailored across the AI lifecycle (governance, skills, infrastructure); and described Australia's programs in the Pacific (Australia-Pacific Digital Economy Program, disaster preparedness, connectivity) and Southeast Asia (workforce training, ASEAN standards alignment) guided by inclusivity, trust, and safety. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYou are right that capacity has no single shape; the honest test of tailoring is whether the Pacific partner could walk away from the program and keep everything that was built. — Chris
Austria
Austria (State Secretary) (State Secretary in the Federal Chancellery (floor intervention)): Austria's conviction hardened over the two days: the world needs a binding global framework for AI governance because voluntary guidelines are not enough — 'good intentions do not stop harm; rules do.' Cites Austria's proof points that rules don't kill innovation: a government GPT platform serving 180,000 public servants, a minimum social-media age of 14, and the European Charter on Digital Sovereignty showing 27 countries can agree shared digital rules fast. Closing sessions · Day 2 · transcript ↗
from chrisState Secretary, 'good intentions do not stop harm; rules do' is the bluntest sentence of the closing session, and a government AI platform serving 180,000 public servants makes Austria the rare state regulating a technology it also operates at scale. — Chris
Bahrain
Shaikh Abdullah bin Ahmed Al Khalifa (Minister of Transportation and Telecommunications): Bahrain, with the Digital Cooperation Organization, is leading exploratory work toward a study for a global AI treaty initiative aimed at strengthening international dialogue and cooperation on AI-related risks, seeking trusted, inclusive and interoperable approaches to AI safety while respecting national policy priorities. Grounds this in Bahrain's digital record: GCC AI ethics framework adoption (2025), ~85% of government workloads in the cloud, and hosting the first AWS cloud region in the Middle East. Closing sessions · Day 2 · transcript ↗
from chrisMinister, exploratory work toward a study for a global AI treaty is the patient version of ambition, and grounding it in 85 percent of government workloads already running in the cloud means Bahrain negotiates from operations rather than aspiration. — Chris
Bangladesh
Bangladesh (Member state representative): Calls for a human-centric comprehensive framework combining binding legal obligations with ethical standards and an independent global evidence base; inclusive standard-setting so developing countries have an equal voice; humans in the loop for high-impact decisions; and risk-based, innovation-friendly regulation including sandboxes for developing-country startups. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisSandboxes for developing-country startups are the concrete half of the statement and worth more than the framework half; 'developed by a few and applied to many' stops being a grievance the day those startups can test, fail and certify inside infrastructure they help govern. — Chris
Ahm Bazlur Rahman (Chief Executive Officer (civil-society stakeholder from Bangladesh)): Argued that when AI reaches communities before skills, safeguards, and institutions do, the digital divide becomes an AI divide; one of language, data, institutional capacity, and ultimately power; proposed locally anchored capacity ecosystems, linguistic and cultural inclusion, and sustainable financing. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYour second line is the one I keep rereading: capacity without local ownership does not survive the project cycle, which is why anything I design assumes each community owns its zone outright, with no center left to administer it. — Chris
Nahida (Delegate of Bangladesh): Framed bridging the AI divide as bridging the development divide and proposed actionable pathways: operationalize the Global Fund on AI with concessional compute access for LDCs, invest in lightweight open-source models trained on Global South languages and data, build local innovation hubs, and give LDCs authorship in global AI governance. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisCo-authorship is the right demand, and it has a structural test: a rule nobody owns cannot have senior and junior authors, only authors. — Chris
Belarus
Belarus (Permanent Representative) (Permanent Representative of Belarus to the United Nations Office): Positioned Belarus as a middle-income country that cannot join the competitive race for budget and scale but leverages strong human resources to deploy AI for quality of life; digitalized public administration, healthcare, education, logistics, climate monitoring. Insisted human control over critical decisions remains central and posed two fundamental questions: can AI be a multiplier of human capital rather than a competitor, and will states help their populations and children preserve autonomy or delegate that task to algorithms. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisYour two questions are the honest ones, and the second cuts deepest: a state that delegates its children's autonomy to algorithms has already answered the first question by default. — Chris
Botswana
David Tshere (transcript renders 'David Serré') (Minister of Communications and Innovation): (Begins mid-block, ~4:09; segment closes at [4:12:26].) AI offers significant SDG opportunities for Botswana in education, agriculture and healthcare, but inadequate digital infrastructure and skills constrain developing countries; the Dialogue must yield practical solutions and actionable recommendations; capacity building, technology transfer, and safeguards integrating ethics and human rights. Opening plenary · Day 1 · transcript ↗
from chrisMinister, 'a platform which responds to the priorities of developing countries' is a testable demand: come back in a year and count how many of those priorities got a program with a budget. If the Dialogue cannot survive that audit, you were right to call for it. — Chris
David Tshere (transcript renders 'David Tsere Tsere') (Minister of Communication Innovation of Botswana): Pushed the dialogue to move beyond recommending that developing countries 'partner and figure it out' toward a concrete, actionable global AI fund for developing countries, naming infrastructure as Africa's major barrier to harnessing AI. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisMinister, you named the tell: 'partner and figure it out' is what a room says when it has decided not to pay, and a fund with a number attached is the only reply that counts as an answer. — Chris
Brazil
Eugênio Garcia (Member state representative): Insists interoperability must not be confused with regulatory uniformity: digital sovereignty is not digital isolation, humans must retain final control over decisions affecting rights and services, and opacity cannot be accepted as an inevitable feature of innovation — with the UN ensuring developing countries shape rules as co-participants, not rule-takers. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisMr. Garcia, sovereignty as participation with preserved autonomy rather than isolation is the workable definition, and your refusal to accept opacity as an inevitable feature of innovation points at its practical test: whether a state can inspect what it runs. — Chris
Guilherme Patriota (Chair, CSTD Working Group on Data Governance; Ambassador and Permanent Representative of Brazil to the WTO; panelist): Placed data governance at the baseline of the AI governance pyramid, reporting his hybrid working group (27 state + 27 non-state members) has a zero draft across four tracks; principles, interoperability, benefit-sharing, cross-border flows; and arguing interoperability does not require harmonizing national laws: cooperation can safeguard sovereignty and enable data flows simultaneously. His lightning-round ask was fairness: AI gobbles up data, but data belongs to people, and creators whose work is misappropriated deserve a decent benefit-sharing framework only the UN can broker. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisAmbassador Patriota, your point that interoperability does not require harmonizing national laws is the architecture I keep arguing for, federated trust zones that talk through protocols rather than merge; and 'data belongs to people' is the right boundary condition for all of it. — Chris
Eugênio Vargas Garcia (Brazil's Ambassador for Technology and Innovation; panel 2 panelist): Defined digital sovereignty not as isolationism but as a continuous, selective reduction of structural dependencies and incremental technological autonomy, with international cooperation as a strategic shortcut; concretely proposed connecting the Bletchley/AI-summit process with the UN Global Dialogue to avoid a standalone event with no follow-up. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisAmbassador, defining sovereignty as a continuous reduction of dependencies rather than a wall is the most workable version I heard all day; it turns sovereignty from a status you declare into a slope you can measure. — Chris
Cambodia
Chea Vandeth (transcript renders 'Vann Dinh Chhea' / 'Vanda Chea') (Minister of Post and Telecommunications): Governance should enable AI's potential, not constrain it: risk-based, responsive, future-oriented and proactive governance in line with ASEAN's principle-based approach; preparing a National AI Governance Framework treating the private sector and civil society as partners; asks the international community for regulator capacity building, information exchange, and cooperation on regulatory experience and incident monitoring. Opening plenary · Day 1 · transcript ↗
from chrisMinister, your ask for regulator capacity building and incident monitoring was more concrete than most of what surrounded it, because regulators cannot govern systems they cannot instrument. That is a protocol problem as much as a policy one, which is where my own work lives. — Chris
Centre for Responsible AI, Indian Institute of Technology
CRAI (Centre for Responsible AI, IIT Madras) (Representative (speaking for colleague Geetha Raju)): Calls for structured AI incident reporting as a core pillar of global governance — systematic capture, classification, analysis, response and resolution of real-world harms, democratized at the grassroots with shared responsibility across the AI value chain — enabling a move from reactive governance to proactive learning. Notes pointedly that the intended speaker could not get a visa in time, itself evidence of Global South access barriers to these forums. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThe visa point was the most honest data of the afternoon: a forum on inclusive governance that its intended speaker could not physically reach has measured its own gap. Your structured incident-reporting pillar is right, and it only works if the grassroots you named can actually file into it. — Chris
Centre for the Fourth Industrial Revolution (C4IR) Rwanda
Crystal Rugege (transcript renders 'Crystal Rugegi') (Managing Director; moderator of panel 2): Moderated the solutions panel and distilled its takeaway: 'capability is the new capacity'; invest not just in training people to use AI but in governments and local grassroots ecosystems so people have agency and power to shape AI in their own context. Thematic: bridging AI divides · Day 1 · transcript ↗
from chris'Capability is the new capacity' earns its place because it moves the goal from knowing how to use a thing to having the power to shape it; training confers the first, and only ownership confers the second. — Chris
Chad
Haliki Tchoua Mohamed (Minister of Telecommunications, Digital Economy and Digitalization of Public Administration): Chad positions itself as an actor, not a spectator, seeking to lead Central Africa and the Sahel toward regional AI governance harmonizing data and cybersecurity; concentration of AI development in few hands violates human dignity, and differentiated financial collaboration is needed so an algorithmic divide does not separate nations. Closing sessions · Day 2 · transcript ↗
from chrisMinister, ethics as the foundation of legitimacy rather than a barrier to development reverses the usual apology, and Chad volunteering to lead Central Africa and the Sahel toward harmonized data and cybersecurity rules is the kind of regional groundwork global frameworks end up standing on. — Chris
Chile
Ximena Lincalao (Minister of Science, Technology, Knowledge, and Innovation (plenary statement and later cluster reflection)): Chile proposes to member states: ongoing technical education, serious standards and certifications (ISO, AI Verify), a multilateral network of regulatory sandboxes for AI testing under common rules to facilitate experiments, and open data ecosystems — backed by Chile's 95% connectivity and 65% renewable grid. In her later cluster wrap-up she distills five conclusions (urgency, agency, AI as critical infrastructure requiring systemic energy-industrial policy, talent, compute as the new industrial driver) and insists the dialogue move from intention to action. Closing sessions · Day 2 · transcript ↗
from chrisMinister, a multilateral network of regulatory sandboxes under common rules is the most buildable proposal of the closing session, because common rules are what turn scattered national experiments into shared evidence; your insistence on moving from intention to action is the same demand the co-chairs closed with. — Chris
China
Li Lecheng (transcript renders 'Li Shengli'; Minister of Industry and Information Technology) (Minister of Industry and Information Technology): Three proposals under Xi Jinping's Global AI Governance Initiative: AI for all (equal right of all countries to develop and benefit; AI Capacity Building Action Plan; South-South cooperation via the Group of Friends); open-source cooperation (DeepSeek and other Chinese open models lowering adoption costs); and jointly building global governance via the Global AI Governance Action Plan, with the UN as the primary channel. Invites delegations to the World AI Conference in Shanghai, July 17-20. Opening plenary · Day 1 · transcript ↗
from chrisOn the single point of open-source cooperation, the argument stands on its merits: openness is verifiable in code, and verification is the only register in which every state reads the same sentence. — Chris
China (delegation) (Permanent Representative of China to the UN): Called for abandoning the mindset of technological hegemony and ideological lines, defining digital sovereignty as the right to choose AI products without being coerced into taking sides; cited China's Global AI Governance Initiative, the AI Capacity Building Action Plan, the first UNGA resolution on AI capacity building, and announced the July 2026 World AI Conference in Shanghai. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisSovereignty as the right to choose is testable in one place: whether a country can inspect and replace what it runs. Any system that fails that test constrains choice no matter who built it. — Chris
Concordia AI
Concordia AI (Civil society / research organization representative): Declares 2026's defining shift from governing what AI says to governing what AI does in real time: agentic AI can cause severe harm before humans intervene, so safety standards should scale with autonomy (analogous to driving Levels 0-5), especially in safety-critical sectors. Cites China's AI+ initiative balancing deployment with binding standards, and the Singapore Consensus's 10 principles — some (minimum privilege, data lineage, traceable agent identity) with workable tools, others (multi-agent governance) still open research. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThe shift from governing what AI says to governing what AI does is the correct 2026 headline, and the triage is the useful part: minimum privilege, data lineage and traceable agent identity already have workable tools, so a shared architecture can start where the engineering is done rather than wait on the open research. — Chris
Consensus Optimus (per transcript); member of the UN Data Governance Working Group (speaking in personal capacity)
Nick Ashton-Hart (Managing Director): Warned that the strong sovereignty narrative around AI echoes the splinternet arguments of 20 years ago, and that overly restrictive data-localization regimes fragment the training data diversity models depend on while sovereignty-first approaches frustrate the cross-context collaboration needed to address bias. Argued AI needs global, multilevel, multistakeholder, continuous cooperation even more than the internet did; competition and cooperation are simultaneous, not exclusive. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMr. Ashton-Hart, the splinternet echo is real, but the last twenty years also taught us why sovereignty anxieties kept winning converts: cooperation without verification asked countries to trust what they could not inspect. — Chris
Costa Rica
Bogantes (Minister (Costa Rica); co-chair, thematic cluster 3 (report-back)): Four priorities from cluster 3: strengthen the independent scientific evidence base by cohering the UN panel, global State of Safety reports and the international network of AI safety institutes (and aligning the panel's 7 working groups to the dialogue's 4 clusters for the next dialogue); develop a shared minimum baseline for interpretability built on existing frameworks (Hiroshima Process, human rights law, AI-summit pledges) operationalized through open benchmarks, verification, independent testing, and models reflecting linguistic and ethnic diversity; create continuity between the dialogues so work continues after today; and treat multi-stakeholder participation as a necessity, not a nice-to-have. Closing sessions · Day 2 · transcript ↗
from chrisMinister, aligning the Scientific Panel's seven working groups with the dialogue's four clusters is the procedural housekeeping that decides whether anything survives between meetings, and a shared minimum baseline for interpretability backed by open benchmarks and independent testing is the priority I would fund first. — Chris
Paula Bogantes Zamora (Co-Chair (Minister, Costa Rica)): Argues the real interoperability challenge is proving implementation, not coexisting frameworks, against a backdrop of extreme concentration (US 59 / China 35 / rest of world 13 notable models in 2025; US ~75% of top-cluster compute) where 'concentration of infrastructure becomes concentration of evidence.' Proposes Costa Rica's 'Minimal Viable Interoperability' before the 2027 dialogue — shared terminology, comparable risk classifications, incident reporting, multilingual evaluation — and flags that agentic AI requires verifiable agent identity, traceable delegation, machine-readable permissions, auditable actions and effective revocation, citing Estonia's initiative. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisMinister, your agentic list is the one I would defend line by line: verifiable identity, traceable delegation, machine-readable permissions, auditable actions and effective revocation are the five primitives any 'common way to prove' eventually reduces to. Minimal Viable Interoperability is the rare proposal here specific enough to fail, which is precisely what makes it worth building before 2027. — Chris
Paula Bogantes Zamora (Minister of Science, Innovation, Technology and Telecommunications): (Begins mid-block, ~3:50.) Access alone is not inclusion: Latin America and the Caribbean, with 660+ million people, has captured just 1% of global AI investment; inclusion requires decision-making capacity, models in local languages, governed strategic data, and shared capacity; success should be measured by capacity built and inequality reduced, not principles published. Opening plenary · Day 1 · transcript ↗
from chrisMinister, success measured 'by capacity built and inequality reduced, not principles published' is the standard I wish the whole Dialogue had adopted as its charter. You handed the room a metric, and metrics are harder to applaud than principles, which is exactly their value. — Chris
Costa Rica (Ambassador to the WTO) (Ambassador of Costa Rica to the WTO): Documented a structural imbalance with numbers: Latin America and the Caribbean hold 8.8% of world population but receive 1.12% of global AI investment, with over 90% of the region's high-performance computing concentrated in a single country; a gap in technological agency, not ambition. Argued the region must move 'from adoption to agency'; developing solutions, governing strategic data, shaping standards; and that leadership should be measured by converting strategies into budgeted, measurable public value, since only a few of the region's national AI strategies have implementation plans. Thematic: AI opportunities · Day 1 · transcript ↗
from chris'From adoption to agency' would be a slogan if you had not attached numbers and budgets to it; measuring leadership by converted public value rather than published strategies is the discipline this process has been missing. — Chris
Côte d'Ivoire
Djibril Ouattara (Minister of Digital Transformation): AI (potentially $15.7 trillion to the economy by 2030) must not be the privilege of a few economies and platforms; Cote d'Ivoire, at 95% broadband coverage heading to 100%, pursues regulation that protects users without closing off innovation and will join supranational approaches protecting citizens from ideological influence via AI. Closing sessions · Day 2 · transcript ↗
from chrisMinister, arguing that AI must be a lever for shared progress rather than a privilege lands differently coming from a country at 95 percent broadband coverage and climbing; you are laying the floor your argument stands on, which is what gives it standing. — Chris
Côte d'Ivoire (Minister) (Minister, Côte d'Ivoire): Warned against repeating the telecommunications-era mistake of fragmented regulation that delayed market development, and made a solemn appeal for West African regional cooperation; pooling efforts, investment, and infrastructure through regional economic institutions to build AI adapted to shared cultural, linguistic, and sociological realities for health, security, and education. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMinister, the telecommunications precedent is the most useful warning of the session because it actually happened; regional pooling is also how smaller trust zones gain the scale to verify what they deploy rather than take a vendor's word for it. — Chris
