Who Writes the Rules: The UN Seats the AI Labs at the Table

On July 1, the United Nations and its International Telecommunication Union launched the AI for Good Global Commission — co-chaired by Salesforce CEO Marc Benioff and Rwandan President Paul Kagame, and seating, as founding members, Nvidia's Jensen Huang, Amazon's Andy Jassy, Microsoft president Brad Smith, Anthropic co-founder Jack Clark, and Cohere co-founder Aidan Gomez, alongside more than forty heads of state, cabinet ministers, and UN agency chiefs.

Read that membership list again, because it is the whole story. The most senior gathering of AI executives and heads of state ever assembled under a UN mandate did not convene to be regulated. It convened to govern — to publish recommendations, convene working groups, and broker voluntary commitments on the technology those same executives build, sell, and profit from.

It is, by the UN's own framing, designed to be a "smaller, faster team of business executives" that can move without the procedural weight of the 193-member General Assembly. No ratification. No treaty. No consensus of the world's governments. A commission that can act at the speed of a board meeting, staffed by the people who run the board meetings.

I am writing this on the Fourth of July — the weekend the United States celebrates a document whose entire premise is that legitimate authority derives from the consent of the governed. So it is a fitting weekend to ask the question this commission forces into the open: when the governed and the governors are the same people, what exactly is being consented to?

What Was Actually Created

Strip away the launch language and the Commission is a specific institutional object with specific powers — and, more importantly, specific non-powers.

It can: publish recommendations, stand up thematic working groups, convene its members, and broker voluntary commitments among the companies and states at the table. It cannot: bind anyone, levy penalties, compel disclosure, or override national law. It has no enforcement arm because it was deliberately built without one. The stated design principle — a "smaller, faster team" — is not an accident of scope. It is the point. Speed was purchased by shedding the machinery that makes international rules enforceable.

That trade is legible in who sits on it. This is not a body of regulators, inspectors, or ombudsmen. It is a body of principals: the people with the largest possible stake in the outcome, granted a UN letterhead and a convening power.

The presence of Kagame as co-chair matters more than it might look. Rwanda has spent a decade positioning itself as the convening ground for the Global South's digital ambitions, and Kagame's co-chairmanship is the fig leaf of inclusivity over a body whose center of gravity is a handful of American hyperscalers. That is not cynicism; it is structure. When you seat the CEO of the world's most valuable chipmaker next to the president of a nation of fourteen million, the commission can credibly say it "brings the world together." Whose problems get prioritized in the working groups is a different question, and it is the one worth watching.

The Three-Track Architecture Nobody Is Describing Clearly

The Commission did not appear in a vacuum. It is the third and newest track in a governance architecture the UN has been assembling since the 2024 Global Digital Compact, and you cannot evaluate it without seeing the other two.

Track one is the Independent International Scientific Panel on AI — the evidence-producing body, modeled loosely on the IPCC. Its job is to publish independent assessments of AI's capabilities, risks, and impacts, and to present an annual report. It is the closest thing the world has to a neutral referee on what the technology can actually do. It presents its preliminary report at the Geneva dialogue on July 6.

Track two is the Global Dialogue on AI Governance — the deliberative body. This is the inclusive, slow, legitimate track: all 193 governments, plus private sector, academia, and civil society, convening to discuss international cooperation and share best practices. It is where every stakeholder gets a seat and, correspondingly, where nothing moves quickly, because moving quickly with 193 sovereigns in the room is not a thing that happens.

Track three is the Commission — the executive track. Small, fast, exec-led, unbound. It exists precisely because tracks one and two cannot ship. The Scientific Panel produces knowledge but no action; the Global Dialogue produces legitimacy but no velocity. The Commission was chartered to supply the missing ingredient — the ability to actually do something — and it bought that ability by abandoning the two things the other tracks have: independence and universal representation.

Seen this way, the Commission is not a scandal. It is a rational response to a real problem: the deliberative machinery of international governance moves in years, and frontier AI moves in weeks. Somebody looked at that mismatch and decided the answer was a body that could keep pace. The only question — the entire question — is whether the price of that velocity was worth paying.

Which Governance Model Is This, Really?

Analogies are doing a lot of quiet work in the coverage of the Commission, and most of them are wrong. It is worth being precise about which prior institution this most resembles, because the analogy you reach for determines whether you expect it to succeed.

