On 29 September 2026, President Donald Trump and the leaders of six artificial intelligence companies signed a voluntary one-page document at the White House, the White House Accord on Super Intelligence. It contains no penalties, no enforcement mechanism, no requirement to report safety incidents, and no requirement to publish audit results. Those absences are confirmed fact, from the accord’s text as reported contemporaneously. Everything else here — whether the absences make the pact meaningless or sensible, whether self-policing can work, whether the public will accept it — is argument, attributed throughout to the people making it.

This article does not repeat the full account of the pact. For who signed, the four layers of controls, and what the accord does and does not require, see our explainer on the White House AI safety pact and Super Intelligence accord. This article takes the enforcement question on its own: what “no enforcement” means in practice, what critics say is missing, what supporters and the White House answer, what the “codify later” clause promises, and what the public told pollsters a week before the signing. Both sides are presented in their own terms, with no verdict imposed.

The pact in brief: the thing being debated

The accord was signed by President Trump and the leaders of Anthropic, Google, Meta, Nvidia, OpenAI and xAI (Dario Amodei, Sundar Pichai, Mark Zuckerberg, Jensen Huang, Greg Brockman for OpenAI, and Elon Musk). President Trump called it “morally binding”; it is not legally binding — a premise both sides accept, drawing opposite conclusions.

The accord asks each company training and deploying frontier models to implement four layers — in its own word, “should” implement them: robust internal controls monitoring model capabilities, including cybersecurity, biosecurity and chemical-threat risks, ensuring models do not hack or access systems unintendedly; an internal team checking controls work and problems are fixed; an independent external auditor assessing whether controls work; and an independent board committee receiving reports and overseeing fixes. The companies also agreed to meet regularly to set standards and best practices.

What “no enforcement” means, concretely

“No enforcement” is precise and checkable: the accord creates no mechanism for any outside body to compel compliance, verify it independently, or punish non-compliance.

No penalties. The accord sets no fine, sanction or consequence for a company that fails to implement the four layers, implements them late, or abandons them. CoinDesk, drawing on AP reporting, stated it directly: the pact “has no enforcement mechanism, disclosure requirement or implementation deadline”, and carries “no penalties if they fall short”.

No deadline. The accord sets no date by which any layer must be in place. Some steps are ones the companies already take in some form, according to AP reporting cited by CoinDesk — so compliance and business as usual are difficult to distinguish from the outside.

No publication duty. The accord does not require companies to publish audit results, name their auditors, or disclose the standards auditors apply. Findings go to each company’s board committee, not to a regulator or the public, so from the accord alone the public cannot know whether an audit happened, what it found, or whether problems were fixed.

Companies choose their auditors, with no standards and no government role. The accord leaves auditor choice with each company, sets no independence test or accreditation requirement, and gives the government no role in setting standards, appointing auditors or verifying implementation. It sets no standard for how strong controls must be, how often audits should happen, or what “robust” means. The reported 10-member board and new White House AI official were not part of the accord’s text, which creates no board; as of 28 October 2026, no charter, membership or appointment for either had been verified for the companion explainer.

“Should”, not “shall”, and no power to stop a model. Companies “should implement” the four layers — expectation, not obligation. Nothing restricts developing, training or releasing any model, and no outside body can block a release.

The critics’ case

Critics argue a commitment with no enforcement, no deadline and no publication duty changes little, however good its intentions. Four strands, each attributed to named critics:

Self-policing by the companies that create the risk is not a practical fix. Tom Wheeler, a visiting fellow at the Brookings Institution, put it sharply, as reported by ASEAN NOW: “The people who are creating the problems are designing the solutions.” The argument is structural: organisations profiting from rapid deployment judge their own safeguards, choose who checks them, and decide what the public learns.

Voluntary pledges have been abandoned before. Anthony Aguirre, president of the Future of Life Institute, said in a statement reported by ASEAN NOW that the companies involved have “jettisoned safety commitments countless times before” and would do so again. For critics, a new voluntary commitment with fewer enforcement features than some earlier pledges should not be expected to hold where earlier ones did not.

The United States needs laws, not pledges. Ruth Whittaker of Third Way said the approach “continues to fall short”: the US needs “actual laws on the books” — mandatory pre-release testing, power to block unacceptable-risk models. Hakeem Jeffries called it “entirely unenforceable”.

Existing law arrives after the harm. The White House’s reply is that existing authorities — the Department of Justice and the FTC above all — already provide a backstop. Critics answer that enforcement after the event is not prevention: as advocates quoted by Reuters put it, the Department of Justice “shows up after the harm has already happened”, and “nothing in the September pact is built to stop a large-scale failure before it occurs, only to document one after the fact”.

The supporters’ case and the White House reply

Supporters argue a voluntary, industry-led approach suits fast-moving technology better than binding regulation; legal enforcement is not the only accountability.

Cooperation beats regulation for a frontier technology. President Trump told reporters the accord was “morally binding” and that the companies understood “they have to self-police”, according to CoinDesk. Infosecurity Magazine reported that Trump, “who has pushed back on calls for new AI safety rules, told reporters the accord was morally binding and said AI firms should police themselves”, and that “the agreement creates no new regulatory requirements”. Builders understand the systems best, regulators would struggle to keep pace, and cooperation can be agreed in days where legislation takes years.

Existing authorities are the backstop. The Administration points to existing powers, including the Department of Justice and the FBI. The day after the signing, the FTC confirmed investigations of OpenAI, Anthropic and others over safety — a probe reported to have opened in the summer, before the accord — which supporters cite as evidence existing powers are real and active.

