OpenAI, Anthropic and Google negotiate a joint AI standards body — as Trump attacks regulation and safety divides the industry

OpenAI, Anthropic, and Google negotiate a joint AI standards body — self-regulation put to the test by Trump and mistrust

🔎 The three rivals who can no longer afford to ignore each other

According to information reported by The Information via WeSearch on September 14, 2026, OpenAI, Anthropic, and Google have been holding working group meetings since July. The stated goal: the creation of an industry-led AI standards body tasked with establishing unified guidelines for the development and deployment of advanced models.

The timing is no coincidence. On Saturday, September 12, Dario Amodei issued a public call for cooperation on auditing and testing — in the wake of his essay "We Must Pace the Frontier". Sam Altman responded as early as the following week, at an internal town hall, that he supported a testing and auditing body. Demis Hassabis, for his part, had already proposed in July a self-regulated body modeled on FINRA, the American financial watchdog.

But the political context is explosive. Trump has sharply attacked calls for greater regulation, as well as Amodei himself. Mike Johnson, Speaker of the House, declared that same week that Congress would not lead on AI safety — the priority being not to lose the race against China. The three labs are therefore negotiating a safety pact in a country whose executive branch wants no part of it.

The Essentials

• Confirmed: Anthropic, OpenAI and Google have been holding working group meetings since July 2026, made up of executives below CEO level, to discuss the creation of an industry AI standards body (Seoul Economic Daily).
• Unconfirmed: nothing has been created. No charter, no finalized governance, no official announcement. These are discussions.
• Three visions are clashing: Amodei wants a "FAA for AI" (a federal agency able to block a release), Hassabis a "FINRA for AI" (an industry body overseen by the federal government), Altman an "IAEA for AI" (an international forum) (InvestingLive).
• The operational scope under discussion remains narrow: pre-release third-party evaluations (roughly 30 days before deployment according to Hassabis's proposal), formal safety checks, unified protocols for critical risks (TechRound).
• The underlying question: is a body designed and funded by the very companies it is meant to audit, in a political climate hostile to any regulation, a guarantee or a shield?

Recommended Tools

A standards body will serve no purpose if developers don't know what they're deploying. Until unified protocols arrive, the bare minimum is to test the models yourself — including via free endpoints.

What do we actually know? Separating facts from speculation

What is confirmed: discussions, nothing more. Three corroborating sources — The Information, PYMNTS, and the Seoul Economic Daily — describe regular meetings since July 2026. They involve executives below the CEO level. Amodei, Altman, Hassabis, and even Musk have endorsed the effort at the level of principles.

What is not confirmed: everything else. No charter. No legal structure. No timeline. TestingCatalog explicitly reminds us of this: the discussions are still ongoing, with no finalized governance. Presenting this as an "agreement" would be a distortion — and it is precisely the kind of distortion this industry loves to inflict on itself.

The sequencing is nonetheless revealing. According to PYMNTS, the working group discussions began before Amodei's public call on September 12. In other words: private coordination already existed, and the weekend of public statements merely made it visible. This is not a change of course. It is a staging of a process already underway.

Three visions of regulation that are not compatible

Beneath the apparent facade of unity, the three labs are pushing radically different architectures. This is the most interesting point of the story — and the one that most coverage glossed over.

Amodei: the "FAA for AI." A federal agency with real power, able in theory to block the release of a model deemed dangerous. This is the most interventionist vision — and paradoxically the one most likely to displease the current White House. Anthropic leans overall toward government partnership.

Hassabis: the "FINRA for AI." An industry-funded body overseen by the federal government, modeled on Wall Street's self-regulation. According to CNN Business, his July essay described a public-private partnership staffed with independent technical experts and open source representatives, testing advanced models before deployment. The stated trajectory: from voluntary to mandatory.

Altman: the "IAEA for AI." An international forum certifying countries, companies, and standards. This is the most abstract vision — and the least binding in the short term. OpenAI leans toward pure industry voluntarism, believing the major labs will have to create the body themselves without support from the US government.