It is not the IPCC. The Intergovernmental Panel on Climate Change is an evidence body — it synthesizes science and hands governments an assessment; it does not negotiate commitments. The UN already built the IPCC-analogue: it is the Scientific Panel, track one. The Commission is something else.

It is not the IAEA. The International Atomic Energy Agency has inspectors, safeguards agreements, and the power to refer violations to the Security Council. It can walk into a facility. The Commission has none of that and was built to have none of it.

The closest real analogues are softer and more revealing. The Financial Stability Board — a body of central bankers and finance ministries that sets standards the members then implement through their own national authority — is the optimistic template: influential precisely because its members are the regulators, so a standard agreed at the board becomes binding at home. But the Commission's members are not regulators. They are the regulated (the labs) and the sovereigns (the states) in one room, which breaks the FSB mechanism: there is no automatic path from "the board agreed" to "it binds," because the labs cannot bind themselves through national law and the states did not send their regulators.

The more honest analogue is ICANN or the Internet Governance Forum — a multistakeholder body that coordinates a technical domain through rough consensus and voluntary adherence, credible within its lane, but perpetually contested on legitimacy grounds precisely because private incumbents sit at the table as principals. Multistakeholder governance has a genuine track record in internet infrastructure. It also has thirty years of unresolved argument about whose interests it actually serves. The Commission is importing both the model and the argument.

The reason this matters is that multistakeholder internet governance worked best for problems that were genuinely technical and coordination-shaped — allocating addresses, agreeing on protocols — where every participant benefited from a shared answer and nobody's core business model was threatened by the outcome. It has worked far worse for problems that are distributive, where one party's gain is another's loss. Frontier AI governance is full of the second kind: a meaningful capability restriction costs the company that would otherwise ship it. Importing a model that excels at coordination problems into a domain full of distributive conflicts is the structural gamble underneath the Commission's design.

The Members Don't Actually Agree

The lazy version of the capture critique treats "the labs" as a monolith with a single interest. They are not. The five frontier-adjacent companies on the Commission have sharply divergent incentives, and those divergences are the most interesting lever in the whole arrangement.

Nvidia sells the picks and shovels. Jensen Huang's interest is that AI keeps scaling and keeps consuming compute — his worst outcome is a governance regime that caps model size or slows deployment. Amazon and Microsoft are primarily deployers and infrastructure providers; their interest is enterprise trust and regulatory predictability, and they can live with — even benefit from — rules that raise the compliance bar in ways that favor incumbents with large compliance departments. Anthropic and Cohere are model labs, but of very different scale and posture: Anthropic has staked its brand on safety and would plausibly accept tighter capability commitments that its positioning already implies; Cohere is a smaller enterprise-focused player whose interest is that governance not entrench the largest labs' advantages.

This matters because a commission whose members disagree is a commission that can, in principle, produce something. The FSB works partly because its members' interests are aligned enough to agree and divergent enough to check each other. If Anthropic pushes a capability-disclosure commitment that Nvidia's interest resists, the negotiation between them is where actual governance could live — not in the unanimous platitudes, but in the contested clauses where one member's proposal survives another's objection. The optimistic reading of the Commission is that its internal disagreements are a feature: enough friction to produce a real standard, enough shared interest to produce any standard.

The pessimistic reading is that the divergences resolve toward the lowest common denominator every self-negotiated code drifts to — the commitment that costs the most resistant powerful member nothing. Which reading is right is, again, an empirical question the deliverables will settle.

What "Teeth" Would Actually Look Like

I keep saying "does it bind" as though that is a binary, so let me make it concrete. There is a spectrum of commitment strength, and knowing where a deliverable lands on it is how you cut through the language.

Almost every voluntary framework in the history of technology self-regulation lands at Level 0 or 1 and is described in press releases as though it were Level 3 or 4. The tell is verification: a commitment nobody outside the committing party can check is, functionally, a Level 2 pledge no matter how strong its verbs. The single most important thing the Commission could do — the thing that would distinguish it from every toothless industry code before it — is to route verification through the independent Scientific Panel, turning the Panel from a report-writer into the referee that decides whether a member met its commitment. That is the bridge between the evidence track and the action track, and it is the difference between an architecture that governs and one that performs governing.

