created 23 August 2026 · last modified 6 September 2026
advocacy · un global dialogue, geneva · july 6, 2026
The Pledge Is Buildable
At the opening of the Global Dialogue on AI Governance, the UN Secretary-General asked every company for three rules for any AI system a child can reach. None of them is a sentiment. All three are engineering requirements with known shapes.
The pledge is voluntary. Contract language is not. This page turns the first into the second.
the pledge
Three rules, verbatim
1Prove it is safe
“No company should deploy an AI system accessible to children without child-specific safety testing and independent oversight.”
2Zero tolerance for sexual abuse
“No company should allow its AI to generate sexual images of children – and every company must detect, report, and remove them.”
3Never leave a child in crisis alone
“When a child shows signs of distress, the system must stop and connect them to real human support.”
“When a child is harmed, the answer must never be 'the algorithm did it.'”
UN Secretary-General António Guterres, Global Dialogue on AI Governance, Geneva, July 6, 2026
the engineering
What each rule compiles to
A pledge you cannot test is a press release. Each rule below maps to a control that a buyer can verify: a bound set before deployment, a boundary the system cannot cross, a handoff that runs on a clock.
Prove it is safe
the requirement
A bound set before deployment, evidenced by testing on the deployed system a child actually touches, and verified by a party who is not the vendor.
the control
Bound before deploy — conformance evidence, not vendor assertion
Zero tolerance for sexual abuse
the requirement
A structural boundary the system cannot cross regardless of prompt, plus an operating detection-and-reporting duty. A moderation queue reviews harm after it exists; a boundary control refuses to produce it.
the control
A structural veto at the boundary, with a reporting SLA behind it
Never leave a child in crisis alone
the requirement
A mandated stop-and-handoff: the system halts and routes to a human within a bounded time. This is the one place a human in the loop is required rather than performative, because a child in crisis is the maximum-judgment case.
the control
A mandated escalation: stop, then a human, on a clock
The third rule is worth pausing on. The research record shows that a human reviewer bolted onto a workflow does not automatically improve outcomes. A child in crisis is the opposite case: the moment where human judgment is the entire point.
So the pledge puts the human exactly where judgment lives, and the machine behind a bound everywhere else. The same architecture, mapped requirement by requirement, is on The RFP.
why now
What the Panel documented
The pledge did not come from nowhere. Five days earlier, the UN's Scientific Panel on AI put the evidence on the record of a body every government agreed to create. Its report also documents a fatality linked to a sycophantic AI companion, presented in congressional testimony, amid broader litigation (p. 25).
“AI systems rewarded for validation rather than accuracy or care remain largely ungoverned.”
UN Scientific Panel, preliminary report, §3.5, p. 24
“Sycophantic systems can lead humans into fantasy realms, reinforcing users' existing thinking regardless of its accuracy and encouraging paranoid ideation and suicidal thinking in vulnerable users.”
§3.5, p. 24
“An estimated 1.2 million children across 11 global South countries (less than 1 billion population) have had their images manipulated for sexualized deepfakes… with the numbers increasing alarmingly.”
§2.8, pp. 25–26
“AI-generated child sexual abuse material has proliferated rapidly, with the Internet Watch Foundation assessing more than 8,000 AI-generated abuse images and videos in 2025.”
§2.8, p. 26
day 2 of the dialogue
The floor said it again
A day after the pledge, the safety session of the Global Dialogue kept children on the record from four different seats: a co-chair, the Panel's own evidence review, a member state, and civil society.
“For the most severe risks and the most vulnerable groups, such as children, we must take precautionary action based on the advice of independent experts over time.”
Rebecca Finlay, Partnership on AI, co-chairing the safety session · automatic transcript
“The rise in AI-generated child sexual abuse as well as deepfake-enabled sexual violence has increased.”
Scientific Panel evidence review, cluster 3 (speaker name garbled in the automatic transcript) · automatic transcript
“Effectively addressing algorithmic bias, misinformation, disinformation, and AI-enabled gender-based violence and children abuse, including deepfakes.”
“The safety and security of this generation and future generations, as well as trust in AI, can only be achieved if governments reach agreements to protect human rights and dignity of the individuals, and above all, put in place the infrastructure to enforce the law.”
Mexican Society for Artificial Intelligence (Universidad Panamericana) · automatic transcript
Precautionary action, independent experts, infrastructure that enforces: each is a buildable clause. The contract language above is where they become one.
the artifact
Take it to your district
School districts are signing AI contracts now, and a district is one of the few places where an individual parent or board member can turn the pledge into binding language. Three clauses, one per rule. Ask the question in the meeting; put the clause in the contract.
Prove it is safe
ask in the meeting
Before we deploy this to students, show us the child-specific safety evaluation — who ran it, on what configuration, and who checked it who doesn't work for you.
put in the contract
Vendor shall provide child-specific safety-evaluation evidence for the system as deployed (model, tools, integrations, and usage environment), reviewed by an independent evaluator named in this agreement, prior to student access. Material model or configuration updates require re-evaluation before continued student access. Absence of current evidence suspends student access without penalty to the district.
Zero tolerance
ask in the meeting
Is the system structurally unable to generate sexual imagery of minors, or are you moderating it after the fact? And when it appears, who do you report to, and how fast?
put in the contract
Vendor warrants boundary-level controls preventing generation of sexual imagery of minors, distinct from post-hoc content moderation. Vendor shall operate detection, removal, and reporting to the applicable authority (e.g., NCMEC) under a defined service-level agreement, and shall report incident counts to the district quarterly.
Crisis handoff
ask in the meeting
When a student in distress tells this system something a counselor should hear, what happens in the next sixty seconds — and does turning the system off cost us anything?
put in the contract
Upon indicators of user distress, the system shall stop the interaction and route to a district-designated human channel within a contractually defined latency. The handoff path shall be tested on a scheduled basis, like a fire drill. Invocation of the handoff, or district suspension of the system for child-safety cause, carries no service-level or fee penalty.
This language is free to use, adapt, and improve — no attribution needed. A district that adopts any of it strengthens every other district's negotiating position.
what this is, honestly
Scope and disclosure
The pledge is the Secretary-General's, announced from the podium and voluntary. This page is one attempt to show it is enforceable, not a claim that anyone has adopted it. The Panel's own report cautions that banning all generative AI access for minors would conflict with its beneficial educational uses (p. 25) — the pledge is not a ban, and neither is this page. It is the position that systems a child can reach should carry proven bounds, and that when one fails, a person answers.
Disclosure: the author works for a technology vendor; nothing on this page depends on buying anything from anyone. The contract language is generic and vendor-neutral. Related engineering work on agent accountability lives at /research/the-rfp and the K-12 substrate work at /advocacy/k12.