Speed matters, and China is the comparator. The Administration argues binding restrictions would slow American companies and hand ground to China. Senator Ted Cruz praised the accord as “inspired leadership”, saying Americans must “win the race for AI” against China. The argument is not that risks do not exist, but that falling behind is the greater risk.

Public commitment creates reputational accountability. Supporters argue a public, signed commitment is itself a form of enforcement: any future incident will be judged against the structure endorsed. Jennifer Huddleston of the Cato Institute said voluntary agreements can still incentivise safer products. Brad Carson of Americans for Responsible Innovation called the accord a “promising start” but warned enforcement details would be decisive and government oversight ultimately needed.

The “codify later” clause: what it promises and what it does not

The accord states that “over time, it may make sense to codify these steps into laws or regulations”. Infosecurity Magazine reported: “the accord says it may eventually make sense to codify the measures into laws or regulations, but the signatories committed to them regardless.”

The clause is important and weak: it records that legislation is a legitimate destination for these measures, but “may make sense” is not a commitment, “over time” sets no date, and no bill or timetable is identified. As ASEAN NOW reported, critics noted it “still leaves regulators without clear powers now” — a promise to consider legislation gives no regulator a power today.

Binding rules elsewhere illustrate the gap: California signed two AI workplace laws on 30 September 2026 (SB 947, AB 1883), and the EU AI Act’s highest penalty tier carries fines up to 7% of global revenue, as Startup Fortune reported; the US has no federal equivalent. Whether Congress codifies the accord’s steps remains open.

What the public said: the Reuters/Ipsos poll

The pact was signed eight days after a Reuters/Ipsos poll showing deep unease. Published 22 September 2026, the four-day online poll of 1,277 US adults (±3 points) found 73% worried AI companies “haven’t gone far enough to prevent AI from causing serious harm to society”. Thirty-nine percent thought AI was having a negative impact, up from 36% the previous month and the highest since Reuters/Ipsos began asking in March 2026; only 11% said positive. Many lacked confidence in self-regulation, most said federal officials should play a major role in safety standards, 55% said slowing development would be good (13% bad), and 73% put safe, responsible development above staying ahead (23%).

Critics cited the poll as evidence the public wanted more than a voluntary pact. Startup Fortune put it directly: “eight days later, the administration’s response to that anxiety was a pact with no enforcement role for government, no requirement that audit results ever see daylight, and no consequence if a signatory simply ignores it”. Reuters reported advocates calling the timing “an attempt to be seen addressing voter fear without touching the industry’s ability to move fast” — their characterisation. The figures explain why enforcement drew attention: confidence was low, appetite for federal standards high, and a majority favoured slowing development. For wider polling context, including the President’s approval rating and the generic ballot ahead of the 2026 midterm elections, see our report on the Reuters/Ipsos poll and the generic ballot. The debate will be settled not by argument about the text — both sides agree it contains no enforcement — but by what happens next: a named auditor, published findings, a serious incident, a state law becoming the standard, a bill with enforcement, or the reported board with powers.

Frequently asked questions

Does the AI safety pact have any enforcement?
No. Signed 29 September 2026, the accord has no penalties, no enforcement mechanism, no implementation deadline and no requirement to publish audit results. Companies choose their own auditors and address shortcomings themselves.

What do critics say is missing?
Critics — including Ruth Whittaker of Third Way, Anthony Aguirre of the Future of Life Institute and Hakeem Jeffries — say the pact lacks binding laws, mandatory pre-release testing, a power to block risky models, independent auditor appointment, publication of findings and any consequence for non-compliance. They argue existing law arrives only after harm.

What does the White House say in reply?
President Trump called the accord “morally binding”; companies must “self-police”. The Administration points to existing authorities (Department of Justice, FTC), argues binding restrictions would hand ground to China, and says cooperation with the builders is faster than new regulation.

What did the public think before the pact was signed?
A Reuters/Ipsos poll published 22 September 2026 found 73% of Americans worried AI companies had not gone far enough to prevent serious harm, with most saying federal officials should play a major role in safety standards. Fifty-five percent said slowing AI development would be a good thing. The poll surveyed 1,277 US adults, margin of error 3 points.

Updated: 28 October 2026. This article covers the enforcement debate about the 29 September 2026 accord. For the accord itself — signatories, four layers and what it does not require — see the companion explainer linked in the first section.

Sources

• CoinDesk (Associated Press reporting), “OpenAI, Google and Meta pledge outside AI audits under voluntary White House deal”, 30 September 2026 — https://www.coindesk.com/tech/2026/09/30/openai-google-and-meta-pledge-outside-ai-audits-under-voluntary-white-house-deal
• Infosecurity Magazine, “Trump, Six AI Giants Sign ‘Super Intelligence’ Safety Accord”, 30 September 2026 — https://www.infosecurity-magazine.com/news/trump-ai-giants-super-intelligence/
• ASEAN NOW, “Trump’s AI Self-Policing Deal Slammed as Unenforceable”, 2 October 2026 — https://aseannow.com/topic/1406299-trumps-ai-self-policing-deal-slammed-as-unenforceable/
• BusinessWorld (Reuters), “Three out of four Americans say AI firms not doing enough to prevent disaster, Reuters/Ipsos poll finds”, 22 September 2026 — https://bworldonline.com/technology/2026/09/24/780864/three-out-of-four-americans-say-ai-firms-not-doing-enough-to-prevent-disaster-reuters-ipsos-poll-finds/
• Startup Fortune, “Trump doubles down on voluntary AI safeguards as public fear of AI grows”, 3 October 2026 — https://startupfortune.com/trump-doubles-down-on-voluntary-ai-safeguards-as-public-fear-of-ai-grows/