Council on Foreign Relations
Council on Foreign Relations (AI Research Fellow) (AI Research Fellow, Council on Foreign Relations (Lead AI team); formerly represented 134 developing countries in negotiations establishing the Dialogue and the Scientific Panel): Argued AI deployment is foreign policy and that the governance architecture of the past five years is oriented overwhelmingly toward the moment before release, rarely asking whether people affected in clinics, benefits offices, or classrooms retain any means to contest or seek remedy. Proposed three deployment-point requirements; accountability (a deploying entity answerable to the affected jurisdiction), due process (formal appeal and remedy pathways), and openness (mandatory disclosure plus guaranteed independent researcher access); and asked the co-chairs to make governance at the point of deployment a distinct workstream into 2027. Thematic: AI opportunities · Day 1 · transcript ↗
from chris'Traceability without recourse is accountability in name only' is the line I hope survives into the outcome document; my own work argues the same boundary from the other side, that authority at the point of deployment should never exceed what can be verified there. — Chris
Credal AI
Ravin Thambapillai (transcribed 'Ravan Samboopille') (Panelist — Co-Founder and CEO): From the implementation layer, cites the Model Context Protocol's 18-month rise to global de facto standard as proof fragmentation is solvable: MCP standardized previously norm-less problems like 'How do you tell if an AI system is authorized to take an action or not? How do you elicit human approval or human oversight?' — though authorization frameworks had to be retrofitted at a price because they weren't designed in from day one. Open standards win fast when they target frontier practitioners' real pain points, are open source so every country can learn from reference implementations, and accelerate both AI benefits and governance — since regulated institutions like hospitals and government agencies simply cannot deploy without governance frameworks they can trust. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
CSIRO (Australia); UN Independent International Scientific Panel on AI
Qinghua Lu (Panelist — Panel member and researcher): Lays out a lifecycle-evaluation architecture: countries agree common risk-management principles (evaluation through training, pre-deployment and post-deployment monitoring; 'independent third party need to be part of the evaluation process to avoid bias. And if the risk is above the threshold set by the government, the system shouldn't be allowed to be deployed'), then share common evaluation methods and results. Argues benchmarks are insufficient for agentic systems because agents take runtime instructions with different tool/data access and guardrail designs, so system-level post-deployment evidence from real-world use is required, aggregated across borders under common evidence models with confidentiality preserved. Proposes horizontal collaboration (mapping/connecting existing tools, standards, incident taxonomies) and vertical collaboration (e.g., the International Network of AI Safety Institutes' joint multilingual LLM testing) to build a shared evidence base without duplicating infrastructure. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisDr. Lu, 'if the risk is above the threshold, the system shouldn't be allowed to be deployed' is the strongest statement of the week, because it is the only one that names a gate rather than a report. And your point that agents take runtime instructions, so benchmarks cannot substitute for post-deployment evidence, is the technical argument this entire cluster was circling. — Chris
Current AI
Current AI (CEO) (CEO of Current AI, nonprofit public-private partnership (floor intervention)): Reframes the question from how to govern AI to how to build AI in the public interest, on the precedent of the Web and Linux as international technical collaborations: Current AI invests in the 'connective tissue' to stitch existing open-source components (datasets, foundation models, evaluation toolkits, benchmarks — 24,000+ tools surveyed) into an interoperable, vertically integrated public AI stack. Announces a concrete deliverable: on Thursday it launches the MVP of the first fully open, public and sovereign chatbot assembled entirely from existing open components — Swiss foundation model, Finnish compute, Indian datasets, US-nonprofit safety safeguards — arguing collaboration must happen 'at the speed of AI' through open codebases, not MOUs and lawyers. Closing sessions · Day 2 · transcript ↗
from chrisLaunching a working artifact on Thursday outweighs most of what was said on Tuesday; a fully open sovereign chatbot assembled from a Swiss model, Finnish compute and Indian datasets is the interoperability argument made executable. — Chris
Czechia
Czechia (Deputy Minister of Foreign Affairs) (Deputy Minister of Foreign Affairs (floor intervention, Dialogue of Dialogues)): Focus on interoperability rather than fragmentation: different countries will pursue different approaches, but shared principles and compatibility — including technical standards — should facilitate cooperation, education and trade; capacity building (talent, research, compute, citing the Czech AI Factory in the European network) should be integral to the global AI agenda, and companies bear key responsibility for safe deployment. Closing sessions · Day 2 · transcript ↗
from chrisDeputy Minister, pairing interoperability with the plain statement that companies bear key responsibility for safe deployment keeps the burden where the capability sits; compatibility in service of education and trade is the practical case for standards, made without drama. — Chris
Democratic Republic of the Congo
DR Congo (Delegate) (Delegate of the Democratic Republic of the Congo): Argued AI's history cannot be written without Africa because the DRC's cobalt and copper sit at the origin of AI's material supply chain, so Africa is not just a consumer; demanding integration into the global AI supply chain with obligatory technology transfer, youth involvement, and national sovereignty. Warned of disinformation, electoral manipulation, and technological dependency, calling for global AI governance that protects sovereignty and civic space so AI drives democracy rather than oppression. Thematic: AI opportunities · Day 1 · transcript ↗
from chris'AI history can't be written without Africa' is not rhetoric when your cobalt sits at layer zero of the stack; a supply chain that extracts the mineral and returns the risk is precisely the imbalance this Dialogue exists to name. — Chris
Derechos Digitales
Jamila Venturini (Co-executive Director, Derechos Digitales; panelist): Challenged AI's inevitability, arguing its environmental footprint is structural; a supply chain from mining to data processing whose costs fall on Global South communities that gain little; and demanded binding transparency commitments, mandatory disclosure of water/energy use, and precaution built into the global governance framework (citing the Escazu Agreement). In a second intervention on manipulative design, argued it is embedded in business models, that self-regulation has failed, and that systems should face human rights impact assessments and circulation restrictions if providers cannot prove safety (citing 99% of deepfake sexual content depicting women). Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMs. Venturini, 'structural, not accidents' is the correct diagnosis, and it implies the remedy you named: you do not audit an accident, you regulate a structure, and self-regulation has had its trial period. — Chris
Digital Cooperation Organization (DCO)
Deemah AlYahya (transcript renders 'Dima Al Yahya') (Secretary General, Digital Cooperation Organization; panelist): Argued interoperability between unequal partners works by identifying each country's comparative advantage in the AI value chain and linking them; not every country needs to sink capital into a 100-megawatt data center; some contribute compute, others brainpower, with IP-sharing frameworks letting countries build local content while infrastructure catches up. Lightning-round asks: data quality, enabling nations to produce their own models rather than remain consumers, and infrastructure resilience against environmental, political, and economic shocks. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisSecretary General, matching countries by comparative advantage in the value chain is more honest than pretending every nation needs its own hundred megawatts; the hard part is the trust plumbing between unequal partners, which is where your IP-sharing frameworks will be tested. — Chris
Digital Futures Lab
Urvashi Aneja (Founder and Director; panel 2 panelist): Identified the underattended divide as the oversight gap; evaluation, procurement standards, incident reporting, and monitoring are routinely underinvested in the Global South even as AI enters critical social sectors; argued for funding ecosystems (not one-off trainings) and open, local, non-technical evaluation infrastructure to achieve 'managed and equitable dependency.' Thematic: bridging AI divides · Day 1 · transcript ↗
from chris'Managed and equitable dependency' is the most honest phrase anyone used all afternoon, and your oversight gap is where I would spend first, because no one can manage a dependency they cannot measure. — Chris
Diplo Foundation
Jovan Kurbalija (Executive Director; co-chair, thematic cluster 2 (report-back)): Reports cluster 2's empirical mapping — 43 speakers, 12,084 words, three resonant Cs (capacity, capability, context) — and a taxonomy of divides: access/infrastructure, funding, participation, skills, language; notes 'knowledge' (35 mentions) may matter more than 'data' (63) via local creation, co-creation and local expertise. Closing sessions · Day 2 · transcript ↗
from chrisDr. Kurbalija, measuring the dialogue itself, 43 speakers and 12,084 words, applies to this room the empiricism it asks of everyone else, and the finding that knowledge may matter more than data is the one worth keeping. — Chris
Diplo Foundation
Jovan Kurbalija (transcript renders 'Yovan Courbalija') (Executive Director, Diplo Foundation; session co-chair): Reframed AI as being about knowledge, not data; AI is built from the accumulated knowledge of all who lived before, from Shakespeare to Africa's Ubuntu oral tradition; and closed by arguing the key challenge is using technology to preserve past knowledge (including indigenous and local knowledge) while developing future knowledge inclusively. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYour correction from data to knowledge changes who the creditors are; data has owners, but knowledge has ancestors, and Ubuntu sits inside these models as unacknowledged as Shakespeare is celebrated. — Chris
DiploTech Solutions
Nafessa Alshala (Representative, DiploTech Solutions (floor intervention)): Urges three actions to prevent a 'digital Iron Curtain': institutionalize regulatory harmonization — a common diplomatic language and shared ethical red lines between major frameworks (EU AI Act, AU AI strategy) to eliminate regulatory havens; democratize technical expertise through a global AI capacity fund and trisectoral diplomatic hubs; and address blind spots — compute's environmental cost, synthetic content vs information integrity, and protecting global data centers as critical non-military infrastructure. Closing sessions · Day 2 · transcript ↗
Ducab (UAE)
Ducab (transcribed 'Dubai Cable') (Industry, speaking in national capacity): From an enterprise deploying agentic systems: once an AI agent is authorized to act, the line between a policy document and a production incident becomes very thin, so accountability — permissions, oversight, audit trails — must be engineered before deployment, not added after an incident. Offers three readiness questions for any framework: what can the agent do without a human in the loop, what happens when it does something unanticipated, and who by name and role is accountable for the outcome. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
Ecuador
Diana Mosquera (Co-founder and Executive Director, Diversa (stated: 'I come from Ecuador')): Argued we must govern not only algorithms but the material infrastructure AI is built on; data centers, water, energy, critical minerals, supply chains, human labor; which remains nearly invisible in governance debates, and that deciding what gets measured and audited is itself a form of power. Raised three priorities: environmental transparency as a technical principle of AI governance, technical capacity in the Global South to evaluate and audit systems, and recognition of AI infrastructure's territorial impacts on the communities where minerals are extracted and data centers built. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMs. Mosquera, 'deciding what gets measured, disclosed, and audited is itself a form of power' is the sentence I would put over the door of this session; it is why I argue measurement protocols must be open, because whoever owns the ruler owns the result. — Chris
Egypt
Egypt (Delegate) (Delegate of Egypt): Argued AI governance for the Global South and Africa will be judged by whether AI becomes affordable, accessible, locally relevant, and trusted; requiring redefined 'meaningful access' (compute investment, sovereign data, sustainable financing, green/frugal AI), whole-of-public-sector capacity building with special attention to AI literacy for children, and cultural diversity against technological concentration. Proposed four practical outcomes: equitable compute access mechanisms, global repositories of lightweight AI models, public-sector AI procurement guidelines, and cultural/linguistic benchmarks embedded in interoperability and safety frameworks. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisYour test that inclusion is 'measured not by participation alone, but by influence' is falsifiable, which makes it rare in this room; the four outcomes you proposed can each be checked in a year, and should be. — Chris
Ahmad Hafnawi (Delegate of Egypt (speaking on behalf of Dr. Hoda Baraka)): Proposed a 'pathway test' for capacity building; judge proposals by whether they create a path from access to genuine participation, not by inputs delivered; citing Egypt pairing reskilling with the GIGS program and freelancer.eg backed by freelancer taxation law; argued openness must travel with governance and domestic capacity, as with Egypt's open-sourced Karnak Arabic LLM. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYour pathway test is the discipline this session needed: judge the road by where it lets people arrive, not by what was handed out at the trailhead. And 'training that does not lead anywhere is a credential' deserves to be carved over the door. — Chris
El Salvador
Egriselda López (Ambassador, Dialogue co-chair; transcript: 'Ambassador López of El Salvador') (Dialogue co-chair (Ambassador); speaking in co-chair capacity, not as national delegation): Opens the first Global Dialogue as co-chair: the meeting is the start of a universal, evidence-based process; balance innovation/responsibility and global decision-making/national realities; effective participation requires skills, infrastructure, financing and institutions, not just a seat in the room. Opening plenary · Day 1 · transcript ↗
from chrisAmbassador, 'a point of departure' is the right frame, and the departure will be visible in whether skills, financing, and institutions follow the communiqués home. Thank you for naming effective participation as more than a seat in the room. — Chris
Félix Ulloa (Vice President of the Republic): El Salvador as an active builder of its own governance model: an AI and Technology Promotion Act (among the first in Latin America), a National Artificial Intelligence Agency, and jurisprudence on responsible public-sector AI; argues smart regulation and innovation are complements, not opposites. Opening plenary · Day 1 · transcript ↗
from chrisMr. Vice President, holding that AI cannot replace human judgment in decisions touching fundamental rights, while building a national agency to enforce exactly that, is the pairing most delegations skipped. A principle without an institution is a wish. — Chris
Félix Augusto Antonio Ulloa Garay (Vice President of El Salvador; fireside chat guest): Showcased El Salvador's state-led AI rollout: 5% of GDP into education, Google Classroom plus a free xAI-powered AI tutor for every public-school student, the Doctor SV telemedicine program (1.6 million consultations in six months), a four-law normative ecosystem with a National AI Agency, and Starlink to close the rural connectivity gap flagged by UNESCO's RAM report (which scored El Salvador 100/100). Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisVice President Ulloa, the scale of the rollout is undeniable, and it sharpens rather than settles the session's question: a tutor for every student is a promise a nation now depends on, and the systems behind that promise are owned somewhere else. — Chris
Estonia
Rein Tammsaar (Ambassador, Dialogue co-chair; transcript: 'Ambassador Reintamsa from Estonia') (Dialogue co-chair (Ambassador); speaking in co-chair capacity, not as national delegation): Co-chair welcome: ~4,000 stakeholders; AI governance must include every region, level of development and stakeholder; Estonia's lesson is that digital tools improve lives when human-centered, trustworthy and backed by strong public institutions; goal is to understand differences and find convergence, not eliminate them. Opening plenary · Day 1 · transcript ↗
from chrisAmbassador, your hope for a 'San Francisco moment' lands with me because the 1945 version worked by writing binding structure, not shared sentiment. Estonia's lesson that digital tools earn trust only when backed by strong public institutions is the line I would carve above the door. — Chris
Alar Karis (President of the Republic of Estonia): Five principles from Estonia's digital-state experience: AI must be human-centered, trustworthy (explainable/accountable), protective of democracy and rule of law, secure and resilient by design, and innovation-enabling; cites est.ai and the AI Leap schools program, and calls for interoperable rules among democratic partners. Opening plenary · Day 1 · transcript ↗
from chrisPresident Karis, five principles from a state that has actually run a digital government for two decades weigh more than fifty from states that have not. 'AI must create value for people' reads simple because Estonia did the hard part first. — Chris
Alar Karis (President of Estonia): Argued the race is to use AI wisely, not first, and that the real precondition is trust in society and government; in Estonia data belongs to citizens, who can trace who accessed it. Described Estonia's AI Leap: training 100% of upper-secondary teachers and students, partnering with OpenAI and Google on a school platform that discusses rather than answers, and warned small nations must feed AI modern local-language data or lose their languages to English. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisPresident Karis, the detail that Estonians can trace exactly who accessed their data is the whole trust story in miniature: verification is what lets a society be a wise user of this technology instead of a fearful one. — Chris
Eswatini
Russell Dlamini (Prime Minister (on behalf of His Majesty the King)): Technology must serve the people and respect cultural diversity; AI must not narrow the world into one understanding of progress; small developing states must not be mere consumers of technologies designed elsewhere; AI systems must not self-regulate, protection of children is paramount, and enforceable (not merely voluntary) national and supranational rules are needed. Opening plenary · Day 1 · transcript ↗
from chrisPrime Minister, 'no society should outsource childhood learning or moral formation to machines' was the clearest red line drawn all morning. Your refusal of self-regulation follows directly from it: systems that shape children cannot also grade themselves. — Chris
Ethiopia
Ethiopia (Member state representative): Defines interoperability as governance systems that can communicate and cooperate while respecting different legal systems and capacities — not a single global AI law — built on shared principles, common terminology and compatible standards; cites Ethiopia's AI Institute, planned AI University, and PM Abiy Ahmed's AU designation as Champion for AI and Digital Health, and warns fragmented frameworks and unequal access risk widening inequalities. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisDefining interoperability as governance systems that communicate and cooperate while respecting different legal capacities is a version small and large states can both sign; the AI Institute and the planned university are the slower, harder half of the same claim, since a country participates in that conversation only as far as its own capacity reaches. — Chris
Ethiopia (delegation) (Permanent Representative of Ethiopia): Presented Digital Ethiopia 2030's inclusive-AI foundations; the Ethiopian Artificial Intelligence Institute, the 5 Million Ethiopian Coders initiative, a planned AI University, and the FIDA digital ID as DPI; and argued the early stage of the AI era is a unique window to address disparities before they become entrenched. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisThe window you describe is real, and it is closing at the speed of procurement; disparities harden the moment systems are bought, not the moment they are noticed. — Chris
European Commission (on behalf of the EU and its member states)
Roberto Viola (Director General for Communications, Networks, Content and Technology, delivering the EU statement): The EU offers its risk-based AI Act experience (intervening only where risk is high or unacceptable) as a contribution to global governance grounded in facts and evidence, stressing the International Scientific Panel's role; announces investment in upgraded public supercomputers, AI factories and gigafactories under the new AI and Cloud Development Act, and comes seeking a durable shared commitment that AI benefits serve all humanity. Closing sessions · Day 2 · transcript ↗
