Rivals as Referees: Anthropic, OpenAI and Google Have Spent Months Quietly Designing an Industry-Run AI Safety Body — and Musk Wants to Make It Mutual

While Washington argues about whether AI risk is real, the three labs that actually build frontier models have been quietly building the industry’s answer — a shared, industry-run standards body that would test advanced models before they ship.

According to a report by The Information published over the weekend and corroborated by The Washington Post on Monday, Anthropic, OpenAI and Google have been holding working-group meetings since July to design an industry-led body that would set common safety standards for frontier AI, run pre-deployment testing of the most advanced models, and coordinate on risks that no single lab can see alone. Demis Hassabis, Google DeepMind’s CEO, has argued the body should test advanced models before deployment and take a public-private form, according to accounts of the discussions.

The timing is not accidental. The talks surfaced into public view in the same week that Anthropic CEO Dario Amodei published his call to “pace the frontier,” resignations from frontier safety teams made headlines, and President Trump — phone call to Jensen Huang and all — declared the entire safety discussion “a hoax.” The industry body idea is the labs’ bet that if government will not regulate, and the White House is actively hostile to the notion, the only credible referee left is the industry itself.

What the body would actually do

Details remain fluid, but the reported scope is more concrete than a talking point:

• Shared pre-deployment standards. A common bar for what testing a frontier model must pass before release, so that safety cannot become a competitive race to the bottom between labs.
• Testing of advanced models before deployment. Hassabis in particular has pushed for the body to evaluate frontier systems pre-release — the step that today happens only inside each lab, behind closed doors.
• Third-party and cross-lab evaluation. Reports mention testing third-party models for prompt injection and other robustness failures — echoing the kind of adversarial probing labs rarely do on each other’s systems today.

There is precedent, and it is recent. In August 2025, OpenAI and Anthropic published findings from a first-of-its-kind joint alignment evaluation, testing each other’s models for misalignment and failure modes. That pilot proved cross-lab review is technically feasible; the proposed body would institutionalize it.

Musk’s version: rivals as referees

Enter Elon Musk. At the All-In Summit in Los Angeles this week, alongside SpaceX President Gwynne Shotwell, Musk revived and sharpened a proposal he first floated in a July interview with The Economist: the leading AI labs should peer-review each other’s frontier models before release.

Musk’s framing, as reported by Fortune and others, has two notable edges. First, he argues no lab will actually slow the pace of capability research — but they will withhold new capabilities from the public, which is where review should bite. That is a subtler position than a blanket pause, and closer to what Amodei’s “pacing” essay actually proposed. Second, Musk frames the whole thing as industry self-policing: government should step in only as a last resort if labs fail to hold each other accountable.

It is a striking posture for a man who runs xAI, competes directly with all three labs in the talks, and once warned AI could “crush all humans.” It also aligns neatly with the political moment — White House AI adviser David Sacks told CBS this week that safety is the labs’ job, not the government’s, and House Speaker Mike Johnson said industry leaders should lead. Peer review is the one version of “the industry will handle it” that comes with a mechanism attached.

The counterweight: Nvidia, Broadcom, and “nothing to slow down”

Not everyone is on board. An industry peer-review regime would, by design, add friction at the exact point where the hardware layer is monetizing speed. At the same summit, Nvidia and Broadcom executives dismissed the idea that AI poses any near-term threat and signaled that demand for AI infrastructure remains durable regardless of the slowdown debate, Yahoo Finance reported. Their position is consistent with commercial reality: Nvidia’s guidance implies roughly 70% revenue growth next year, and a genuine capability pause would land directly on that line.

The split is now unusually legible. The model layer — Anthropic, OpenAI, Google DeepMind, and even Musk — is converging on some form of mutual restraint and mutual inspection. The compute layer — Nvidia, Broadcom — sees no reason for any of it. And the political layer has picked its side loudly, with Trump framing the AI race as one to win, not to pace.

Will it work?

An industry-led standards body has one structural strength and one structural weakness, and they are the same fact: the members are competitors.

The strength is incentive. Every lab in the talks has publicly endorsed some form of pacing — Amodei’s essay, Musk’s “Dario is right,” Altman’s agreement — which means each wants assurance that restraint will be symmetric. Nobody wants to be the only lab holding a capable model back while a rival ships. A shared body with real testing authority is the only mechanism that makes restraint verifiable rather than trusting.

The weakness is enforcement. A body the labs build themselves can set standards, run evaluations, and publish results — but it cannot compel a member to delay a launch, and it has no obvious answer for a capable lab that simply declines to join, or a state actor that never intended to participate. The Information noted that the labs involved scored between 29 and 42 out of 100 on an independent safety index earlier this year; the body is partly an attempt to fix that credibility problem from the inside.

There is also the antitrust question, which has barely been aired: an industry body that coordinates on release timing and capability thresholds sits uncomfortably close to coordination on the industry’s most commercially sensitive decisions. Structuring it to evaluate safety without touching commercial pace will be the design challenge of the year.

What to watch

Three signals will tell you whether this is real. First, whether xAI and Meta — both absent from the reported talks — join or are invited. Second, whether the body’s evaluations are ever made public, or whether they remain confidential like today’s internal red-teaming. Third, whether Congress, currently in a standoff over AI guardrails legislation, treats an industry-run body as an adequate substitute for statute or as evidence the labs cannot police themselves.

The labs have spent two months designing a referee. Whether the referee ever gets a whistle is now the most consequential open question in AI policy.