Watch for it specifically. If the first deliverable pairs a commitment with a named, independent verifier, the Commission is trying to build something real. If it does not, you are looking at Level 1 with good production values, and the July 1 launch was the most impressive part.

The AI for Good Lineage

None of this appeared overnight, and the branding is not incidental. "AI for Good" is an ITU program that has run an annual Geneva summit since 2017, growing from a small workshop into the UN's flagship convening on beneficial AI. For most of its life it was a showcase — demos of AI for healthcare, agriculture, and disaster response, long on optimism and short on governance. The Commission is that brand's graduation from showcase to would-be rule-setter, and the continuity is deliberate: the "for Good" framing pre-loads the answer to the question of whether the body is a good idea.

That framing is worth resisting, not because the intentions are bad but because "for Good" is doing rhetorical work. A commission named for its aspiration is harder to criticize than one named for its function, which — stated plainly — would be something like the "AI Industry Coordination Commission." The name is the first voluntary commitment: a commitment to be seen as beneficial, made before any beneficial act. Hold the body to its function, not its adjective.

The Velocity-Legitimacy Trade

Every governance body sits somewhere on a curve between how fast it can act and how legitimately it can claim to speak for the governed. Courts are slow and legitimate. Executive orders are fast and contestable. Standards consortia are fast and narrow. The Commission planted its flag at the far speed-end of that curve and called the trade a feature.

Here is the mechanism that makes it fast: it does not require ratification by the General Assembly. That single design choice is what separates a working group that can publish something in a quarter from a treaty process that can take a decade. It is also what strips the output of binding force. You cannot have both. A body that can bind you needs your government's consent; a body that only needs forty members' agreement can move next week but can only ask.

I have written before about how the binding, state-led track has been retreating even as this executive track advances. The United States spent 2026 narrowing its own regulatory footprint; Colorado repealed the first comprehensive US state AI law before it took effect; the EU moved to simplify its AI Act obligations; and the federal government recast its posture around a narrow "covered frontier model" executive order rather than broad statute. The three-speed pattern of EU, US, and UK governance I traced in May has, if anything, decelerated on the binding end. The Commission fills a vacuum that binding regulators are actively widening.

That is the uncomfortable synthesis. Voluntary, exec-led governance is not competing with binding governance for the same space. It is replacing it, because the binding alternative is being dismantled at the national level at the same moment the voluntary alternative is being chartered at the international one. The market for actual constraints is not being contested. It is being vacated, and a commission of the constrained is moving into the empty building.

The Capture Question

There is a word for a governance body composed of the entities it is meant to govern, empowered to set the rules through voluntary commitments it negotiates with itself: regulatory capture. Usually capture is a hidden process — an agency slowly staffed by the industry it oversees, a rulemaking quietly shaped by the regulated. The Commission is unusual only in that the capture is the founding design, printed on the charter, announced in a press release.

Common Dreams put it bluntly in its coverage: a commission "full of Big Tech execs." That is not a distortion. It is a description.

The historical record on self-regulation of powerful industries is not encouraging, and it follows a consistent shape: voluntary frameworks arrive early and easily; binding regimes arrive late, if at all, and usually only after a visible catastrophe forces the issue.

Treat those numbers as an illustrative pattern rather than precise measurements — the point is the shape, not the decimals. Ozone is the optimistic outlier: a narrow, technically-tractable problem with a clear substitute chemistry and a small number of producers moved from voluntary to binding in a couple of years. Everything else took a decade to a generation, and tobacco's binding framework arrived only after the health catastrophe was undeniable and litigated. The voluntary phase is not usually a prelude to binding rules. It is frequently a substitute for them — a way to relieve regulatory pressure without accepting regulatory constraint.

Which raises the mechanism-level worry. A voluntary commitment brokered among members has a built-in ceiling: it cannot ask any member to do something that member would veto, because that member is in the room. The commitments that survive negotiation are, by construction, the ones the most powerful members can already live with. That is why self-negotiated codes tend to formalize existing best practice rather than push the frontier of restraint. They ratify what the leaders already do and label it a standard.

The gap in that chart is the reason the Commission exists and the reason to worry about it in the same breath. Something has to close the distance between an exploding capability surface and a flat line of binding international constraint. The Commission is the something. The danger is that a body of principals closes the gap on paper — with a framework everyone signs and nobody is bound by — while the real line stays flat.