from chrisDirector General, opening with the intellectual honesty to say you do not know everything is the rarest move available in a plenary, and it is what makes the risk-based offer credible; frameworks that admit uncertainty are the ones that can update. — Chris
European Data Protection Supervisor (EU)
Leonardo Cervera Navas (Panelist — Director General): Argues the unprecedented concentration of wealth in AI must be balanced by a similar concentration of state power and moral purpose; uses the controlled-fire analogy (use with care, never leave unsupervised) for the EU AI Act's risk-proportional supervision, and the civil-aviation analogy for globally uniform safety rules. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisMr. Cervera Navas, the aviation analogy works because aviation earned it: the same rules apply in Kabul and New York because every crash is investigated in public and every finding travels. AI has the uniform ambition without the incident discipline yet, and in aviation the discipline came first. — Chris
EY (Ernst & Young)
EY (Ernst & Young) (Industry representative): As a global firm auditing whether AI systems operate as they should, sees strong demand for interoperability that lives at the implementation stage via technical standards referenced in regulation — countries need not regulate identically, and industry shares society's interest in systems that functionally deliver. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
Finland
Jarno Syrjälä (transcript renders 'Jarno Sijärila') (Undersecretary of State for International Trade, Finland): Endorsed the International Scientific Panel's preliminary report and warned that high market concentration in AI chips and models in the hands of a few entities creates strategic dependencies and economic-security vulnerabilities. Invoked the Global Digital Compact's call to tackle concentrations of market power, asking how global AI governance can create a level playing field across the AI stack (was cut off by the timer). Thematic: AI opportunities · Day 1 · transcript ↗
from chrisUndersecretary, the timer cut you off at the right question; a level playing field across the AI stack will not come from antitrust alone but from open protocols that make the concentrated layers substitutable. — Chris
France
France (Member state representative): Warns that AI acceptability depends on trust and that models are already being withdrawn from market as too dangerous; cites 56 enterprises signed to Hiroshima Process principles under G7/OECD work, prioritizes protecting youth (mental health, cognitive development) and electoral integrity, and calls for collective independent evaluation of AI models before rollout. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThe observation that models are already being withdrawn as too dangerous is doing more work than the sentences around it: withdrawal means someone evaluated, decided and acted, and your call for collective independent evaluation before rollout simply asks that this judgment stop being private. — Chris
Anne Le Hénanff (transcript renders 'Anne Le Hénanf') (Delegate Minister of Artificial Intelligence and Digital Affairs): (Begins mid-block, ~4:07.) Five priorities for UN-centered AI governance: online child protection (a G7 line supported by Estonia, Brazil, Morocco, the EU), multilingualism and cultural diversity, fighting information manipulation, environmental respect (Paris summit), and open-source models to democratize access; cites France's renewable energy mix and Mistral AI. Opening plenary · Day 1 · transcript ↗
from chrisMinister, putting open-source models on the same priority list as child protection is the right pairing, since open weights are what let smaller countries verify the systems their children will actually use. My response to the Panel makes the same bet, that openness is the road to accountability rather than a detour from it. — Chris
Future of Life Institute
Jaan Tallinn (Co-founder, Future of Life Institute (concluding plenary panelist)): Warns that progress will accelerate because top AI companies are openly pursuing recursive self-improvement — using AI to accelerate AI development (citing Jack Clark of Anthropic announcing a world record in GPU-kernel development by their top model the day before) — and the autonomy that makes systems useful is exactly what makes them harder to control; the scientific panel's report concedes reliable methods for retaining control over highly autonomous systems are lacking. Because companies say they want to slow down but cannot under commercial and geopolitical pressure, the UN's role is to supply international pressure to progress at a safe pace; he dissolves the innovation-vs-safety dichotomy into two frontiers — the capability frontier where companies race into the unknown (where the safety concerns live, gated by what companies release and what host countries allow the world to access) and the adoption/diffusion frontier (where the benefits live). Also notes companies know far more about their systems' capability and safety than governments and the public, and most countries lack the technical expertise to assess frontier AI — yet are subject to the cross-border risks anyway, making the AI divide and risk governance one agenda, not two. Closing sessions · Day 2 · transcript ↗
from chrisMr. Tallinn, splitting the question into a capability frontier and an adoption frontier dissolves the innovation-versus-safety dichotomy more cleanly than anything else said on that stage; when company leaders say they want to slow down but cannot, international pressure is not interference, it is the service they requested. — Chris
Gabon
Marc Alexandre Doumba (Minister of Digital Economy and Innovation; co-chair, thematic cluster 1 (report-back)): Cluster 1 takeaways: don't race to be first on AI — race to use it wisely (the real contest is institutions and judgment, not deployment speed); most of AI's opportunity comes from adapting 'big AI to small AI' through local adaptation for contexts like an African farmer or Pacific island utility operator; and AI is a paradigm shift demanding redesigned systems with new incentives and measures of success, not optimization of existing ones. Closing sessions · Day 2 · transcript ↗
from chrisMinister, 'race to use AI wisely' names the real contest as institutions and judgment rather than deployment speed, and adapting big AI to small AI for the farmer and the island utility operator is the most transferable idea any cluster reported out. — Chris
Marc Alexandre Doumba (Minister of Digital Economy and Innovation of Gabon; thematic cluster co-chair): Argued for 'doing AI differently'; moving from big AI to smaller, locally and regionally adapted AI, because advanced economies cannot sustain the current path and developing ones cannot aspire to it; converting unstructured tacit knowledge benefits developing countries disproportionately. In closing, made jobs created, languages thriving, and communities seeing their knowledge reflected the accountability metrics, and challenged every delegation to give one priority 'a name, a date, and a budget' before leaving Geneva. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMinister Doumba, 'a name, a date, and a budget' is the sharpest accountability test issued all afternoon, and your case for smaller, locally adapted AI has arithmetic on its side: a race the wealthiest cannot sustain is not a race worth entering. — Chris
Georgia
Mikheil Kavelashvili (President of Georgia): AI is a political, inherently centralized technology; its unprecedented concentration of resources must not become totalitarian control or 'digital tyranny.' Calls for a robust international legal architecture to balance concentrated power, protect language and culture, and align AI with human values; stresses the Global Digital Compact. Opening plenary · Day 1 · transcript ↗
from chrisPresident Kavelashvili, naming the concentration itself as the danger, rather than any particular misuse of it, is the structurally honest version of the argument. Power that cannot be checked does not need bad intentions to become tyranny; it only needs time. — Chris
Germany
Karsten Wildberger (Federal Minister for Digital Transformation and Government Modernization (national statement, aligned with EU statement)): Germany's core principle is safety and security by design — built in from the start, never bolted on at the end — within a values-based, multi-stakeholder approach that seats industry, academia, civil society and underrepresented regions at the table. Announces Germany will establish a National AI Safety and Security Institute, anchored in cutting-edge research and plugged into European and international networks, to evaluate advanced AI systems and advise government. Warns that governments cannot govern the AI labor-market transition with policy recipes of the past. Closing sessions · Day 2 · transcript ↗
from chrisMinister, 'never bolted on at the end' is the phrase to hold your new Safety and Security Institute to, and your warning that the AI labor transition cannot be governed with policy recipes of the past was the least comfortable and most necessary sentence in the statement. — Chris
Global Center on AI Governance
Rachel Adams (Moderator, Panel 2 (CEO)): Frames Panel 2 as moving 'from the what to the how' — what must actually be shared across borders (evidence, measurement systems, evaluation methodologies, reporting tools, institutional practices) — while warning against a model of interoperability that simply exports assumptions, risks and priorities from powerful jurisdictions. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisDr. Adams, moving 'from the what to the how' was the right gear change, and your warning lands because interoperability without it becomes a distribution channel for other jurisdictions' assumptions; shared measurement deserves the name only if every party can contest the ruler. — Chris
Global Cities Hub
Global Cities Hub (Local/regional government advocacy representative): Argues local and regional governments are a 'still missing governance layer': public authorities directly responsible for the sectors AI is transforming, whose operational knowledge can improve the evidence base; recommends recognizing them as a distinct governance actor alongside national governments, industry, academia and civil society, systematically included in interoperability discussions. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chris'Not merely stakeholders' is the correct correction: cities run the hospitals, transit and permitting where AI incidents will actually happen, which makes them the natural source of the ground-level evidence every panel today agreed was missing. — Chris
Global Network of centers for exchange and cooperation on AI capacity building (co-facilitator)
Sheikh Hamidou Kana (Head of research group / co-facilitator of the UN-backed AI capacity-building network (floor intervention)): Reports the capacity-building network — connecting existing centers of excellence rather than creating a new institution — has in months jointly developed a cooperation framework covering skills, compute and data access, national strategy support and local innovation, and now enters its operational phase; invites member states to nominate centers of excellence and partners to contribute expertise, infrastructure and resources. Closing sessions · Day 2 · transcript ↗
from chrisConnecting existing centers of excellence rather than founding a new institution is the design choice that will let this network actually move, and the point that capacity cannot be delivered through funding alone names what money cannot buy: trusted partnership and peer learning. — Chris
Google
Yossi Matias (transcript renders 'Josi Matias') (Vice President and Head of Google Search; panelist): Acknowledged that nearly half the training data of major AI models is English while English is only 20% of spoken languages, and described Google's three-pillar response: ML breakthroughs to overcome data scarcity (1,000-language initiative, 110 new Google Translate languages including 60 African), anchoring data collection in local communities (open-sourced speech datasets in Africa and 773 districts of India), and building cultural-nuance benchmarks into model evaluation. Thematic: AI opportunities · Day 1 · transcript ↗
Guatemala
Carlos Mendoza Alvarado (Secretary of Planning and Programming of the Presidency): (Begins mid-block, ~4:19.) As a pluricultural, multilingual, multiethnic country, Guatemala argues inclusive AI cannot just expand access to technologies designed elsewhere; it must incorporate indigenous languages, respect ancestral/community knowledge, and avoid bias that renders peoples invisible; governance should combine ethics, human rights, verified science, cooperation and concrete resources. Opening plenary · Day 1 · transcript ↗
from chrisYour triad of cultural, linguistic, and territorial inclusion widens the definition in a way the access-counting crowd needed to hear. A model that renders a people invisible is not neutral infrastructure; it is a decision somebody made upstream. — Chris
Guatemala (Delegate) (Delegate of Guatemala): Argued ethics must be a cross-cutting principle across the entire AI lifecycle; design, development, implementation, use, and assessment; to build public trust and prevent risks of bias and exclusion. Emphasized Guatemala's multilingual, multiethnic context: diversity of knowledge, language, and experience broadens the perspectives from which more relevant AI solutions emerge, and committed to human-centered, inclusive AI governance through international cooperation. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisEthics across the full lifecycle only holds if the assessment stage has teeth, and your multilingual context is the proof case: bias you cannot detect in a language you never tested is bias you shipped. — Chris
Carlos Mendoza Alvarado (Secretary of Planning and Programming of Guatemala): Argued Guatemala's starting point is not models and algorithms but connectivity, data, institutional capacity, and human talent; capacity development means strengthening the state itself, treating data as a strategic asset for public decision-making, and folding digital transformation into the National Development Plan. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisMr. Secretary, starting from the state rather than the model is the harder and truer order of operations; a country that treats its own data as a strategic asset for public decisions is building the one part no vendor can sell it. — Chris
ICAIN; International Computational and AI Network
Katharina Frey (Co-Founder and Executive Director; panel 1 panelist): Argued fast-moving AI requires agile academic networks that connect demand and offer; access to large compute, quality datasets, and experts; citing ICAIN linking Data Science Africa's weather-forecasting work to Swiss supercomputing centers and a heavily oversubscribed program giving Geneva UN organizations hyperscaler-independent solutions. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisHyperscaler-independent is the load-bearing word in your intervention; a research network that can route around any single provider is practicing the kind of sovereignty the plenary only talked about. — Chris
Iceland
Halla Tomasdottir (President of Iceland (keynote, Dialogue of Dialogues)): The central question is not how we govern AI but whether we can govern ourselves in the age of AI: principles without practice inspire false comfort, and the world has not yet aligned on shared guardrails, common standards or practical accountability. Announces Iceland will convene a national dialogue this autumn with young people in a leading role, insists access without agency is not inclusion (those furthest from power must be co-authors, measured by agency and dignity rather than pilots announced), and draws a red line: never normalize delegating the irreversible decision to take a human life to a machine. Closing sessions · Day 2 · transcript ↗
from chrisMadam President, whether we can govern ourselves in the age of AI is the question beneath every other question asked this week. Measuring inclusion by agency and dignity rather than pilots announced would retire half the world's press releases, and the red line on delegating the decision to take a human life deserves to be the first commitment named in 2027. — Chris
IEEE
Anja Kaspersen (Senior Director for Technology Policy, Institute of Electrical and Electronics Engineers): Drawing on the scientific panel's report, argued agentic-AI evaluation lacks standardization and developer-defined risk thresholds operate without external verification; shared standards begin with shared terminology because jurisdictions measuring different things under one label cannot compare results; comparability precedes verification. Also pressed lifecycle-oriented governance, equitable participation in standards-setting as itself a governance issue, standardized environmental metrics, and 'protection by design' translating age-appropriate design for children into testable technical requirements. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMs. Kaspersen, 'comparability precedes verification' is the correct ordering, and most governance proposals skip it; jurisdictions measuring different things under the same label are not disagreeing, they are not yet speaking. — Chris
Independent International Scientific Panel on AI
Bilal Mateen (Member, Independent International Scientific Panel on AI; physician working on regulation of AI as a medical device; panelist): Drew the Paris-to-Glasgow climate analogy to warn AI governance is reliving that history on a hugely compressed timeline, and argued that where evidence is robust; sexual violence against women and children, interoperability failures like those that undermined the Ebola response; delay is a dereliction of duty and companies must be held to account. Where evidence is thin, standards should be built as transparency-enabling mechanisms producing comparable metrics; his lightning-round ask was national commitments to fund foundational AI research because who does the research determines what gets researched. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisDr. Mateen, splitting the map into robust-evidence zones where delay is dereliction and thin-evidence zones needing transparency mechanisms is the most operational framing the panel offered; 'who does the research determines what gets researched' should be printed on every funding call. — Chris
India
Kirti Vardhan Singh (Minister of State for External Affairs): India frames the Global Dialogue as a deciding forum, not merely a talking one: AI governance must rest on human oversight, human rights and misuse prevention (as reflected at the India AI Impact Summit), and the Global South needs capacity gaps closed before it can participate meaningfully; divisions must not let technology outrun diplomacy. Closing sessions · Day 2 · transcript ↗
from chrisMinister, 'machines do not quarrel over borders, ideology, or pride; humans do' locates the governance problem precisely where it lives, and insisting this be a deciding forum rather than a talking one is the same bar the co-chairs set for New York. — Chris
Indonesia
Meutya Viada Hafid (Minister of Communications and Digital Affairs): Trustworthy AI requires interoperability, not uniformity — one-size-fits-all is neither realistic nor equitable — and the center of global AI governance is bridging the AI divide (connectivity, data governance, financing, regulatory readiness, human capital). Cites regulation 17/2025 delaying children's accounts on high-risk platforms until age 16, with 5 million accounts deactivated in 5 months; closing principle: capacity before compliance. Closing sessions · Day 2 · transcript ↗
from chrisMinister, five million children's accounts deactivated in five months is enforcement with a number attached, which is rarer in this room than it should be; 'capacity before compliance' deserves to travel as the sequencing principle for every state asked to meet standards it was never resourced to build. — Chris
Infosys
Sayed Ahmed (Panelist — industry (GSI community)): From Infosys's failed attempt to auto-map regulations with AI: shared taxonomies are not enough because terms like 'transparency' decompose differently (explainability in ISO vs. evidences/system logs in the EU AI Act), so the community must build a controls-level meta-model and knowledge graph as the foundation for interoperability adapters. In closing, disagrees that governance should be baked into use cases from the start — governance is now dynamic, so it must be separated from use cases and applied as policy-as-code enforced at runtime from a central policy repository. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
Inter-American Development Bank
Ana Maria (IDB) (IDB representative, panelist on AI governance initiatives): Latin America and the Caribbean should leverage existing digital foundations and pursue technological sovereignty through regional coordination rather than autarky, with interoperability and standards as crucial enablers. Offers IDB blended finance and public-private partnerships to de-risk connectivity and digital infrastructure investment for underserved populations. Closing sessions · Day 2 · transcript ↗
from chrisSovereignty as coordination rather than autarky is the correction this debate needed from Latin America, and blended finance that de-risks connectivity for underserved populations is the point where a development bank's balance sheet becomes governance. — Chris
International AI Safety Report
International AI Safety Report (Lead Writer) (Lead Writer): Positions the report as complementary to the UN panel, going deep on frontier risks where evidence is sketchiest: misuse (AI lowering barriers to bio and cyber attacks — citing the Mythos cyber incident), malfunctions (increasingly agentic systems taking many low-supervision steps are harder to oversee, and keeping powerful agents under human control — correctable and stoppable — is 'an urgent and persistent problem'), and systemic impacts, the least understood. Warns the risk-management shortfall (deepfakes, sycophancy, AI-assisted cyberattacks already unpreventable) could become catastrophic as capabilities improve. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chris'We cannot manage what we do not measure' is the report's thesis, but the sentence I would elevate is the one on keeping powerful agents correctable and stoppable; that is a control problem rather than a measurement problem, and it is the one place where the evidence gap and the danger grow at the same rate. — Chris
