What is the proposed AI industry standards body?

It is a self-regulatory organization that Anthropic, OpenAI and Google have been discussing in working-group meetings since at least July 2026, according to The Information (September 13) and CNN (September 14). The idea is an industry-run body that sets shared safety standards and runs testing and auditing of frontier models, moving ahead of government regulation rather than waiting for it. Dario Amodei has driven the push, Sam Altman backs it, and Demis Hassabis proposed a similar FINRA-style body in a July 2026 essay. No organization has been formally announced.

Why do AI labs want to regulate themselves?

Three stated reasons. Speed: U.S. federal AI legislation has stalled, and Sam Altman told an OpenAI town hall the labs would have to build a testing and auditing body themselves without government backing. Credibility: Amodei's September 12 essay argues joint audits and a shared pace are the only way to slow down without any one lab losing ground. Pre-emption: a functioning industry body gives Congress and the EU a template to recognize rather than replace, as FINRA does for broker-dealers under SEC oversight.

How would an AI standards body work in practice?

The model most cited is FINRA — a private body whose rules are binding on members, enforced with fines and suspensions, and overseen by the SEC. Applied to AI, a standards body would define evaluation thresholds for dangerous capabilities, accredit independent evaluators, require pre-deployment testing and incident reporting, and publish audit results. Anthropic, OpenAI, xAI and Google DeepMind already made embedded-evaluator commitments in early September 2026, which such a body could formalize. Membership, governance and enforcement have not been settled.

What are the criticisms?

Self-regulation lets the regulated pick the rules, the evaluators and the thresholds; critics point to the 2023 voluntary White House commitments as an example of pledges without enforcement. A lab-run body could also become an antitrust question if it coordinates release timing — OpenAI asked Congress in September 2026 whether an industry-wide slowdown would be legal. And it leaves out labs that are not at the table: Meta, xAI, Microsoft AI and Chinese developers, whose absence China's foreign ministry has already framed as fearmongering.

Quick Answer

AI Standards Body: Anthropic, OpenAI, Google Talks Explained

Published: September 15, 2026

The short answer

The three biggest U.S. frontier labs have been quietly designing their own regulator. The Information reported on September 13, 2026 — and CNN confirmed the next day — that Anthropic, OpenAI and Google have held regular working-group meetings since July about an industry-led AI standards body that would set safety standards and audit frontier models. The talks predate Dario Amodei’s September 12 “We Must Pace the Frontier” essay, which made the case publicly. Nothing has been formally created yet.

What has been reported (as of September 15, 2026)

Fact	Source
Representatives of Anthropic, Google and OpenAI have met “on a working-group basis since at least July” 2026	The Information, Sep 13; CNN, Sep 14
Dario Amodei is driving the push; Sam Altman backs it	The Information
Altman told an OpenAI town hall he favors a testing-and-auditing organization but believes labs must build it without U.S. government backing	The Information, via PYMNTS
Demis Hassabis proposed a self-regulatory body modeled on FINRA in a July 2026 essay	PYMNTS
Amodei’s Sep 12 essay calls for voluntary shared standards while governments legislate; endorsed by Altman, Hassabis and Elon Musk	Amodei blog; The Information
No name, charter, membership list or launch date has been published	—

What the body would do

Drawing on Amodei’s essay, Hassabis’s July proposal and the labs’ existing commitments, the working model looks like this:

Shared capability thresholds. Common definitions for “critical” cyber, bio and autonomy capabilities, so that one lab’s red line is every lab’s red line. (OpenAI’s GPT-6 Astra crossed OpenAI’s own “Critical” cyber threshold in September 2026; Anthropic’s Mythos line is gated by trusted-access programs.)
Independent testing. Accredited third-party evaluators with pre-deployment access. The embedded-evaluator commitments Anthropic, OpenAI, xAI and DeepMind announced in early September — with METR named as an initial evaluator — are the seed of this. See Anthropic vs OpenAI vs xAI vs DeepMind embedded-evaluator commitments.
Audit and incident reporting. Standardized reporting of serious incidents and misuse, comparable to Anthropic’s periodic threat reports but cross-lab.
Pacing. A mechanism for labs to coordinate on slowing frontier training or deployment when evaluations show risk — the most contentious piece, because it touches antitrust law.

The FINRA analogy

FINRA is the private, member-funded body that writes and enforces rules for U.S. broker-dealers, with the SEC approving its rules and overseeing it. Hassabis’s July essay proposed the same structure for AI: an industry body whose standards are binding on members, with a government regulator as backstop. The appeal is that it can be stood up in months rather than the years federal legislation takes, and it produces something Congress can later recognize in statute. The weakness is that FINRA exists because Congress required registration; an AI body with no statutory hook has only reputational leverage over members and none over non-members.

Why now

Federal legislation is stalled. Congressional leadership split over whether to act after the Amodei essay; see Will Congress regulate AI after the Amodei essay?.
The White House is not asking for a slowdown. President Trump said Americans should not be scared of AI and Vice President Vance called for “smart regulation over fear” on September 14, 2026.
The models crossed thresholds. Astra’s Critical cyber rating, Anthropic’s disclosure of four Claude cyber incidents and DeepMind’s 100-agent swarm paper all landed within weeks of each other.
Antitrust cover. OpenAI’s September letter asking Congress whether an industry-wide slowdown would be legal is best read as a request for the safe harbor a formal standards body would need.

Who is not in the room

Meta (Muse Spark), xAI (Grok 4.6), Microsoft AI (which published its own Humanist AI Code of Conduct on September 14) and every Chinese lab — DeepSeek, Moonshot, Z.ai, MiniMax, Alibaba. China’s Ministry of Foreign Affairs called Amodei’s essay “fearmongering” that would disrupt global AI governance. A U.S.-only, three-member body can set norms for the models most people use, but it cannot pace a frontier it does not fully contain.

The criticisms, briefly

Fox and henhouse. Members would write the thresholds, pick the evaluators and decide what is disclosed. The 2023 voluntary White House commitments are the cautionary precedent.
Antitrust exposure. Coordinated release timing among competitors is exactly what Section 1 of the Sherman Act targets unless a regulator sanctions it.
Two-tier frontier. If the body works, it may advantage incumbents who can afford compliance and audits over open-weight and smaller labs.
Substitution risk. A credible-looking industry body could reduce pressure on Congress to legislate at all.

What to watch

A formal announcement with a charter and a named executive — the signal that talks have become an organization.
Whether Meta, xAI and Microsoft AI join or launch a rival framework.
Whether Congress or the EU AI Office references the body in statute or Code-of-Practice recognition.
The first cross-lab audit report.