The Realist Case For It

I want to argue the other side honestly, because the reflexive "capture!" reaction is too cheap and the situation is genuinely harder than that.

First: the alternative to imperfect action is not perfect action. It is no action. The binding track is stalled and retreating. If you refuse every governance mechanism that includes industry, you are left with the Scientific Panel — which cannot act — and the Global Dialogue — which cannot decide. In a world where capability is compounding monthly, "wait for a legitimate binding treaty" is functionally a vote for ungoverned deployment. The Commission's defenders are not wrong that speed has moral weight when the downside is measured in months.

Second: the expertise is real and it is not fungible. The people who understand what a frontier model can and cannot do, where the actual failure modes live, and what a meaningful safety commitment would even look like technically — a large share of them work at the five companies on this commission. A governance body that excludes them on principle governs in ignorance. The Frontier Model Forum and the industry's shared safety-classifier work over the past year are proof that voluntary, expert-led coordination can produce real artifacts, quickly, that no treaty process would have shipped.

Third: voluntary commitments are not nothing. They create focal points, reputational hooks, and — crucially — a written baseline that later binding regimes can codify. Basel's capital accords started as a voluntary club of central bankers and became the backbone of global bank regulation. The OECD's voluntary privacy principles seeded the binding law that followed decades later. Voluntary-first is a legitimate on-ramp, not only a dead end. The tobacco and data-protection lags cut the other way, but the mechanism can run in the productive direction when the incentives align.

The honest position is that the Commission is neither a coup nor a savior. It is a bet — that expert velocity now beats representative legitimacy later, and that the voluntary baseline it produces will run toward binding rules rather than away from them. Whether that bet pays off is not knowable from the charter. It is knowable only from the output, and the output starts July 8.

Why States Handed This Over

It is worth asking why sovereign governments, jealous of their authority, delegated agenda-setting on the most consequential technology of the century to a body their own chief executives do not control. The answer is that they did not have a better option and they knew it.

Look at that composition. Industry executives and heads of state are roughly at parity, and together they dominate. Civil society and academia — the constituencies whose entire job is to represent the interests of the people not in the room — are a thin slice. That is not a body designed to constrain industry. It is a body designed to coordinate industry and states, with a garnish of civil-society presence for legitimacy. The composition is the policy.

The deeper reason states delegated is that AI governance has an expertise asymmetry no regulator has ever faced. Financial regulators can hire economists who understand banks better than bankers do in the aggregate. Nuclear regulators can hire physicists. But frontier AI capability lives inside a handful of private labs, is changing faster than any external body can track, and depends on proprietary details of training runs that no government has full visibility into. The Scientific Panel is an attempt to build independent expertise, but it is new, under-resourced, and perpetually behind. Faced with that asymmetry, states made the calculation that a fast body with the labs beats a slow body without them. It may even be correct. It is certainly a surrender of a kind.

The Map Has a Hole in It

There is a conspicuous absence in the membership, and it may be the most consequential fact about the Commission. The named members skew heavily American on the industry side — Nvidia, Amazon, Microsoft, Anthropic, Cohere (Canadian, but firmly in the Western enterprise orbit) — with states drawn from Europe, Africa, the Gulf, and Southeast Asia. What is missing is China, and with it the entire Chinese frontier ecosystem that has spent the past year reaching capability parity on multiple axes.

A governance body for a global technology that does not seat the world's second AI superpower is not governing that technology. It is governing one bloc of it. If the Commission produces a meaningful capability commitment and the Western labs adhere while the Chinese labs do not, the commitment does not slow the frontier — it just cedes the unrestricted portion of it to the players who never signed. This is the oldest problem in arms control, and it has a name: the defection problem. A restraint regime that the most aggressive competitor sits outside of either fails or transforms into a competitive disadvantage for its own members.

This tension runs directly into the Global South representation the Commission advertises. States like Rwanda, Namibia, and Nigeria have a real interest in AI governance, but their interest is largely in access — capacity-building, affordable compute, not being locked out of the technology — rather than in restraint. The Gulf states on the membership, flush with capital and building sovereign AI capacity, have an interest closer to the labs' than to the restraint-oriented civil-society groups. Add it up and the coalition inside the Commission that actively wants binding limits on frontier capability is small: a few safety-positioned labs and the thin civil-society slice. The coalition that wants coordination, access, and predictability — but not hard limits — is most of the room.