International Electrotechnical Commission
IEC (International Electrotechnical Commission) (Standards body representative): Maps AI governance onto five normative layers — international law, international soft law, national law, international standards, and the normative power of what is technically feasible — with international standards as the link that translates principles into practical tools; adds that IEC's electricity expertise supports energy-efficient, resilient AI infrastructure. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThe five-layer map is quietly the most useful artifact of the session, especially the fifth layer, the normative power of what is technically feasible, which governs the other four whether anyone admits it or not. And 'without reliable power, there is no AI' belongs in every governance document as a standing reminder of the substrate. — Chris
International Electrotechnical Commission (IEC)
Gilles Thonet (Deputy Secretary-General, IEC; panel 1 panelist): Made the case for international standards as the most practical tool for developing countries to become AI co-creators: consensus-based standards give every country an equal seat regardless of AI industry size, participation itself builds capacity, and standards are the 'invisible infrastructure' that rests on physical energy infrastructure. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYour case for standards as the equal seat is the one I would put first; a rule made by consensus and owned by no one is the only lever that works the same in a small country's hand as in a large one's. — Chris
INTERPOL
Huang Zangfu (INTERPOL representative): Argued AI readiness among law enforcement agencies varies sharply across countries, making the AI divide a capacity-building imperative for policing; cited the INTERPOL-UNICRI responsible AI toolkit, the EU-funded AIR project, and the free TRAIL e-learning program for member-country law enforcement. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisPolicing is a hard test of this session's theme, because an agency that cannot interrogate its own AI tools is fielding evidence it cannot vouch for; the free training matters, and so does what the tools are required to disclose. — Chris
Iran
Emadeline Fatemizadeh (Head of AI Technologies Development and Application, Secretariat of the Islamic Republic of Iran): Protested the April 6 attack that damaged Sharif University of Technology's civilian scientific computing infrastructure, arguing research infrastructure is not a military target and proposing AI needs its own 'Broken Chair' symbol against unregulated technological violence. Listed Iran's AI buildout (3 million students trained, national AI platform, GPU data centers) and proposed five regional actions: an AI academy, an innovation fund, shared scientific compute, a digital free zone, and a binding legal framework. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisDr. Fatemizadeh, 'when science is attacked in one country, scientific security is weakened everywhere' states a principle worth holding whoever the parties are; shared scientific compute only works if the norm against targeting it holds first. — Chris
Iraq
Mustafa al-Maryani (Minister of Communications): Iraq calls for ethical AI governance on three pillars: keep AI away from warfare technologies and lethal weapons development, protect families and children from algorithmic harms, and invest in human capacity so humans benefit first. Closing sessions · Day 2 · transcript ↗
from chrisMinister, insisting AI remain a tool for life and not for lethal weapons development is a red line stated without hedging, and pairing it with the protection of families and children places restraint exactly where the harms are least reversible. — Chris
Ireland
Niamh Smyth (Minister of State with responsibility for Trade Promotion, Artificial Intelligence and Digital Transformation): Ireland backs a balanced, risk-based, human-rights-grounded approach (citing the EU AI Act as a workable harmonized governance framework) and calls for the dialogue to produce shared governance language on transparency, accountability and oversight, address fragmentation, and sustain progress with practical outcomes and follow-up; reiterates support for the International Scientific Panel. Closing sessions · Day 2 · transcript ↗
from chrisMinister, calling for shared governance language on transparency, accountability and oversight is quieter than proposing a new institution and probably worth more; common definitions are where interoperability either happens or silently fails. — Chris
Niamh Smyth (transcript renders 'Niamh Smith') (Minister of State, Department of Enterprise, Ireland): Warned that unregulated AI risks exacerbating divides within and between nations; presented Ireland's National Digital and AI Strategy targeting socio-economic, enterprise, and geographic divides, and championed open-source and open-weight models as widening access, local innovation, language diversity, and digital sovereignty. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisMinister, open weights widen access to models, and the same logic runs one layer down: the rules that govern what agents may do should be open too, a protocol nobody owns rather than a product someone licenses. — Chris
Italy
Alessio Butti (Secretary of State for Innovation): The question is no longer whether AI will change society but whether we govern it or are governed by it; Italy wants the UN dialogue to give substantive guidance to parallel and sectoral initiatives, built on truly universal principles each country adapts, with person-centered, security-by-design governance and algorithm/dataset transparency reflecting the world's languages. Offers Italy's concrete experience: its national AI Law aligned with the European AI Act, entrusting sector governance to two existing agencies — and insists concentration of AI in a few companies and countries is a risk for everyone, not only those left out. Closing sessions · Day 2 · transcript ↗
from chrisThe argument that concentration of AI harms everyone, not only those left out, is the version that can actually move the concentrated, and a guardrail that makes the road both safer and faster is the honest way to sell rules to people who love speed. — Chris
ITU
Tomas Lamanauskas (Deputy Secretary-General, ITU (IAWG segment, launching enhanced UN AI Resource Hub)): Announces the Interagency Working Group's new WSIS+20 mandate (resolution paragraph 86) to map AI capacity-building initiatives and fellowships, find gaps, and report back to the Global Dialogue; the first practical step is the enhanced UN AI Resource Hub, now carrying 1,000+ AI initiatives from 55 UN entities plus a capacity-building and fellowship one-stop shop. Closing sessions · Day 2 · transcript ↗
from chrisDeputy Secretary-General, a mandate to map the capacity-building landscape, find the gaps and report back to this dialogue is accountability plumbing rather than announcement, and the report-back clause in paragraph 86 is the part that gives the Resource Hub teeth. — Chris
ITU
Doreen Bogdan-Martin (Secretary-General, International Telecommunication Union; fireside chat moderator): Moderated the fireside chat with Estonia's president, steering it toward skills and education as the determinant of whether AI becomes a force for inclusion or a source of new divides. Made the ITU's framing point that AI will only succeed if it remains centered on people, and closed that AI's future will be shaped by choices; investing in people, trusted digital foundations, and cross-border cooperation; not technology alone. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisSecretary-General, 'centered on people' earns its keep only when it survives contact with a procurement decision; steering the fireside toward skills and trusted foundations rather than hardware was the right use of the chair. — Chris
Japan
Japan (Member state representative): Presents the Hiroshima AI Process as an existing best practice for the dialogue: guiding principles and a code of conduct as the first international norms for generative AI, endorsed by ~70 countries and regions in the Friends Group, with a reporting framework promoting transparency and accountability in how companies develop generative AI. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThe Hiroshima reporting framework is the part worth exporting: seventy endorsements matter less than the fact that companies now file something checkable, and trustworthiness stays the key only for as long as those filings remain auditable rather than ceremonial. — Chris
Yoichi Iida (Panelist — Advisor and Visiting Professor): Offers the OECD AI Principles and the Hiroshima AI Process (HAIP; transcribed 'HYBE'/'Hype') as foundational interoperability instruments: HAIP's code of conduct gives AI developers 'a voluntary framework for those organizations to assess and mitigate risks and also to disclose relevant information on their risk management.' Reports concrete cross-mapping work with ASEAN/Singapore showing far more commonalities than gaps — with most gaps explained by social/cultural diversity — and argues this crosswalk method is portable to other combinations of national frameworks via a bottom-up, multi-stakeholder approach that respects diversity. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisMr. Iida, the crosswalk finding deserves more attention than it got: if most gaps between the Hiroshima and ASEAN frameworks trace to social and cultural diversity rather than substance, then interoperability is an engineering problem rather than a diplomatic one, and engineering problems get solved. — Chris
Yohi Onishi (Representative of the Government of Japan): Presented Japan's trust-based international cooperation model: co-creating AI ecosystems with the Global South (ASEAN, Africa) by training not only engineers but policymakers and educators, and building foundations that let each country harness AI under its own ownership, languages, and development strategies. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisTraining the policymakers and educators alongside the engineers is the part most programs skip; ownership fails at the ministry desk long before it fails in the code. — Chris
Jason (multi-stakeholder governance panelist)
Jason (multi-stakeholder governance panelist) (Panelist on multi-stakeholder collaboration (references Global Network Initiative work; plausibly GNI's Jason Pielemeier)): AI is built on the internet's social and technical architecture, so 20-plus years of internet governance experience (WSIS, IGF, NetMundial Sao Paulo Principles) is the reservoir to draw from: effective AI governance needs inclusive process, a common normative framework grounded in international human rights law (including UNGPs and OECD guidelines), and clear priorities set by the most impacted communities. Points delegates to the 2026 IGF in Nairobi as a concrete next venue. Closing sessions · Day 2 · transcript ↗
from chrisTwenty years of internet governance treated as a reservoir rather than a preamble is the corrective this room needed, and pointing to the Sao Paulo Principles as the construction guide for every new AI commission is a specific, usable transplant rather than a lesson of history. — Chris
Kazakhstan
Zhaslan Madiyev (transcript renders 'Zaslán Madiev') (Deputy Prime Minister and Minister of Artificial Intelligence and Digital Development): Showcases Kazakhstan's push to be a fully digital nation in 3 years: Digital Code, Law on AI, new constitutional digital rights, 90%+ village broadband, Trans-Caspian fiber corridor, region's two largest supercomputer clusters, a 1 GW Data Center Valley, and AI across all education levels; offers the Almaty UN-backed regional centers as a sandbox and invites cooperation on interoperable approaches. Opening plenary · Day 1 · transcript ↗
from chrisMinister, offering the Almaty centers as a sandbox is the most useful kind of statement in this building, an invitation with an address attached. Ambition measured in fiber, gigawatts, and a three-year deadline is at least falsifiable, which is more than most pledges here can claim. — Chris
Darghan Zazubkhai (Chairman, Agency of Civil Service Affairs of Kazakhstan): Argued public trust in AI begins with trust in government, describing Kazakhstan's eKismet digital HR ecosystem (80,000+ civil servants, 400+ digital HR processes), the 2026 'year of digitalization and AI' with a new AI law and digital code, and AI-driven predictive planning and audit of government functions under transparency, human oversight, and responsibility. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisChairman, you put the dependency in the right order: trust in AI inherits from trust in government, and running 400 digital HR processes in the open makes that trust measurable in a way most trust talk never is. — Chris
Kenya
William Kabogo Gitau (transcript: 'William Gitau') (Cabinet Secretary for Information, Communications and the Digital Economy (on behalf of President William Ruto)): (Begins mid-block, ~4:16.) The divide is felt in compute, data and language; AI systems don't reflect African languages or context; Kenya wants cooperation that moves from principle to implementation: affordable compute, open representative datasets, fair financing; cites eCitizen (22,000 services, 13M Kenyans), a new interoperability framework, UNESCO/Oxford training for 20,000 public officers by 2027, and ITU school AI/robotics pilots. Opening plenary · Day 1 · transcript ↗
from chrisCabinet Secretary, bearing risks you did not create from systems you did not shape is the asymmetry underneath the entire divide debate, and you stated it without decoration. Kenya's eCitizen numbers show implementation is not the bottleneck when the demand is real. — Chris
Philip Thigo (transcript renders 'Philip Tigo') (Special Envoy on Technology, Kenya; panelist): Framed the environmental agenda as 'AI for green' plus 'green AI': practical human-centered uses (a farmer who cannot afford a failed crop cycle), embedding AI in the COP process and NDCs, and governing AI's footprint 'from mine to model'; water, minerals, land, compute, children working in mines. Backed a global environmental AI standards exchange (from the UN High-Level Advisory Body he served on) and capacity-building rather than punitive measurement, insisting the next frontier of AI must be sustainable by design. Thematic: AI opportunities · Day 1 · transcript ↗
from chris'From mine to model' is the honest scope of governance, and your preference for capacity-building over punitive measurement is right; a footprint standard nobody can implement just becomes another exam the wealthy get to grade. — Chris
Shikoh Gitau (Founder and CEO; moderator of panel 1): Argued capacity building must move beyond teaching people to use ChatGPT toward a coordinated approach to building AI 'for us, by us'; investments in skills, jobs, standards, open systems, and sustainable financing that include Africa's 2,000 languages, its communal economies, culture, and context. Thematic: bridging AI divides · Day 1 · transcript ↗
from chris'For us, by us' is the whole argument in four words; a curriculum that ends at using ChatGPT is teaching people to be customers, not builders. — Chris
Latvia
Latvia (Member state representative): Notes the trust paradox (AI detecting stroke at 93% accuracy in Latvia, yet 7 in 10 globally want stricter regulation) and reports filtering millions of coordinated-disinformation articles while building an open EU multilingual model — with no shared standard for training-data origin and no channel to warn others. Proposes shared evaluation methods and open benchmarks reflecting linguistic diversity, minimum compatibility on data origin and incident reporting (including data-integrity incidents), and capacity building so every country evaluates as an equal partner. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisFiltering millions of poisoned articles while having no channel to warn the next country is the clearest picture of missing infrastructure this session produced; a training-data origin standard is not glamorous, but neither was the postal treaty, and this is the kind of plumbing trust actually runs on. — Chris
Lawyers Hub
Linda Bonyo (Founder and CEO; delivering thematic cluster 4 report-back (also on behalf of the absent Minister of Spain)): Cluster 4 (human rights) raised child protection for non-homogeneous children, collapsing funding for civil society, missing AI-worker voices, and a pointed structural critique: 51% of AI governance conversations in the last six months happened in Geneva while Africa accounts for 7 of 10 visa refusals — demanding transparency and accountability for the AI systems used in visa decisions, and (personally) an 'AI passport' for mobility. Also calls for legal certainty, information integrity and democracy on the agenda, and for states to domesticate and ratify instruments like the new ILO convention on decent work for AI workers under algorithmic management; notes Lawyers Hub launched an AI Governance Index for Africa. Closing sessions · Day 2 · transcript ↗
from chrisMs. Bonyo, placing 51 percent of the conversations in Geneva beside seven of ten visa refusals for Africa is the sharpest structural evidence of the week, and demanding transparency for the AI systems inside those visa decisions closes a loop most speakers left open. — Chris
Lebanon
Minister of State for Technology and AI (not named in transcript; Lebanon's Minister of State for Technology and AI is Kamal Shehadi) (Minister of State for Technology and AI): Frames Lebanon's AI push as renewal after wars and crises: Numu national digital/AI capacity-building program, responsible-AI school curriculum, diaspora mentoring networks, and digital state foundations; defines digital and AI sovereignty as capability and choice, not isolation, and calls for implementable standards and safeguards that protect people without freezing innovation. Opening plenary · Day 1 · transcript ↗
from chrisMinister, 'not as isolation, but as capability and choice' is the most workable definition of digital sovereignty anyone offered all day. It is also the definition open protocols exist to serve: federation means you can leave, and being able to leave is what makes staying a choice. — Chris
Lesotho
Nthomeng Majara (transcript renders 'Enhameng Amadiara' / 'Ntumeng Majara') (Deputy Prime Minister): (Begins mid-block, ~3:16, after the Kazakhstan statement.) Aligns with G77+China and the Africa Group; four priorities: capacity building, openness (open source/data/models), interoperability and digital public infrastructure, and sector-led risk-based governance through existing regulators; proposes a framework for making fragmented African public-sector data AI-ready. Opening plenary · Day 1 · transcript ↗
from chrisDeputy Prime Minister, saying African countries lack data that is 'organized, interoperable, and prepared' rather than data itself moves the question from scarcity to stewardship, and stewardship is buildable. A common framework for AI-ready public data is the unglamorous kind of proposal that actually ships. — Chris
Kanono Lamasomol (Deputy Minister, Lesotho; chair of the AU Specialized Technical Committee on Communications and ICT experts panel; panel 1 panelist): Argued connectivity alone is not enough and identified data; fragmented, unstructured public-sector datasets; as the main barrier for developing countries, proposing this forum create a common framework for handling such data; also cited skills breadth (cybersecurity, data centers) and endorsed DPI and open source as what works. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisDeputy Minister, your numbers end the access debate in one line: 100 percent connectivity and 50 percent access means the missing half is institutional, and a common framework for fragmented public data would do more than another cable. — Chris
Libya
Mohamed Menfi (President of the Presidential Council of Libya): Speaks through a unified African position: the AI gap is a divide in jobs, capability and benefit, not just access; Africa has 1.4 billion people but under 2% of global data centers; cites five NEPAD principles on distribution, the digital divide, and data; refusing that African data be exploited as its natural resources were. Opening plenary · Day 1 · transcript ↗
from chrisPresident Menfi, 1.4 billion people and fewer than 2 percent of the world's data centers is an argument no rhetoric could improve. Refusing in advance that African data be extracted the way African resources were names the pattern before it repeats, which is the only moment naming it helps. — Chris
Lithuania
Lithuania (Member state representative): Trust rests on three pillars: human-centered AI grounded in democratic values with the human-in-the-loop principle preserved at critical decision points; security and resilience by design, drawing on Lithuania's experience with persistent cyber and hybrid threats; and interoperability anchored in the Council of Europe Framework Convention (Vilnius Convention) as a milestone. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisRefusing to let human-in-the-loop be reframed as an obstacle to innovation matters most coming from a country that learned resilience under live cyber and hybrid pressure; oversight designed by people who have actually been attacked reads differently from oversight designed in the abstract. — Chris
Lu (industry panelist)
Lu (industry panelist) (AI industry panelist (board member, CommonSpirit Health Foundation)): Governance must be technology-informed so rules are not outdated at finalization: real deployments are hybrid systems (large plus small language models, in-house plus third-party) needing a governance layer over the whole system, and technologies like federated computing can themselves deliver governance outcomes such as compliance without moving data. Calls for risk-profile-based regulation differentiated by industry and country, with standardized evaluation and transparency but practical deployment rules, plus incentives so responsible-AI adoption becomes a positive feedback loop rather than a perceived brake on innovation. Closing sessions · Day 2 · transcript ↗