These three models don't diverge over details. They diverge over the central question: who has the power to say no? An FAA can say no to a release. A FINRA can escalate to a regulator. An IAEA publishes reports. These are not three versions of the same body — they are three different relationships to authority.

Why now? The political and financial context

Three converging pressures explain this communication and negotiation offensive.

Political pressure is going down, not up. A draft executive order for an AI standards body was indeed prepared at the White House, but it ran into internal conflicts — David Sacks, co-chair of PCAST, opposing any government-led body (Seoul Economic Daily). Meanwhile, Mike Johnson made it clear that Congress will not legislate on AI safety, explicitly handing responsibility back to the labs. The message: figure it out on your own, but don't slow us down.

The markets struck back. Several tech stocks dropped after AI CEOs' public warnings about the risks. When your safety warnings cost your investors stock points, you'd better channel the message. A formal body makes it possible to say "we take the problem seriously" without multiplying individual statements that move the markets.

The security pressure is real and documented. The unprecedented collective call signed by OpenAI, Google, Microsoft, Anthropic and more than 100 companies about the imminence of a wave of AI cyberattacks has shifted the debate: the risks are no longer theoretical. The scope discussed for the body — cyberattacks, bioweapons, deceptive behaviors — reflects exactly this threat hierarchy (TechRound).

Add a human factor: the war for talent is raging, as shown by the ongoing exodus at Google DeepMind, with Nobel laureate John Jumper leaving for Anthropic. Organizations tearing each other apart over recruitment have every interest in presenting a united front on safety.

The precedent that should make us wary: the Frontier Model Forum

This is not the first time these players have announced a coordination effort. The Frontier Model Forum, created in 2023 by the same players, already had a founding mission that mentioned standards. It gave rise to the AI Safety Fund, endowed with more than $10 million (InvestingLive). More recently, the Linux Foundation's Agentic AI Foundation (December 2025) brought together Anthropic's MCP, Block's Goose, and OpenAI's AGENTS.md.

The analysis by Santage is scathing and worth quoting: the real question is not whether a new body is created, but what more it would actually impose. The industry's current self-assessment results do not make the case for it. In the FLI AI Safety Index 2026:

Three below-average grades for the three presumptive founders of the future standards body. One can debate FLI's methodology, but one cannot ignore the signal.

That same month, Microsoft — via Satya Nadella — endorsed a purely voluntary pacing framework. When the entire industry converges on voluntary commitments at the very moment its own independent assessments judge it insufficient, the coordination looks more like narrative control than risk control.

Shield or safeguard: the question that will decide everything

This is the heart of the problem, and it has no technical answer.

The conditions for a credible safeguard. For a lab-designed body to be a guarantee rather than window dressing, three conditions are necessary according to Santage: an evaluator truly independent from the funding labs, concrete consequences in the event of a failed audit, and a mandate resistant to commercial pressure. None of these three conditions appears in what is known of the discussions.

The contradictory signals. On one side, Amodei is pushing for benchmarks aligned across labs and — a crucial detail — official antitrust protection for safety coordination (TechRound). Requesting antitrust immunity to coordinate three actors that dominate the market is the most interesting request in the whole file: it shows that the labs know their cooperation brushes against the legal boundary, and it would set a formidable precedent if granted without verifiable reciprocal commitments.

On the other, Hassabis's concrete proposal — pre-release evaluations roughly 30 days before deployment — has an obvious flaw: who tests, with what criteria, and what happens if the test fails? A test 30 days before a marketing release already underway is not an audit. It's a formality.

The European angle changes the game. The AI Act already imposes binding evaluations for models with systemic risk. In Europe, a voluntary American body would be a mere regulatory complement. In the United States, where the regulatory vacuum is embraced, it would become a proxy-regulator — with all the power asymmetry that implies.