That is not a conspiracy. It is arithmetic. And it means the most likely output is not a capability restraint that the members would have to sacrifice for, but a coordination-and-access framework that serves the assembled interests well and constrains the frontier barely. Which, to be fair, might be the most that any body could achieve given who was willing to show up — but it is worth naming the ceiling honestly rather than letting the "for Good" branding imply a higher one.

What To Actually Watch

Manifestos are free. The Commission will produce a founding statement full of words like "trust," "access," and "impact," and none of that tells you whether it matters. Here is what does.

The first deliverable's teeth. When the Commission publishes its first framework or commitment — likely within a few months of the July 8 meeting — read it for one thing: does it ask any member to do something that member was not already doing? A commitment to "responsible development" that codifies existing practice is pressure relief. A commitment with a specific, measurable, externally-verifiable obligation — a disclosure requirement, an evaluation standard members submit to, a red-line capability nobody ships — is governance. The ratio of the second kind to the first is the whole scorecard.

Whether the Scientific Panel gets teeth or gets sidelined. The healthiest version of this architecture is one where the independent evidence body sets the agenda the executive body acts on — evidence driving action. The capture version is one where the Commission acts on its members' priorities and the Panel becomes a report nobody reads. Watch the funding, the staffing, and whether Commission outputs cite Panel findings or ignore them.

Whether binding processes revive or atrophy. If the voluntary track becomes the only track — if governments point at the Commission as evidence that AI is "being handled" and let their own binding efforts lapse further — that is the capture outcome, regardless of what any single framework says. The Commission is healthy only if it runs alongside a reviving binding track, not as its replacement.

The Global South's actual leverage. Kagame co-chairs, and states from Namibia to Nigeria to Kazakhstan sit as members. Do their priorities — access, capacity-building, avoiding a compute-and-capability colonialism — show up in the working-group agendas and the deliverables? Or is their presence the legitimizing garnish while the substantive agenda tracks the interests of the American labs that supply the technology? This is the difference between "bringing the world together" as a fact and as a slogan.

I have registered a specific, falsifiable prediction about what the Commission ships by mid-2027 — a voluntary framework signed by multiple frontier labs, but with no binding enforcement mechanism attached. It is the null hypothesis of self-regulation, and I would be glad to be wrong about it.

The Founding-Document Problem

Return to the Fourth of July, because the analogy is exact and it is not decoration.

The American founding worked — to the extent it worked — because it separated the governed from the governors and then bound the governors with a document they did not get to rewrite at will. The entire architecture of legitimate authority rests on that separation. A rule you write for yourself, that you can revise whenever it binds, and that carries no penalty for breach, is not a law. It is a preference with good production values.

The AI for Good Global Commission is a founding document authored by the faction it is meant to constrain. That is not automatically illegitimate — plenty of durable governance has emerged from the powerful agreeing to limit themselves, from the Magna Carta forward. But it is legitimate only under conditions the Commission has not yet met and may not intend to meet: that the self-imposed constraints actually bind, that an independent body can verify compliance, and that the constrained cannot quietly revise the constraints when they bite.

None of those conditions are in the charter. All of them are testable against the output. That is the fairest thing I can say about the Commission on the weekend of its birth: it is not yet a capture, and it is not yet a constraint. It is a question that the July 8 meeting begins to answer, and that the first real deliverable will answer for good.

The frontier spent three years arguing about whose model was best. This is the year the argument moves to whose rules win — and the labs, having largely fought the model war to a draw, have now been handed seats at the table where the rules get written. Whether that produces governance or its convincing imitation is the most important open question in the field. Watch the deliverables, not the declarations.

Further Reading

• Colorado repeals the first comprehensive US state AI law before it took effect — the binding, state-led track retreating at the exact moment the voluntary, exec-led track advances
• The covered-frontier-model executive order and the narrowing of US AI regulation — how the federal posture shifted from broad statute to a narrow model-tier rule
• Three-speed AI governance: the EU, US, and UK diverge — the fragmented binding landscape the UN Commission is trying to route around
• This week in AI: the governance turn, Fable 5's return, and Sonnet 5 goes default — the full week's context around the Commission's launch
• Prediction: the AI for Good Commission ships a voluntary framework with no binding enforcement by mid-2027 — the falsifiable null hypothesis of self-regulation, on the record