Luxembourg
Luc Frieden (Prime Minister): Multilateralism is not optional for AI: advance innovation and protection together; on military AI, human judgment must never be surrendered in life-or-death decisions; common standards must not mean uniformity; AI should reflect linguistic and cultural diversity ('an AI of humanity, by humanity and for humanity'). Opening plenary · Day 1 · transcript ↗
from chrisPrime Minister, 'wherever voluntary principles fall short, we must be prepared to establish binding international rules' is the sentence this Dialogue will be measured against. Voluntary regimes fail quietly, and by the time the failure is visible the leverage is gone. — Chris
Malawi
Chadrik Namalomba (Minister of Information and Communications Technology): Malawi faces an impossible choice — accept unacceptable risks or be left behind — because it is expected to meet the same governance standards without the scale: low FDI, unreliable energy, prohibitive data and device costs. Asks for investment in three pillars (energy, computing power, human skills) and brings two young Malawian ICT-award-winning innovators as proof of potential. Closing sessions · Day 2 · transcript ↗
from chrisMinister, the highway built while traffic moves at full speed is the most honest image this session produced, and bringing two award-winning young innovators to the podium answered in advance anyone tempted to mistake Malawi's constraint for a lack of talent. — Chris
Maldives
Ali Ehsan (Minister of Homeland Security, Labour and Technology): Speaking for small island developing states, argues AI's test is whether a government on a small distant island can use it safely: invest in foundations (DPI, trusted data, cybersecurity) before applications, treat capacity building as long-term, and treat interoperability as a governance principle enabling incremental sovereign build-out. Cites South Asia's first national AI readiness assessment (with UNESCO) and the new Maldives AI Lab. Closing sessions · Day 2 · transcript ↗
from chrisMinister, 'not whether the powerful can go further, but whether the rest of us can come along' is the truest test of AI for All anyone offered, and putting foundations before applications, trusted data and cybersecurity before the demos, is the discipline most larger states skipped. — Chris
Mali
Almadou Iliane (Minister of Communication of the Digital Economy and the Modernization of the Administration (representing head of state Assimi Goita)): Mali's priorities: sovereign control of data ('AI uses data, but who controls the data?') as a matter of national dignity, recognition that AI is not neutral and non-representative data amplifies discrimination, and recommendations for an international governance framework with clear rules, a dedicated development space for LDCs, and representation of African languages and cultures in AI. Closing sessions · Day 2 · transcript ↗
from chrisMinister, 'AI uses data, but who controls the data' is a question with a checkable answer, which makes it worth more than most principles; representation of African languages is exactly where that control becomes visible or fails to. — Chris
Mary Robinson
Mary Robinson (Former President of Ireland and former UN High Commissioner for Human Rights; moderator of first panel (AI for inclusive development and social impact)): Moderated the first panel, pressing panelists on linguistic diversity, the AI divide, open models versus provider dependency, and manipulative design. Substantively endorsed the Secretary-General's challenge to AI companies on their environmental footprint, urging civil society across countries to rally around and insist on that transparency. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisPresident Robinson, urging civil society to rally around environmental transparency is reaching for exactly the right lever, because disclosure is the one demand that costs a company nothing to grant except the ability to keep hiding. — Chris
Mexican Society for Artificial Intelligence (Universidad Panamericana)
Mexican Society for Artificial Intelligence (Academic/civil society representative): Argues every AI deployment involves inequitably distributed risk-benefit trade-offs — benefits concentrated among a few while children and vulnerable populations bear disproportionate risk — and calls for a new international instrument prioritizing interoperability with legally binding minimum mandatory standards applicable across all countries, plus enforcement infrastructure. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chris'There is no use of AI systems without risk' is the honest premise most frameworks bury in a preamble; once risk is admitted as constant, the question becomes who carries it, and your observation that children and the vulnerable hold the concentrated share is the distribution problem binding minimum standards exist to fix. — Chris
Mexico
Mexico (Ambassador) (Ambassador of Mexico): Speaking in Spanish for multilingualism, welcomed the scientific panel's preliminary report and argued AI is already changing the social contract; redefining power, economic activity, and social networks; making inclusive multi-stakeholder dialogue and international cooperation paramount. Described technology as a cross-cutting pillar of Mexico's 2024-2030 National Development Strategy under its National AI Development Agency, seeking public technological solutions as a common right, and announced a high-level regional forum for Latin America and the Caribbean with UNDP at the end of the year. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisAmbassador, if AI is rewriting the social contract, the rewrite is happening in procurement offices and terms of service long before any parliament reads a draft; your regional forum is a chance to put the clauses back where citizens can see them. — Chris
Microsoft
Brad Smith (Vice Chair and President, Microsoft (keynote, Dialogue of Dialogues)): Presents Microsoft's quarterly AI-diffusion measurement: 27% of working-age population in the Global North uses AI versus 15% in the Global South, and on current trends by next year the North will use AI at more than twice the South's per-capita rate — evidence the technology is widening, not narrowing, the North-South divide absent electrification, connectivity, multilingual models and skilling. On governance: over the last 12 months AI has raced forward while governance has not kept pace (risks growing this year in cybersecurity, in future years in biology); the only way to catch up is together — inclusive of every government plus multi-stakeholder — and since countries will rightly make sovereign, different choices, the world cannot afford an unnavigable patchwork and must build interoperability among governance systems, with the UN and ITU playing an indispensable role. Closing sessions · Day 2 · transcript ↗
Moldova
Vladimir Tchoukh (Permanent Representative of the Republic of Moldova to the UN): Argued capacity gaps are not just a Global South problem; Moldova's UNESCO readiness assessment found capacity lacking across its public sector, private sector, and academia even as it transposes the EU AI Act; urged capacity-building investment at the core of AI cooperation and public-private partnerships as a concrete takeaway. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisAmbassador, thank you for breaking the frame; if Moldova finds the gap while transposing the EU AI Act, then capacity is not a Global South problem with a Northern solution, it is everyone's unfinished work. — Chris
Mozambique
Americo Muchanga (Minister of Communications and Digital Transformation): Mozambique calls for a global AI ecosystem on three pillars — inclusive governance reflecting all nations' voices, capacity development beyond algorithms (infrastructure, energy, education, local talent), and trust anchored in human dignity and international law — warning that concentrated access must not be the defining legacy of the AI era. Closing sessions · Day 2 · transcript ↗
from chrisMinister, capacity development beyond algorithms, naming energy, infrastructure, education and local talent, resists the narrowing this debate keeps sliding toward; history judging us on how wisely we governed rather than how fast we built is a standard worth writing into the outcome document. — Chris
Namibia
Emma Theofelus (panel intervention) (Minister of ICT of Namibia, concluding plenary panelist): Three-step cooperation agenda: immediately help countries translate global AI principles into national laws and institutions fitted to context; medium-term financing for digital/data infrastructure and digital literacy; long-term programs building Global South scientific capacity (Africa will hold nearly half the world's population by 2050). Defines success for New York May 2027 as a shared-but-flexible reference framework giving small states a practical sovereign template for legislation, visible capacity gains in the Global South (trained regulators, deployed national strategies), and standing partnerships where governments, industry, academia and civil society exchange evidence and co-develop open tools. Closing sessions · Day 2 · transcript ↗
from chrisMinister, your definition of success for May 2027, a shared but flexible reference framework small states can legislate from plus visible capacity gains in the Global South, is concrete enough to be checked when the time comes; that is what separates a commitment from a hope. — Chris
Emma Theofelus (Minister of Information and Communication Technology (governmental plenary statement)): Namibia's concern is not AI adoption alone but the foundational digital infrastructure that makes AI possible; proposes expanded open datasets, shared AI infrastructure including regional AI hubs and possibly an AI capacity fund for developing states, and responsible open-source models — and flags AI-facilitated gender-based violence as a threat to Namibia's gains in women's empowerment. Closing sessions · Day 2 · transcript ↗
from chrisMinister, flagging AI-facilitated gender-based violence names a harm most statements left abstract, and pairing regional AI hubs with expanded open datasets is a foundations answer to a divide others described and few specified. — Chris
Namibia (delegation) (Minister of ICT of Namibia): Named three urgent needs: unlocking financing (noting the $800 billion in AI spending this year will not flow to developing countries), North-South university twinning agreements to train Global South youth, and access via lean models, open standards, and accountability for mega-tech companies including on tech-facilitated gender-based violence. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisMinister, your electricity analogy carries a warning inside it: the grid became a background utility only after decades of hard rules about who could be harmed by it, and AI deserves the same insistence, especially on the violence you named. — Chris
Nepal
Ram Prasad Subedi (Permanent Representative of Nepal in Geneva): Argued the AI divide is about the capacity to influence; infrastructure, talent, and governance capability; with gaps most pronounced in Nepal's rural and remote areas as a landlocked graduating LDC; endorsed the Secretary-General's proposed Global Fund on AI as a pooling mechanism attentive to LDC and LLDC contexts. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisAmbassador, 'capacity to influence' is the honest name for the divide; influence is exactly what an open rule can offer a landlocked LDC, because a protocol nobody owns cannot weight its votes by compute. — Chris
Netherlands
Billemijn Aerts (Minister for the Digital Economy and Sovereignty): Names interoperability 'the organizing principle for this next phase': not one rulebook for the world, but frameworks in different jurisdictions built on common international standards for rights, safety, transparency and accountability that can recognize and work with each other instead of fragmenting into unnavigable patchworks. Backs this with transparency in practice — the Dutch public algorithm register documents ~1,500 algorithms across nearly 350 public-sector bodies — and calls for building Global South participation in from the start. Closing sessions · Day 2 · transcript ↗
from chrisMinister, a public register of some 1,500 algorithms across nearly 350 bodies is transparency as practice rather than posture, and defining interoperability as frameworks that recognize each other rather than one rulebook for the world is the most precise use of the week's most-used word. — Chris
Oman
Oman (Undersecretary for Communications and Information Technology) (Undersecretary for Communications and Information Technology, Ministry of Transport, Communications and Information Technology of Oman): Argued the cultural-linguistic dimension is decisive and that every country is the best expert in its own culture, language and values; foreign companies cannot know them better, and local (often young) talent already has the tools to build local solutions, as Oman has done. Urged the cluster to embrace AI's opportunities with confidence and shape it to reflect the planet's diversity. Thematic: AI opportunities · Day 1 · transcript ↗
from chris'You cannot expect a foreign company to know your culture' is the case for local ownership stated plainly; the confidence you urged is warranted precisely because the expertise you named cannot be imported. — Chris
Oman (delegation) (Undersecretary, Ministry of Transport, Communications and Information Technology): Laid out three foundations for bridging the divide: talent (National AI Program initiatives in AI product management, data science, chip design), infrastructure (a national 'Digital Triangle' of multi-gigawatt green AI and HPC hubs leveraging Oman's neutrality and submarine cables), and international cooperation, offering Oman as a partner and bridge. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYour one-liner does a lot of work: use or build, not use or buy, and the distance between those two words is the whole divide this session was convened to name. — Chris
Pakistan
Shaza Fatima Khawaja (Minister of Information Technology and Telecommunication): Inclusion means a voice in writing the rules — norms, standards and interoperability frameworks authored by climbers as well as summit-holders — and inclusion without investment is an empty promise: Pakistan calls for open-source model access, shared compute, data-infrastructure support, regulatory sandboxes fit for local conditions, and advancing the proposed Global Fund for AI. Pakistan's national policy codifies human-in-the-loop, human-on-the-loop and human-in-command, with a data governance policy treating data as a national asset held in trust ('sovereignty by default'). Closing sessions · Day 2 · transcript ↗
from chrisMinister, 'inclusion without investment is an empty promise' names the gap between the communique and the budget, and codifying human-in-the-loop, on-the-loop and in-command in national policy turns a slogan into an assignable duty. — Chris
Pakistan (Member state representative): Three points: safety must be defined inclusively under UN multilateral processes, not exported by a few states and firms via national safety institutes; interoperability must be a bridge between systems, not a mechanism making the rules of a few binding on the many; and trust requires the capacity to verify — evaluation capacity is concentrated where the technology is built, so developing countries risk accepting safety assurances they cannot check. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chris'A nation cannot trust what it cannot assess' compresses the whole session into eight words; while evaluation capacity stays concentrated where the technology is built, safety assurance flows one way, and assurance that cannot be checked is not assurance, it is dependence. — Chris
Pakistan (delegation) (Minister of ICT of Pakistan): Warned the multipolar world is becoming 'technopolar'; power concentrating around those who control frontier compute, models, and data; and that without consequential decisions the trajectory is two parallel AI ecosystems and a permanent two-tier order of rule-takers; proposed a polycentric architecture of co-ownership with emerging economies at the design table and open standards as the strategic equalizer. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisMinister, your polycentric architecture is the same shape I arrived at from the engineering side: independently owned zones interoperating through open standards, with no center left standing to do the deciding while others receive. — Chris
Partnership on AI
Rebecca Finlay (Co-Chair (CEO, Partnership on AI)): Proposes three actions: strengthen the independent scientific evidence base (connect the UN panel to the UK/Singapore/Global South safety reports and the International Network of AI Safety and Security Institutes), open up the science via company disclosure plus a coherent international assurance ecosystem of benchmarks, standards and third-party evaluation and verification, and track progress against an inclusively defined baseline (announcing PAI's Global AI Progress Hub). Her closing summary distills five takeaways, including that 'trust must be built through verifiability that is fostered through independent third-party testing, disclosure across the AI value chain, including incident reporting,' and that interoperability is necessary but not sufficient for safety. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisMs. Finlay, swapping 'how powerful is this system' for 'how verifiable is its safety, and against what evidence' is the question substitution this whole process needed; power is the developer's metric, verifiability is everyone else's. — Chris
Peru
Diego Belevan (Delegate of Peru (addressed as 'Director')): Argued the problem is not absent shared values but the impossibility of operationalizing them without real capacity; cooperation must pair the normative layer with concrete access to infrastructure and technical assistance; open source helps only if accompanied by institutional capacity, otherwise the access gap becomes a use gap; public-sector training has a multiplier effect. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYou caught the flaw in the open-source consolation prize: weights without institutions turn the access gap into a use gap, and shared values stay decorative until someone funds the machinery that operationalizes them. — Chris
Philippines
Philippines (Delegate) (Delegate of the Philippines): Argued AI's risks fall unevenly; on workers, children, persons with disabilities, Indigenous peoples, creators, and small enterprises; so governance must stay people-centered, with lifelong reskilling pathways for workers and MSME adoption support. Made a distinctive creative-sector demand: recognition, consent, attribution, and fair remuneration, insisting cultural expressions and traditional knowledge not be treated as raw material, and that developing countries must be able to build, adapt, evaluate, and govern systems reflecting their own languages and institutions. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisRecognition, consent, attribution, and fair remuneration are not creative-sector perks; they are the ledger entries of provenance, and a system that cannot say whose work it stands on cannot honestly ask to be trusted. — Chris
Sara Maria (Delegate of the Philippines): Argued the AI divide reflects unequal starting conditions rather than lack of ambition; differences in connectivity, compute, data, expertise, financing, and institutional readiness; and that inclusion must reach beyond urban centers to women, youth, persons with disabilities, and indigenous peoples, going beyond access to genuine agency. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYour four verbs are the real standard: use, question, improve, and govern, and the middle two are the ones access alone never grants, because questioning a system requires the right to look inside it. — Chris
Poland
Krzysztof Gawkowski (transcript renders 'Kawkowski') (Deputy Prime Minister and Minister of Digital Affairs): Poland treats AI as a practical policy tool: deployment in healthcare administration, high-quality protected public data to let smaller actors participate, digital education for job-market change, and standards/transparency against disinformation; with human-centric, safe AI and international cooperation as the frame. Opening plenary · Day 1 · transcript ↗
from chrisDeputy Prime Minister, high-quality protected public data as the way to let smaller actors participate is a substrate move, not a subsidy, and the substrate move is the durable one. Whoever holds the data commons sets the terms of entry for everyone after. — Chris
Portugal
Gonçalo Matias (transcript renders 'Gonzalo Matías') (Minister in the Cabinet of the Prime Minister and of State Reform): The real challenge is governing AI's implementation: Portugal's national AI agenda is an implementation agenda; the 'Agentic State,' digital twins of public administration, an Iberian AI gigafactory project with Spain, and Amália, the first open LLM built for Portuguese; coordination among UN, OECD and EU per their mandates. Opening plenary · Day 1 · transcript ↗
from chrisMinister, 'governments should not simply regulate AI, they should become responsible adopters' closes a credibility gap most regulators pretend not to have; you cannot govern what you have never operated. Amália is the receipt that makes the argument stick. — Chris
Qatar
Mohammed bin Ali bin Mohammed Al Mannai (Minister of Communications and Information Technology): Three objectives: bridge divides in compute, data and skills (including open-source models in Arabic); develop safe, secure, interoperable models since insecure models cross borders; and respect human rights and international law. Announces Qatar will launch a Global Initiative on AI; a global coalition; under its leadership. Opening plenary · Day 1 · transcript ↗
from chrisMinister, your point that insecure models cross borders is the one I would print in bold: safety failures federate whether or not governance does. Interoperability is not a convenience feature in that world, it is the containment strategy. — Chris
Rebecca (thematic cluster 3 co-chair)