A truce born of necessity, not trust

The most striking thing about this story is who's talking to whom. Anthropic and OpenAI are avowed rivals who publicly attacked each other before their respective IPOs. According to the Seoul Economic Daily, the two companies concluded that the truce and a voluntary body mattered more than their dispute.

This forced reconciliation recalls a classic dynamic: when the regulator steps back, dominant players regulate the market among themselves. That's sometimes better than nothing. It's rarely neutral. Anthropic already proved it with the acquisition of Stainless for over 300 million dollars — a deal that tightens Anthropic's control over SDK access infrastructure, while OpenAI and Google lose a channel in the process. Cooperation on standards and competition over infrastructure are moving forward in parallel, and the latter won't wait for the former.

We should also remember where the internal pressure comes from. No lab has solved the alignment problem — and the calls to slow down come from the labs themselves. This new initiative is the institutional extension of an admission: the race has spiraled to a point that its own participants find uncomfortable. The resignation of an Anthropic researcher claiming that labs don't "behave responsibly" (CNN Business) shows that the admission is also coming from within.

Then there's the AI For Good conference (September 15-17), which is simultaneously exploring practical uses of AI. Two narratives coexist: useful, deployable AI, and dangerous AI that needs to be reined in. The standards body is where these two narratives will have to reconcile — or clash.

❌ Common Mistakes

Mistake 1: Confusing "discussions" with "creation"

The working groups have existed since July 2026. The organization itself does not. No charter, no governance, no name, no timeline. Any analysis that presupposes the organization as a given starts from a false factual basis. The correct reading: what is being negotiated right now is precisely the scope and authority — not the implementation details.

Mistake 2: Believing that self-regulation equals regulation

An organization funded by its own auditees, with no legal mandate, and in a political climate where Congress refuses to act, has no means of enforcement. At best, it standardizes security reports and creates a common language. At worst, it provides a hollow line of defense ("we are audited"). Judge it on its ability to say no, not on its existence.

Mistake 3: Ignoring the requested antitrust protection

This is the most under-covered element of the story. Without antitrust immunity, this coordination between three dominant players is legally fragile. If it is granted without strict transparency and independent evaluation requirements, it sets a precedent whereby cooperation between competitors becomes unchallengeable under the guise of security. This is a competition policy issue as much as a security one.

❓ Frequently Asked Questions

Has an AI standards organization already been created?

No. As of September 14, 2026, these are only working group discussions among executives below CEO level, reported by The Information. No charter, no finalized governance, no official announcement. The discussions are confirmed by several corroborating sources, but their outcome is not.

What power would this organization have?

That's exactly what is being negotiated. Three visions coexist: a federal agency with the power to block releases (Amodei), an industry body supervised by the federal government (Hassabis), an international certification forum (Altman). Without accepted government involvement — and the White House is hostile to it — the enforcement power would likely remain limited.

Why are the three labs suddenly cooperating?

Three converging reasons: a political context in which Congress refuses to legislate, leaving a void to fill; market pressures following the stock market drops linked to the CEOs' warnings; and security risks that have become concrete, particularly in cybersecurity. The discussions began before Amodei's public call.

How does it differ from the European AI Act?

The AI Act already imposes binding evaluations on systemic-risk models in Europe. A voluntary American body there would simply be a supplement. In the United States, in the absence of federal legislation, it would become a de facto proxy regulator — which radically changes its scope and its risks.

Are the covered risks well defined?

The scope under discussion is narrow and operational: third-party evaluations before release, formal safety checks, unified risk protocols across three categories — cyberattacks, bioweapons, deceptive behaviors. Hassabis's proposal mentions evaluations roughly 30 days before deployment. What's missing: the pass criteria and the consequences in case of failure.

✅ Conclusion

Three rivals who have never cut corners are negotiating a common testing and auditing mechanism — a sign of dawning awareness, but the real value of this mechanism will hinge on one thing only: the authority Washington is willing to delegate to a body designed by the very people it will have to audit. Until unified rules arrive, vigilance remains an individual matter: test the models yourself and entrust your judgment neither to the labs nor to their future watchdogs.