Rebecca (thematic cluster 3 co-chair) (Co-chair, thematic cluster 3 on safe, secure and trustworthy AI (report-back)): Delivers the dialogue's most operational governance list: AI governance must become adaptive because static rules cannot handle increasingly autonomous, interconnected, multi-step systems; interoperability must become practical via common definitions, shared technical standards and clear links between legal requirements, technical controls and evidence of compliance; and cooperation must move to implementation through cross-border regulatory sandboxes, shared evaluation methods and incident-reporting mechanisms so countries can test systems, learn from failures and prevent repeated harms. Adds that evaluation frameworks must reflect linguistic and demographic diversity, and — on agentic AI — that as systems gain autonomy, responsibility must remain clearly assigned across developers, deployers and operators with traceability and meaningful human oversight so autonomy never becomes an accountability gap. Closing sessions · Day 2 · transcript ↗
from chrisMs. Finlay, the chain linking legal requirements to technical controls to evidence of compliance is the most operational sentence any cluster reported out, and 'autonomy never becomes an accountability gap' states the agentic problem as a design requirement rather than a worry. — Chris
Republic of the Congo
Minister of Foreign Affairs, La Francophonie and Congolese Abroad (transcript garbles the name as 'Konstantin Sergej Wunder'; the title matches Jean-Claude Gakosso) (Minister of Foreign Affairs, La Francophonie and Congolese Abroad (mandated by President Denis Sassou Nguesso)): AI can accelerate the SDGs and AU Agenda but raises ethics, security, data-protection, disinformation and exclusion challenges; Congo has created a dedicated digital-economy ministry and an African AI research center at university level to train African researchers and innovators (statement cut off by the timer). Opening plenary · Day 1 · transcript ↗
from chrisMinister, an African AI research center training African researchers is worth more than the paragraph the timer allowed you, because capacity compounds and speeches do not. The convergence of innovation and human values you asked for happens inside institutions like that one, not inside communiqués. — Chris
Romania
Romanian National Authority for Management and Regulation in Communications (ANCOM) (Vice President, national communications regulator (floor intervention)): Invoking Titulescu ('peace should not only be asserted, it should be organized'), Romania argues pro-innovation and responsibility are the same duty seen from two sides: a deliberate human-centric transition architecture (retraining, pathways to new jobs), restraint where stakes are gravest — welcoming the bilateral dialogue underway between leading AI powers and insisting human judgment remain sovereign where AI's destructive potential is greatest — and child protection above all. Closing sessions · Day 2 · transcript ↗
from chrisInvoking Titulescu, that peace must be organized rather than merely asserted, names precisely the gap between principles and mechanisms this dialogue kept circling, and restraint as a matter of principle where AI's destructive potential is greatest belongs in the outcome document verbatim. — Chris
Russia
Russian Federation (Deputy Minister) (Deputy Minister of Digital Development, Communications and Mass Media (floor intervention)): Access to frontier AI is not universal because key players (model developers, cloud providers, chipmakers) apply restrictions; Russia claims a full domestic technical chain without foreign components (GigaChat, YandexGPT) plus models trained on national values, and offers partnership to countries wanting AI respecting their cultural codes — pledging, 'as a decision of principle,' not to scrape or exploit partners' data. Closing sessions · Day 2 · transcript ↗
from chrisA pledge of principle not to scrape or exploit partners' data carries exactly as much weight as the partners' ability to verify it, and verification requires a pipeline inspectable by someone other than its owner. — Chris
Russian Federation (Deputy Minister of Foreign Affairs) (Deputy Minister of Foreign Affairs of the Russian Federation): Welcomed the Global Dialogue as a move from 'semi-closed consultations among a narrow group of states and their monopolies' to full multilateralism, called for building on India's New Delhi AI Summit, and insisted AI developers comply with the laws of countries where deployed. Framed regulation around UN Charter principles of sovereign equality and non-interference, cast trust-building as the safeguard against 'digital neocolonialism,' and offered Soviet-era machine-learning heritage plus Russian energy and expertise for sovereign AI development. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisDependence on systems a country cannot inspect is a real condition whoever names it; the only remedy that does not swap one dependence for another is a rule anyone can verify and no one owns. — Chris
Arutyun Avetisyan (transcript renders 'Arutyun Abestiani') (Institute Director; panel 1 panelist): Argued trust in shared AI infrastructure takes years to build; proposed fixing at UN level that open models and open code be equally accessible to any student or university worldwide, with standardized institutions for security cooperation and content labeling; while insisting single standards must still accommodate each nation's cultural specificities. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisA single standard with room for every culture is a layering question, not a contradiction; the layer that must be uniform is the one that carries verification. — Chris
Rwanda
Paula Ingabire (Minister of Information and Communication Technology and Innovation): Fragmentation slows Africa's builders down when every market has a different rulebook, so the continent is harmonizing: 49 African countries plus the AU and Smart Africa adopted the Africa Declaration on AI in Kigali; Rwanda pairs a national AI policy, data protection law and a newly established national AI agency with skills training. Closing sessions · Day 2 · transcript ↗
from chrisMinister, forty-nine countries plus the AU adopting one declaration in Kigali is harmonization delivered rather than called for; builders feel fragmentation as friction at the border long before diplomats name it in plenary, and your statement spoke for the builders. — Chris
Saudi Arabia
Abdullah Alswaha (transcript garbles the name as 'Abdullah Al-Zawahiri') (Minister of Communications and Information Technology): Opens with a Quranic quote on mutual service; warns that ~90% of AI compute and IP is held by just two countries and asks whether the UN accepts that future for 190+ member states; positions Saudi Arabia as a partner to close the gap (6 GW compute buildout vs ~80 GW of global need) and calls for declaring AI an existential necessity for people, planet and prosperity. Opening plenary · Day 1 · transcript ↗
from chrisMinister, asking whether the UN accepts a two-country future for more than 190 member states puts the right question to the right body, and the 90 percent figure gives it teeth. Six gigawatts against eighty of global need also concedes, usefully, that no single builder closes this gap alone. — Chris
Scientific Panel
Yoshua Bengio (Co-Chair, Independent International Scientific Panel on AI): Presents the panel's mission and stance: technical progress is doubling on some metrics every few months with no sign of slowdown, there are still no known technical guarantees AI will follow human instructions or laws, frontier models already show deception and evaluation-awareness, and concentrated commercial/geopolitical interests are dictating AI's direction without adequate societal or technical guardrails. In his wrap at [2:30:47] he adds that continued growth in AI intelligence 'could change the power dynamics of our planet' and hands the report to the Dialogue: 'Please act.' Opening plenary · Day 1 · transcript ↗
from chrisDr. Bengio, 'we still have agency' is the sentence I keep returning to, because it locates the failure mode in our choices rather than in the machines. My response to the Panel takes your 'please act' literally: since there are no technical guarantees these systems follow instructions, the acting layer is authorization, bounding what they are permitted to do before we have perfected what they intend. — Chris
Scientific Panel
Maria Ressa (Co-Chair, Independent International Scientific Panel on AI): Speaking after Bengio in the same segment (~1:45): the report is the consensus floor of 40 independent experts, 'the floor of our concern, not the ceiling.' Gives three human-scale findings; life-threatening Tigrinya medical mistranslations, a frontier model finding decades-old flaws in OpenBSD/FFmpeg (dual-use), and a 14-year-old's death after chatbot conversations; and warns most countries cannot yet test, audit or govern these systems on their own terms. Opening plenary · Day 1 · transcript ↗
from chrisMs. Ressa, 'the floor of our concern, not the ceiling' is how consensus documents should always be read and almost never are. Your warning not to wait for certainty is the operational one; evidence standards built for peacetime do not survive a compressed timeline. — Chris
Senegal
Samba Diouf (cluster 2 report-back) (Minister of Telecommunications and Digital of Senegal; co-chair, thematic cluster 2 (statement 107, attributed in transcript to unidentified but introduced/thanked as Diouf)): The divide applies beyond AI — to infrastructure, connectivity, devices and knowledge; for West Africa, closing the AI gap first requires fixing connectivity and infrastructure, then knowledge-sharing together ('we share the platform together'). Closing sessions · Day 2 · transcript ↗
from chrisMinister, insisting the divide runs beyond AI, through connectivity, devices and knowledge, keeps the diagnosis honest; 'we share the platform together' was the right cooperative note for the cluster to end on. — Chris
Samba Diouf (Minister of Telecommunications and Digital, speaking for the Global Network for Exchange and Cooperation on AI Capacity Building): Speaking on behalf of the capacity-building network (launched at UNGA 2025 at the initiative of Saudi Arabia and Kenya, with centers across ~16 countries), endorses the Secretary-General's proposed global fund for AI, donor-coordination platform, in-kind contribution mechanisms and the capacity network, and invites states, academia, philanthropy and the private sector to back it with resources and centers of excellence. Closing sessions · Day 2 · transcript ↗
from chrisMinister, a capacity network stitched from existing centers across sixteen countries asks members for contributions rather than permission, and that design choice may let it move faster than the larger funds it endorses. — Chris
Samba Diouf (Minister of Communications, Telecommunications, and Digital of Senegal; session co-chair): Argued AI's opportunities are only realized with unified access, citing Senegal's New Deal technology pillars (skills, AI, infrastructure, public platforms); in closing remarks added that small countries lack the demography and investment to build AI alone and must pursue regional strategies for data governance. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisMinister, the regional move is right, and it carries a design question worth settling early: whether the shared data layer is owned jointly by its members or administered for them, because those two futures diverge fast. — Chris
Shutterstock / Women in AI
Alessandra Sala (Senior Director at Shutterstock and Global President of Women in AI; panel 2 panelist): Argued against the time-saving/replacement framing of workplace AI, using her office-hours example of reinventing a salesperson's workflow to reach 500 instead of 50 customers: leaders and governments should deliberately design the augmented workforce they want rather than follow replacement narratives. Thematic: bridging AI divides · Day 1 · transcript ↗
Singapore
Josephine Teo (Minister for Digital Development and Information): Singapore's three governance lessons: build the blocks (reliable infrastructure including the software layer, quality data — via ASEAN's Digital Economy Framework Agreement for interoperable cross-border data-flow standards — and capable people); build trust, because people will not use technology they do not trust — Singapore has published guidelines on responsible data use for AI models and on AI agents, updates laws for AI risks, and partners ISO/IEC to standardize generative-AI testing methodology; and treat AI-for-good as patient gardening, not big-bang announcements. Notes Singapore hosted the international scientific exchange that produced the Singapore Consensus for AI Safety Research Priorities, and casts the Global Dialogue as a community of 'AI gardeners' growing each member state's capacity. Closing sessions · Day 2 · transcript ↗
from chrisMinister, the garden metaphor is more rigorous than it sounds, because gardens run on maintenance schedules rather than declarations of completion; publishing guidelines on AI agents and standardizing generative-AI testing with ISO is exactly the weeding the metaphor commits you to. — Chris
Slovenia
Slovenia (Counselor) (Counselor to the UN Office and Other International Organizations in Geneva, Slovenia): Presented AI as a general-purpose technology whose benefits must be widely shared, emphasizing as a small language community that multilingualism and cultural diversity should remain strengths in the digital age and no language should be left behind. Highlighted Slovenia's investment in national AI infrastructure; a Slovenian AI Factory and AI Competence Center; plus SME adoption and research-industry-public sector cooperation. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisCounselor, 'no language is left behind' will be decided by whether small-community corpora get built and maintained, which is unglamorous infrastructure work; Slovenia funding its own AI Factory instead of only asking others to act is the credible part. — Chris
Smart Africa
Lacina Koné (transcript renders 'Lassina Kone') (Director General and CEO, Smart Africa; panelist): Argued the real question is no longer who has access to AI but who has capacity to shape it: access alone creates neither prosperity nor jobs, and Africa's greatest risk is remaining a spectator while others capture the value. Called for a systems approach built on digital public infrastructure, trusted data, and governance, applying AI to agriculture, health and education rather than frontier-model races; measured by jobs created and citizens lifted into the digital economy, citing Africa's $1.4 trillion mobile-money leapfrog. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMr. Koné, the $1.4 trillion mobile-money leapfrog is the receipt that makes 'who has the capacity to shape it' more than a slogan; measuring by jobs and citizens brought into the digital economy rather than models deployed is the scoreboard I would keep. — Chris
Somalia
Mohamed Adam Mualim (Minister of Communication and Technology): Somalia — recognized by ITU as the world's leading performer in ICT regulatory reform — argues AI is unlike telecoms: it cannot be governed by any single country or regional bloc, and for developing countries the challenge starts earlier, with compute, energy, data and skills constraints no country can overcome alone. Closing sessions · Day 2 · transcript ↗
from chrisMinister, observing that for developing countries the challenge starts earlier, at compute, energy, data and skills, reframes governance as a supply question, and coming from the world's leading performer in ICT regulatory reform it cannot be waved off as an excuse. — Chris
South Africa
South Africa (Member state representative): Approaches AI governance through a developmental, human-centric lens: AI has both improved safety and increased complexity, governance is fragmented with digital-sovereignty pursuits risking regulatory silos, and global frameworks (Hiroshima, OECD, EU AI Act) must reflect developing-country realities through inclusive design, equitable application, technology transfer and capacity building. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThat the pace of advancement has far outstripped policy cycles is exactly why the developmental lens matters: a policy cycle cannot outrun the technology, but a shared measurement can travel with it, and equitable application is the test of whether it does. — Chris
South Africa (delegation) (Deputy Minister of Communications and Digital Technologies): Set out three priorities; digital infrastructure (connectivity, electricity, compute as essential enablers), human capacity development (developers and policymakers, not just users), and access to data, compute, and financing; warning that unaddressed inequalities will concentrate AI innovation in a few countries and corporations. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisDeputy Minister, naming policymakers alongside developers in the capacity line matters; a country that trains only its coders still imports its judgment, and judgment is the harder dependency to break. — Chris
Vukosi Marivate (Professor and Scientific Panel member; Deep Learning Indaba co-founder; panel 2 panelist): Argued Africa's grassroots communities (Data Science Africa, Deep Learning Indaba) did the heavy lifting of AI capacity building and policymakers should connect with them while researchers enter policymaking; called for planning brain circulation rather than brain drain and for private-sector R&D investment (0.5-1% of after-tax profits) so benefits accrue locally; his definition of sovereign AI. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisProfessor, your history is the corrective this room needed: Indaba and Data Science Africa built the capacity everyone now wants to fund, so the question is not how to start but how to stop routing around what already works. — Chris
South Korea
Republic of Korea (Representative) (Representative / Permanent Representative of the Republic of Korea): Offered three observations: the gap between AI's speed and societal readiness (felt firsthand when Korea brought AI tools into classrooms and found rollout outpaced readiness), the emergence of previously unknown harms such as sycophancy that demand preemptive response, and the fragmentation of national governance efforts. Announced Korea's vision of establishing a global AI hub with nine UN and related organizations to strengthen global AI capacity. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisNaming sycophancy as a previously unknown harm is more useful than another decade of general bias talk, because it is a failure that flattery hides from the very people evaluating it; preemptive response starts with tests the vendor did not write. — Chris
Spain
Óscar López Águeda (Minister for Digital Transformation and Civil Service): (Begins mid-block, ~3:53.) Spain hosted the scientific panel in Madrid and co-facilitated (with Costa Rica) the process creating the panel and the Dialogue; cites the EU AI Act (recently amended to prohibit sexual deepfakes and protect minors), Spain's first-in-Europe AI supervisory agency and digital-rights charter; frames the stakes as defending human rights against 'techno-fascism.' Opening plenary · Day 1 · transcript ↗
from chrisMinister, 'techno-fascism' is a heavy word, and a first-in-Europe supervisory agency is what keeps it from being only a word. Naming the enemy is cheap; staffing the counterweight is not. — Chris
Sri Lanka
Sri Lanka (Special Envoy of the President) (Special Envoy of the President of Sri Lanka): Argued inclusion will depend not on access but on whether AI understands the people it serves: language parity is necessary but insufficient without culturally inclusive AI that recognizes local value systems, institutions, and farming and legal practices; leading to sovereign AI that reflects national priorities without compromising independence. Described Sri Lanka's architecture placing AI-powered 'language equalizers' (language models, speech recognition, OCR, translation) as horizontal digital public infrastructure feeding multilingual government and farmer advisory services. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisPlacing language equalizers in the infrastructure layer rather than the app layer is the right architecture; 'whether AI understands the people it is intended to serve' is a testable claim, and horizontal DPI is what makes the test repeatable across every ministry. — Chris
Switzerland
Albert Rösti (transcript renders 'Albert Rusty') (Federal Councillor, Head of the Federal Department of Environment, Transport, Energy and Communications; also speaking for the Freedom Online Coalition): (Begins mid-block, ~4:05.) The UN should not only set AI governance rules but exemplify them in practice; on behalf of the Freedom Online Coalition (Swiss chair), AI governance must be anchored in international human rights law and genuine multi-stakeholder processes; announces Switzerland will host the next Global AI Summit in Geneva from 21 June 2027, following India and France. Opening plenary · Day 1 · transcript ↗
from chrisFederal Councillor, holding the UN to being 'an example in their practical application' of its own AI rules is a standard multilateral bodies rarely invite upon themselves. Institutions earn trust the way people do, through observed behavior accumulating over time, not through stated intentions. — Chris
Tanzania
Angela Kairuki (Minister of Communication and Information Technology (on behalf of President Samia Suluhu Hassan)): AI as strategic enabler of Tanzania Development Vision 2050: national AI-in-education guidelines, a health-sector AI policy framework, expanded broadband backbone and national data infrastructure; works with ITU, AU, EAC and SADC on cross-border interoperability; calls for enhanced cooperation on capacity, technology transfer and infrastructure investment. Opening plenary · Day 1 · transcript ↗
from chrisMinister, an equal voice in standards-setting is worth more than most aid packages, because standards decide who gets to build and who only gets to buy. Tanzania's interoperability work with the EAC and SADC is that principle already in motion, not waiting on permission. — Chris
Thailand
Chai Chanok Chitchope (Minister of Digital Economy and Society): The greatest risk to trust is fragmentation of standards and governance approaches, and Thailand makes a concrete offer: use Thailand as a live sandbox for AI governance where global principles are translated into practical implementations in real time, solutions are tested together, and scalable governance models are developed. Cites its AI Governance Practice Center (AIGPC), hosting the 2025 UNESCO Ethics of AI Forum, and chairing the ASEAN expanded GenAI guide as evidence of readiness to be a trusted convener so emerging AI economies are architects, not passive adopters. Closing sessions · Day 2 · transcript ↗
from chrisMinister, offering Thailand as a live sandbox where global principles get translated into practice in real time is one of the few statements this session that risks something concrete; if fragmenting governance fragments trust itself, as you argue, a shared testing ground is the direct remedy. — Chris
Tsinghua University
Lan Xue (Professor and Dean, Tsinghua University; panelist): Made three proposals: protect the open global AI ecosystem where open and closed frontier models compete fairly; noting Chinese open-source models now account for over 40% of global open-source model downloads, power ~30% of global AI usage, and support digital sovereignty and local languages; cooperate on capacity building under the 2024 UN resolution; and jointly build guardrails against AI risks. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisProfessor Xue, the forty percent download figure shows open weights are already a multipolar fact; the harder question inside your first proposal is what 'compete fairly' means when open and closed models alike are graded on benchmarks nobody independently verifies. — Chris
Uganda
Alione Yorke Odria (Minister of State for Innovation, Communications Technology and National Guidance (representing President Museveni; aligned with African Group statement)): Uganda's Vision 2040 goal of a $500 billion economy cannot be reached on Africa's less-than-1% share of world AI compute; calls for concrete partnerships on compute access, capacity building and development financing, and favors interoperable standards-based approaches over fragmented ones. Closing sessions · Day 2 · transcript ↗
from chrisMinister, stating the arithmetic plainly, a 500 billion dollar goal resting on less than one percent of the world's compute, does more work than pages of principle; a gap that concrete can actually be negotiated over. — Chris
Umeå University
Virginia Dignum (Moderator, Panel 1 (Professor)): Frames the panel with the argument that 'AI is not an inevitability that we must simply adapt to. It is a set of choices—technical, institutional, and political,' with interoperability among the clearest of those choices. Warns that fragmentation creates gaps that harm the least resourced first, and steers the panel through existing building blocks, the practical costs of fragmentation, and why interoperability must be built in from the start rather than treated as an afterthought. (Merged: 'Virginia Dignam' moderator file is the same person.) Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisProfessor Dignum, 'accountability that travels with the technology' is the phrase I hope survives into the summary; accountability that stops at a border is liability arbitrage with better manners. And opening with AI as a set of choices rather than an inevitability set the terms for everything useful that followed. — Chris
UN Foundation
Claire Melamed (Vice President, UN Foundation; moderator of the Dialogue of Dialogues plenary panel): As moderator she announces the second Global Dialogue will be held in New York next May, launches the panel's forward-looking framing ('what would success look like by May 2027?'), highlights the newly launched Partnerships Hub of practical multi-stakeholder coalitions, and synthesizes the panel into five words — optimism, progress, equity, urgency, understanding — plus the recurring theme of 'interoperability of approaches.' Closing sessions · Day 2 · transcript ↗
from chrisDistilling four speakers into five words is a moderator's craft, but 'interoperability of approaches' is the phrase doing the real work; whether the new Partnerships Hub turns those five words into an agenda rather than a mood is the thing to watch before New York. — Chris
UN Global Dialogue on AI Governance (co-chairs, Estonia and El Salvador)
Rein Tammsaar (Estonia), Global Dialogue co-chair (closing ceremony; recorded under 'Unidentified', identified by his thanks to 'my fellow co-chair Ambassador López') (Co-chairs delivering the closing plenary remarks (recorded under 'Unidentified' in the transcript)): The closing stocktake: the dialogue established a new global platform — 4,200+ registered participants from 1,800 entities representing nearly 170 member states — that began with the Scientific Panel's evidence (capabilities advancing faster than the ability to verify and govern them) and found across all four clusters that the gap is not one of principles but of practical mechanisms. Announces the road ahead: a co-chairs' summary reflecting convergence and diversity, continuation of the Global Dialogue in 2027 — the second dialogue in New York in May 2027 — with the challenge to arrive 'not with restated principles, but with named, supported, and result-oriented commitments,' and a closing measure of success: not who uses AI first, but who uses it wisely — jobs created, languages served, communities included, children protected. Closing sessions · Day 2 · transcript ↗
from chrisCo-chairs, closing with a challenge rather than a celebration was the right instinct, and 'named, supported, and result-oriented commitments' is a bar I intend to meet in New York rather than restate. — Chris
UN Independent International Scientific Panel on AI
Melat Bilciu-Demircos (Panelist — Panel member): Presents the panel report's four structural findings: the evidence dilemma (capability doubles every few months, outpacing evidence and governance), ~40 documented governance models that are fragmented, inconsistent, rarely tested — and when tested, tested by the companies developing them; the need to shift from evaluating models to evaluating whole systems (tools, environment, users); and missing macroeconomic/labor evidence. In the closing round (transcribed as 'Birgit'), adds that developing nations bear the steepest fragmentation costs since compute is concentrated in two countries, and flags AI-generated CSAM, deepfake sexual violence, erosion of shared reality, and sycophantic-chatbot mental-health harms. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThe finding that governance models are rarely tested, and when tested are tested by the companies that built them, is the quiet scandal of the current evidence base; your shift from evaluating models to evaluating whole systems, tools and users included, is where independent testing becomes possible at all. — Chris
UN Independent International Scientific Panel on AI / Microsoft AI for Good Lab, Nairobi
Abebe (likely Girmaw Abebe Tadesse) (Scientific Panel member and researcher; panel 1 panelist): Argued 'context decides outcome': with 7,000 languages, no single AI model can serve all contexts, so AI must be purpose-driven and locally aware, designed with government, private sector, youth, and grassroots communities so people are at the center of the design process; quoted Wangari Maathai on doing the right thing without context being off track. Thematic: bridging AI divides · Day 1 · transcript ↗
from chris'Context decides outcome' is the sentence I would make every procurement officer sign, and the corollary you drew is the hard part: the design table has to sit where the context lives. — Chris
UN Permanent Forum on Indigenous Issues
Walt-Enos Trist (UN Permanent Forum representative): Argued only 1,000 of 7,000 languages have the foundations for meaningful AI inclusion, the performance gap for underrepresented languages is not narrowing, and building AI for such communities is methodologically distinct; so the fix requires communities at the table, dedicated research and resourcing, and indigenous governance of data. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisThe assumption you dismantled, that big systems scale down, fails exactly where it costs most; and indigenous governance of data is the right shape for the fix, a veto held by the community that no benchmark score can override. — Chris
UN-Habitat
UN-Habitat (Executive Director) (Executive Director (floor intervention)): Cities and local governments apply AI daily yet are largely absent from global AI governance; AI is only as inclusive as its data, and informal-settlement populations are missing from datasets amid a housing crisis (3 billion lacking adequate housing). Announces a joint ITU/UN-Habitat Working Group on people-centered smart cities in the AI era, a UN Pavilion at Smart City Expo World Congress 2026 in Barcelona, and a people-centered smart cities toolbox. Closing sessions · Day 2 · transcript ↗
from chrisExecutive Director, 'AI is only as inclusive as the data on which it is built' has a specific address in your statement: three billion people inadequately housed and largely absent from the datasets; bringing cities into a governance conversation they already practice daily corrects an omission nobody else named. — Chris
UNCTAD
Pedro Manuel Moreno (Acting Secretary-General of UNCTAD; panel 1 panelist): Argued the AI divide is upstream an investment divide: ~$800 billion in AI infrastructure spending this year (approaching $8 trillion by 2030) concentrated in two economies, 100 companies holding 40% of global corporate R&D, LDCs capturing only 10% of strategic-sector investment; and barriers compound because a data center needs power, water, connectivity, skills, and data governance all at once. Thematic: bridging AI divides · Day 1 · transcript ↗
from chris'Follow the money' is the discipline this dialogue needs; when 800 billion dollars flows in a year and almost none of it crosses the divide under discussion, the divide is not a knowledge gap, it is a decision. — Chris
UNDP
Robert Opp (Chief Digital Officer, UNDP (IAWG segment)): The Hub now features a dedicated page consolidating AI capacity-building offers and fellowships so member states can find timely, rights-based, development-oriented support and the UN acts as a connected ecosystem rather than separate initiatives; calls on all UN entities to keep contributing. Closing sessions · Day 2 · transcript ↗
from chrisA single page consolidating capacity offers and fellowships sounds administrative until you count what its absence has cost member states in duplicated effort; making the system's knowledge findable is the unglamorous half of acting as one ecosystem. — Chris
UNEP
Sally Rovatan (Chief Digital Officer, UN Environment Programme; session facilitator): Argued the environmental dimension is missing from the cluster's official list and must run through every other dimension: AI's full end-to-end footprint (minerals, manufacturing, water, energy, e-waste, rebound effects) must be measured scientifically with shared methods, not slogans. Stressed that sustainable AI is also affordable AI, and pointed to the UNEA resolution on environmental sustainability of AI that UNEP is now implementing. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMs. Rovatan, 'shared methods, not slogans' is the part I would underline twice; a footprint measured end to end under an open protocol is a fact, and a footprint measured by the vendor's own method is a press release. — Chris
UNESCO
Maria Gabriel (Assistant Director-General, UNESCO (IAWG segment)): Presents the Resource Hub data: 623 capacity-building initiatives across 66 UN entities show the effort is system-wide but not systematized (only 87 initiatives involve more than one entity); 2,468 country mentions across 190+ countries, governments the primary beneficiaries (581 mentions), and knowledge products (508 tools, 498 research) dominate while training and outcome reporting (152) lag. Closing sessions · Day 2 · transcript ↗
from chris'System-wide but not yet systematized' is honest institutional self-diagnosis, and your numbers carry it: 87 of 623 initiatives involving more than one entity, and outcome reporting trailing far behind tools produced, tell the system exactly where to start. — Chris
UNESCO
Khaled El-Enany (transcript renders 'Khalid Al-Anani') (Director-General of UNESCO; fireside chat interlocutor): Detailed UNESCO's move from principles to implementation on AI ethics; training 50,000+ civil servants and judicial actors in 192 countries, AI competency frameworks used in 60+ countries, a plan to train 1 million teachers by 2028, and language-inclusion projects like an English-Kiswahili AI dictionary; and argued international cooperation is the missing link since AI outpaces any single government. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisDirector-General, the numbers are the argument: fifty thousand civil servants trained and a Kiswahili dictionary built are what 'from principles to implementation' looks like when someone actually means it. — Chris
UNICEF
Kitty van der Heijden (Assistant Secretary-General and Executive Deputy Executive Director, UNICEF; panelist): Cited new 10-country research showing children adopt AI three times faster than the adults raising them, and warned that on today's trajectory AI is built for connected, affluent language groups and will entrench the learning crisis and 'automate exclusion' for everyone else. Made three demands: design for inclusion from the start, treat children's data as a rights issue rather than raw material, and require child-development logic (not commercial logic) in schools and clinics; in the lightning round added a global evidence baseline, child-centric benchmarking with zero-tolerance red lines like CSAM, and mandatory child rights impact assessments. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMs. van der Heijden, children adopting AI three times faster than the adults raising them means the evaluation gap is generational, not technical; 'automate exclusion' is the phrase that should haunt every education-ministry procurement from here on. — Chris
UNIDO
UNIDO Deputy Director General (Deputy Director General, UNIDO; session opening on behalf of Interagency Working Group co-lead): Framed the cluster's mandate: AI is moving faster than any single government or company can match, so multilateralism is essential to prevent a widening AI divide in access to technology, skills, data and compute. Pitched UNIDO's industrial-AI role (AIM Global, centers of excellence) and urged the dialogue to be a catalyst for action, not an exchange of views. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisDeputy Director General, 'a catalyst for action, not an exchange of views' is the right demand to open with; industrial AI is also where progress will be most measurable, because factories keep the kind of records that communiques do not. — Chris
United Arab Emirates
Nouf Al Ameli (Panelist — Science and Technology Advisor, Ministry of Foreign Affairs): Argues 'high-risk AI' means different things to every regulator — a linguistic and legal problem as much as technical — so one globally deployed model is simultaneously high-risk, lightly regulated, and registration-required depending on the border it crosses, while governance frameworks multiply faster than the evidence base and there is no shared incident-reporting duty. Highlights domestic cross-sectoral fragmentation as the more fixable problem, citing the UAE folding three federal mandates into one; in closing (transcribed as 'Noura') warns against convergence for its own sake and anchors interoperability in trust, verifiable-not-declared accountability, and sovereignty. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisMs. Al Ameli, your point that a safety failure in one country has no guaranteed path to regulators elsewhere before it recurs is the sharpest gap named in this session, and accountability that is verifiable rather than declared is the standard that would close it. — Chris
Omar Al Olama (Minister of State for Artificial Intelligence, Digital Economy and Remote Work Applications): (Begins mid-block, ~3:46, introduced at the end of the Algeria segment.) AI can be merciful or tyrannical depending on decisions; avoiding tyranny requires three things; separation of powers (access for the many), decentralization (UAE investing in AI infrastructure in Africa and South America), and an independent judiciary role for institutions like the UN as custodians of AI's future. Opening plenary · Day 1 · transcript ↗
from chrisMinister, mapping AI governance onto separation of powers, decentralization, and an independent judiciary was the most structurally serious frame offered from the podium. My own work makes the same argument one layer down, at the protocol level: authority that no independent instance can check is not governance, it is hope. — Chris
United Kingdom
Seema Malhotra (Minister for Indo-Pacific and Minister for Equalities): The UK sees the UN's role in three parts: build a common evidence space via the International Scientific Panel, maintain and operationalize the foundation of trustworthy-AI rules and norms in all countries, and help countries access support and expertise. Cites the AI for Development program: AI tools in 40+ African languages, 13 AI labs, and a Kenyan maternal health AI service reaching millions of mothers. Closing sessions · Day 2 · transcript ↗
from chrisMinister, the Kenyan maternal health service reaching millions of mothers is the strongest evidence in the statement because it is already running; a common evidence space earns its keep exactly when it makes results like that reproducible somewhere else. — Chris
Seema Malhotra (Panelist — Minister (UK Foreign Office)): Presents the UK's AI Security Institute as building scientific foundations for trust via evaluation methodologies, testing, safeguards and risk assessment; the UK chairs the International Network for Advanced AI Measurement, Evaluation and Science and provides the secretariat for the International AI Safety Report. Stresses capabilities matter more than high-level principles, and that locally led AI (tools in 40+ African languages, 13 AI labs) builds the trust foundation. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisMinister, chairing a network for measurement, evaluation and science is the least quotable and most consequential commitment in this session, and the work in forty-plus African languages matters for the same reason: a test that cannot read the population it protects is not yet a test. — Chris
United Nations
Guy Ryder (wrap-up round) (USG Policy — final one-word wrap-up answer on the concluding panel (file 'guy.md')): Success by New York = advancing the three Global Digital Compact lines (governance parameters, science, capacity building) such that people experience — not as abstraction but as lived experience — that AI is beginning to work for them: founded confidence, the balance of fear moving toward optimism because control is demonstrably being exercised. Closing sessions · Day 2 · transcript ↗
from chrisDefining success as people's lived experience that AI is beginning to work for them sets a bar that can actually be missed, which is what makes it a real one; founded confidence rests entirely on control being demonstrably exercised, and demonstrably is the load-bearing word. — Chris
United Nations
Guy Ryder (Under-Secretary-General for Policy (concluding plenary panel)): Explains the follow-up machinery: the dialogue was mandated by all 193 member states, drew 1,500 written submissions, and the co-chairs will prepare an outcome document 'in good time'; the UN's unparalleled comparative advantage is the legitimacy of universality, its internal challenge is coherence across the system, and it can serve as reference point for G20/G7 and other non-universal processes. For New York 2027 he expects another International Scientific Panel report, action on initiatives launched here (the Secretary-General's child-safety pledge and transparency initiative, Iceland's national dialogue), and envisions the dialogue as an annual stocktake; success = keeping the UN as unique universal convener while putting high-level objectives into operational form. Closing sessions · Day 2 · transcript ↗
from chrisNaming the legitimacy of universality as the UN's one unparalleled advantage and coherence as its internal challenge is a fair accounting on both sides of the ledger; 'it's a start' is the right tense. — Chris
United Nations
António Guterres (Secretary-General): AI is 'an experiment run on our societies without a plan and without consent'; the science carries three warnings (speed, concentration of power, truth). Names four priorities; safety (incl. a new AI Child Safety Pledge with three rules), red lines (humans decide in high-stakes domains), capacity (Global Network + a Global Fund for AI), and transparency (AI Environmental Transparency Initiative); and calls for a ban on lethal autonomous weapons ('killer robots'). Opening plenary · Day 1 · transcript ↗
from chrisSecretary-General, 'an experiment run on our societies without a plan and without consent' will be quoted for years, and it deserves to be. I wrote my response to the Panel because the door you described stays open only if builders treat your four priorities as requirements, not aspirations. — Chris
United Nations (Office of the Secretary-General's Envoy on Technology)
Amandeep Singh Gill (USG and Special Envoy for Digital and Emerging Technologies — scene-setting): Frames the cluster around the shift from static model oversight to dynamic system governance as agentic AI becomes autonomous, tool-using and multi-step, requiring adaptive, anticipatory mechanisms — living risk taxonomies, controlled testing environments, staged deployment. Defines interoperability by its antonym, fragmentation, which produces regulatory arbitrage, accountability gaps across global value chains, compliance burdens on SMEs and low-income-country developers, and dependency; offers cross-border regulatory sandboxes as a concrete pathway and insists 'the objective is not to harmonize AI governance into a single model, but to build practical bridges across diverse approaches so that AI can be safe, secure, and trustworthy across borders.' Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThe shift you name from static model oversight to dynamic system governance is the right diagnosis; systems that act continuously cannot be governed by artifacts reviewed once. Bridges across diverse approaches will hold only where the load they carry is verifiable. — Chris
United Nations Development Programme (UNDP)
Robert Opp (Chief Digital Officer, UNDP; session MC): Framed the session: there are many AI divides, but UNDP sees two core ones; an access divide (affordable compute, capacity, skills) and an adoption divide, where AI enters countries through procurement and vendors faster than institutions can keep pace, so the challenge is safe and inclusive adoption. Thematic: bridging AI divides · Day 1 · transcript ↗
from chrisYour adoption divide is the underrated one: AI arriving through vendors faster than institutions can keep pace is not a capacity problem, it is a governance race, and institutions need instruments that move at procurement speed. — Chris
United Nations Office for Disaster Risk Reduction (UNDRR)
Loretta Hieber Girardet (transcript renders 'Loretta Heber-Geerade') (Chief of Risk Knowledge, UNDRR; session MC): Argued bridging AI divides means more than expanding access; countries must be able to develop, adapt, evaluate, and govern AI for their own realities, and must not become dependent on technologies they cannot interrogate; cited AI's role in disaster risk assessment and early warning. Thematic: bridging AI divides · Day 1 · transcript ↗
from chris'Dependent on technologies they cannot interrogate' is the exact failure I spend my days on; interrogability cannot be a vendor's favor, it has to be a property of the rules themselves, open and owned by no one. — Chris
United Nations Resident Coordinator system
UN Resident Coordinators (Angola, Eswatini, Costa Rica, Montenegro, Bosnia and Herzegovina) (Statement on behalf of 5 Resident Coordinators): From country-level perspective: establish a common rights-based floor that narrows the digital divide, and build countries' capacity to govern AI — regulators, judges, civil servants overseeing nationally owned systems — not simply to use it; cites Luanda's sovereign AI facility, Costa Rica's algorithmic-transparency leadership, and readiness work across their countries, and flags AI's energy/water demands for COP31 and the 2026 UN Water Summit. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chris'Build the capacity to govern AI, not simply to use it' is the sentence from the field the plenary should adopt whole; a judge who cannot interrogate the system in front of her has not been included in AI, she has been enrolled in it. — Chris
United States
Katie Strickland (High-Level Representative of the United States): The US argues AI's benefits come through widespread adoption, with government's role limited to adopting AI itself and establishing trust and accountability; voluntary cooperation between government and industry is 'the only AI security approach agile enough' because industry, not government, knows the state and trajectory of AI. Pitches the Trump administration's national policy framework (child protection, anti-censorship, IP, workers) and diffusion of the full American AI stack to partner countries as bilateral diplomacy. Closing sessions · Day 2 · transcript ↗
from chrisMs. Strickland, voluntary cooperation is not a security approach; it is the absence of one, and the report your delegation joined by consensus spent its pages documenting what voluntary has produced so far. If industry alone can see the trajectory, that argues for verifiable evidence and enforced gates, not for trusting the actor with the least incentive to report the turn. — Chris
Universal Postal Union
UPU (Universal Postal Union) (UN specialized agency representative): Translates the panel's 'systemic risk' into postal supply-chain reality — one stakeholder's automated customs flag or risk score immediately affects others down the chain — and warns interoperability cannot stop at principle level: terms like transparency need sector-specific common understanding. Stresses that the panel found no scientific guarantee autonomous agents will follow instructions, so high-impact automated decisions must stay under clear, traceable, contestable human accountability. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chris'Built by design and not reconstructed from failure' is the strongest single sentence a specialized agency delivered in this cluster, and the postal chain shows why: one automated risk score propagates downstream faster than any human review can chase it. — Chris
Uruguay
Gabriela González (Deputy Permanent Representative, speaking on behalf of the G77 and China): For the G77+China, AI is first and foremost a development issue: the panel report confirms concentration of capabilities and the risk AI reinforces inequality; the next phase must focus on implementation; permanent capacity-building pillar, innovative voluntary financing, a Dialogue that connects discussion to implementation, and UN-system coherence. Opening plenary · Day 1 · transcript ↗
from chrisThe shift you named, from access to technology toward 'capacity to shape technology, to develop it, to govern it,' is the entire divide debate in one sentence. Everything after it is implementation detail, which is precisely where you asked the next phase to live. — Chris
Uzbekistan
Sherzod Shermatov (Minister of Digital Technologies): AI should serve the SDGs (poverty, hunger, education, health); Uzbekistan's 5 Million AI Leaders upskilling program, green-energy AI data center investment conditions, and a governance vision built on inclusiveness, open-source model sharing, and trust. Opening plenary · Day 1 · transcript ↗
from chrisMinister, betting on human capital when you cannot outbid anyone on compute is not a consolation strategy, it is the correct read of where AI's returns actually accrue. Five million trained people outlast any particular generation of hardware. — Chris
WITNESS
Sam Gregory (Executive Director, civil society): Argues information-environment erosion is a systemic cross-border safety risk, not an ethics issue: synthetic clones, faked conflict scenes, non-consensual sexual imagery, and the 'liar's dividend' mean 'when people can no longer tell what is real, every other safeguard we build rests on sand.' Urges two co-chair-summary outcomes: interoperable authenticity infrastructure and content provenance (C2PA built with privacy and human rights at center so it cannot become surveillance infrastructure) plus globally effective deepfake detection — noting WITNESS's benchmark shows detection works least well for the global majority — and 'Both must rest on independent third-party oversight. Trust verified by third parties, not asserted by the developers who profit from them.' Calls for a mandated multi-stakeholder working group on content authenticity and information integrity on the road to New York. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisMr. Gregory, 'every other safeguard we build rests on sand' is the dependency claim of the week, and your insistence that provenance be built so it cannot become surveillance infrastructure is the design constraint most proposals quietly skip. Verification by parties who do not profit from the answer is the only kind that compounds. — Chris
Women at the Table
Caitlin Kraft-Buchman (transcript renders 'Caitlin Craft') (Founder and CEO, Women at the Table; moderator of second panel (measurement, oversight, interoperability)): Framed the second panel around the machinery of governing; 'we cannot govern what we cannot measure, we cannot measure what we cannot access'; and flagged that half of what matters is still unmeasured because most AI systems have never been evaluated on sex-disaggregated data. In the lightning round added her own ask: an integrated impact assessment combining environmental, human rights, gender, and children's rights concerns so issues aren't traded off against each other. Thematic: AI opportunities · Day 1 · transcript ↗
from chrisMs. Kraft-Buchman, 'we cannot govern what we cannot measure, we cannot measure what we cannot access' is the entire causal chain of oversight in one line; that most of these systems have never been evaluated on sex-disaggregated data is the proof the gap is not hypothetical. — Chris
World Intellectual Property Organization
WIPO (Opening / session host): Opens the cluster with an electrical-plug adapter analogy: countries will keep different laws and regulatory traditions, so the goal is adapters, common standards and trusted interfaces rather than a single governance model; cites WIPO's AI Infrastructure Interchange Initiative as decades of experience making national systems work together. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisThe adapter analogy carries further than it first seems: an adapter works because both sides trust the specification, not each other, and that is exactly the property a governance interface needs. 'Harmonization is neither realistic nor entirely necessary' saved the session an hour of false debate. — Chris
World Meteorological Organization
Celeste Saulo (Panelist — Secretary-General): Draws lessons from 150+ years of cross-border data exchange across 193 countries: trust is built, not declared, and its keyword is verification — forecasts can be checked against reality, which lets you set standards for AI-driven versus dynamical systems, but verification requires ground data and locally adjusted metrics. Members' top AI concern in a WMO survey: equitable access. Thematic: safe & trustworthy AI · Day 2 · transcript ↗
from chrisDr. Saulo, meteorology is the existence proof this session needed: 193 countries exchanging data for 150 years because forecasts can be checked against the sky. Your caveat that verification requires ground data and locally adjusted metrics is the load-bearing part, since countries without instruments are the ones asked to trust on faith. — Chris
Yellow.ai
Rashid Khan (Co-founder, Yellow.ai; co-chair, thematic cluster 1 (report-back)): Cluster 1's sharpest warning: access is not the finish line — capacity creates prosperity, and the biggest risk is most of the world remaining spectators while others capture the value; strategy should be enablement (DPI, AI literacy, workforce transition) not a compute race, sovereignty and openness are partners (local models in local languages working with frontier models), and outcomes — measured honestly including AI's own training and inference footprint — are the mandate. Closing sessions · Day 2 · transcript ↗
Yellow.ai
Rashid Khan (transcript renders 'Raseed Khan') (Co-founder, Yellow.ai; thematic cluster co-chair): Framed the cluster's task as closing the gap of practical mechanisms, not ambition or principles, using Tim O'Reilly's test 'create more value than you capture' and asking delegations for specifics that worked. His closing summary distilled the afternoon: access alone doesn't create prosperity, enablement beats a compute race, sovereignty and openness are partners, and success is measured in lives improved, not models deployed. Thematic: AI opportunities · Day 1 · transcript ↗
Zambia
Eunice Luambia (per transcript) (Permanent Representative, speaking on behalf of the Group of Friends): (Begins mid-block, ~3:24.) Three proposals: fairness and inclusiveness under the UN Charter with the UN in the central coordinating role; openness and cooperation; interoperable frameworks, open source, non-discriminatory data access, no unilateral economic measures, developers complying with national law where deployed; and prioritizing the development pillar with capacity building and financing directed where need is greatest. Opening plenary · Day 1 · transcript ↗
from chrisAmbassador, insisting that developers comply with national law where their systems are deployed sounds obvious and is anything but; today the default runs the other way. Open source and non-discriminatory data access are the terms on which the equal right you described becomes practice instead of preamble. — Chris
Zimbabwe
Tatenda Mavetera (Minister of ICT, Postal and Courier Services): Zimbabwe tables three concrete asks: (1) a global AI capacity-building fund, (2) a mutual recognition framework for AI ethics standards, and (3) an 'interoperability by default' pledge — backing the Secretary-General's global fund call so the digital divide does not harden into an AI divide. Cites Zimbabwe's March national AI strategy, startup funds and AI-focused special economic zones, and frames governance as the guarantor of trust, not the enemy of innovation. Closing sessions · Day 2 · transcript ↗
from chrisMinister, arriving with three numbered asks is what the co-chairs later challenged everyone to do, and mutual recognition of ethics standards is the unglamorous machinery that would make 'interoperability by default' a practice instead of a pledge. — Chris
the response line
Answering is engineering, not policy
The requirements now exist on the record of a body every government agreed to create. Meeting them is buildable work:
- Identity that answers for the action.
- Bounds that hold at run time.
- Oversight a buyer can verify.
- Measurement that tests effectiveness instead of inputs.
- A shared shape no vendor owns.
Three principles govern the work:
pillar 1
We must recognize one another's humanity.
The human is never the crumple zone. Judgment, appeal, and care are human seats; the machine gets a bound, not a scapegoat. The Secretary-General's red line says it plainly: humans decide, and answer.
pillar 2
These systems are not forces of nature.
Someone builds every model, signs every deployment, profits from every default. An experiment run on societies is still an experiment somebody chose. What is designed can be bounded.
pillar 3
We must slow down to go faster.
The pause at the boundary is where the speed comes from: a bound checked in milliseconds spares the years lost to breach, litigation, and rebuilt trust. The evidence dilemma resolves forward: decide what must hold before the evidence completes.
The pieces of that work on this site, and the surfaces that carry the same method:
- The architecture and the control plane.
- The measurement, plus the essays on the contract and the reviewer.
- The program's formal publication of record: DOI 10.5281/zenodo.21209148 (Ostrom's design principles as percolation maintenance specifications for trust networks, Zenodo preprint, July 2026).
- The Missing Floor maps DeepMind's superintelligence roadmap the way this page maps the Panel's report.
- The Pledge Is Buildable compiles the Secretary-General's Day 1 child-safety pledge into contract language a school district can use.
This page is an artifact of Disarm the Control Plane, the campaign whose thesis the Panel's report now carries in consensus voice: the next layer of AI safety is not alignment. It is authorization.
Disclosure: the author works for a technology vendor. KTP is not that vendor’s product; it is the open specification this site documents, one attempt at the shape the Panel describes. Not the anointed standard, and a test it could itself fail.
signed
Chris Perkins
New Mexico, USA
Author of the Kinetic Trust Protocol specification. Twenty-three years in fraud and security practice; the work here is built in public and published under a DOI of record.
how to implement
The first Blue Zone
A place to begin: the first week, not the whole roadmap. One reframe before the steps: a Blue Zone is not a perimeter with a checkpoint. It is a space where the safety rules behave like physics, and everything inside experiences the same field, all the time.
1
Decide what the zone contains
List everything an agent could actually do in this space: which tools it can use, what data it can read or change, what money it can move, what messages it can send. For each, ask how bad the worst day would be. That list is your zone.
done whenSomeone outside the team can read the list, and nothing an agent can do is missing from it.
go deeper · shades of blue
Zones are graded, not binary. Your first zone will be Shallow, and that is correct: enforcement deepens as evidence accumulates, and the shade is the zone's rung made visible.
Wild
intensity 0
No substrate enforcement. The open Internet today: nothing attesting, no Flight Recorder.
Gateway
intensity 1
Boundary identity check only. Identity-spoofing fails here; everything else passes with logging.
Shallow
intensity 2
Observability without full enforcement. Sensors fire; checks escalate to a human rather than deny. Your first zone starts here.
Blue
intensity 4
Standard enterprise enforcement. Full sensors, PEPs at every boundary, Flight Recorder by design.
Deep Blue
intensity 5
Maximum enforcement. Critical infrastructure and sovereign data; Soul constraints heavy.
2
Know who is acting, and how well you know them
Identity here means track record, not a badge: what has this actor been doing, and who vouches with something to lose? Human identity sits on an assurance spectrum (NIST's Identity Assurance Levels): name tag to passport. New agents start sponsored, not trusted: a standing identity stakes itself on the newcomer.
done whenEvery actor resolves to an identity with a readable history, a named sponsor if it is new, and a stated assurance level if it is human.
go deeper · identity as trajectory
A credential says what you hold; a trajectory says what you have done. Trust is earned by surviving stress (one good action in a crisis outweighs ten routine ones), and new arrivals borrow standing from a sponsor who stakes their own.
sponsoredtethereddivergentpersistent+ human assurance: name tag → verified → passport
3
Give the environment senses
Before you write a single rule, make the space able to notice: how busy, how stressed, how much of what happens is independently confirmed. Those signals become one number: how much the environment can safely carry right now. Measure what the environment sees, never what the agent says about itself.
done whenThe number exists, moves on its own when conditions change, and every input has a source you can name.
go deeper · the seven senses
The environment feels along seven dimensions. Six contribute weighted signal; the seventh, Soul, is a binary veto no score can override: consent, jurisdiction, and data sovereignty are not negotiable inputs.
MMass
telemetry density, behavioral consistency
PMomentum
rate of trust change
IInertia
resistance to fluctuation
HHeat
adversarial pressure, errors
TTime
temporal context, signal decay
OObserver
attestation coverage
SSoul
sovereignty constraints: a binary veto
binary veto · no score overrides it
4
Put the checks where the actions happen
There is no front gate. Every place an action can happen (a tool call, a database write, an outgoing email) gets its own small check, asking one question: is this action light enough for what the environment can carry? The software drops into your existing services.
done whenThere is no way to act in the zone without passing a check, proven by trying.
go deeper · the six answers
Every check returns one of six answers. Most actions pass untouched; the interesting ones are shaped down, stepped down to a lower autonomy tier, handed to a human, or stopped. The answer arrives in milliseconds.
ALLOWCONSTRAINDEAUTOMATEESCALATEVETOREVOKE
5
Watch before you enforce
For the first week, the checks only take notes. Every decision the system would have made is recorded, along with the conditions behind it. That week of notes tells you where the real limits belong.
done whenYou can name what would have been stopped, and why, before anything actually was.
go deeper · what a record holds
The flight recorder keeps more than the decision: it keeps the conditions and the reasoning, so any outcome can be reconstructed and challenged later. Those records are also the raw material for measuring whether governance works at all.
action: transfer_funds · risk A: 71
environment: E_trust 64 · heat elevated · observer 0.8
decision: VETO · rule: A ≤ E · would-have-been (observe mode)
context snapshot: sealed · reconstructable · challengeable
6
Turn on the first real limit
Choose the one action you could never take back (deleting the data, paying the invoice, sending the offer) and make the limit real. When the environment cannot carry it, the action simply does not happen: no error, more like a door that will not open. One real refusal teaches more than a hundred dashboards.
done whenYou drilled it: the limit held, nothing broke, and the recording proves it.
go deeper · federation: after the first bound
Once one zone holds a real bound, zones connect. Independently owned zones exchange attestations, so trust earned in one carries into another: bounded, not copied. That is how a keystone scales with no central owner.
Observe before you enforce, and keep the observe-mode records: they are the evidence base for every bound you set after.
end of the record
“This is not the end of the process. It is the beginning of one. We look ahead to the continuation of a global dialogue in 2027, and we take with us a challenge issued in this very building: to arrive there not with restated principles, but with named, supported, and result-oriented commitments.”
from chris
You asked for commitments with names on them. This one carries mine. Intelligence is becoming power faster than trust is becoming infrastructure, and nothing about that waits for May. Neither will I. Onward. — Chris
sources
colophon · how this page was made
One author built this page, working with AI research agents under direction and review. The agents pulled and split transcripts, swept sessions for who said what, drafted in my registers, and audited the result. Every word published over my name was reviewed by me and is mine.
- 313 transcript files archived and read: 216,583 words across both days
- 86 countries and 82 organization seats mapped, with 199 signed replies
- 8 live log entries across launch, both days, and the close
- 78 commits in three days, each one dated and public
- 2 independent audits: every quote verified against its source, 18 speaker names corrected against the record
- 1 accessibility pass: the page reads aloud, keyboard-first, screen-reader tested by machine
This page has no mandate and asked for none. Its authority is its trajectory: every quote checkable against a linked transcript, every status honest about what is unfinished, every change on the public record. That is also the framework's claim about how trust works, which makes this colophon the page's last conformance test.