Situation report: the three biggest frontier labs have been talking to each other about who sets the safety rules. Anthropic, OpenAI and Google quietly discussed forming an AI safety standards body, according to The Information. The talks happened out of public view, and The Information reports them as discussions, not a signed deal.
That’s the whole of what’s confirmed. But the fact that these three sat down at all tells you plenty about where the industry thinks regulation is heading.
Threat Assessment
The threat here isn’t the standards body. It’s the vacuum it would fill.
Right now, frontier AI safety runs on a patchwork. Each lab publishes its own framework. Anthropic has its Responsible Scaling Policy. OpenAI has its Preparedness Framework. Google DeepMind has its Frontier Safety Framework. They overlap, they diverge, and nobody outside the labs can easily compare them.
Governments haven’t closed the gap. The EU AI Act is rolling out in phases. California passed its own frontier model law. Washington has pushed for lighter federal rules and pushed back on state-level action. Companies now face a mess of overlapping and sometimes conflicting expectations.
When regulators don’t standardize, industry tends to do it for them. That’s what this looks like.
Tactical Points
- This isn’t the first attempt. Anthropic, Google, Microsoft and OpenAI launched the Frontier Model Forum in 2023 as a voluntary safety group. If a new body is on the table, it suggests the existing forum hasn’t delivered the shared benchmarks or enforceable norms the labs now want. A standards body implies something more formal: agreed tests, agreed thresholds, maybe third-party checks.
- The timing tracks the regulatory climate. With federal policy leaning toward deregulation and states filling in the gaps, a self-governing body gives the labs a way to say “we already have standards” when lawmakers come knocking. That’s a defensive move as much as a safety move.
- Competitors cooperating is the signal. These companies fight over talent, compute and customers every day. When rivals coordinate on safety, it usually means they see a shared risk that outweighs the competitive edge of going alone. A single high-profile model incident would hit everyone’s licence to operate.
- “Quietly” matters. The Information’s framing suggests the labs didn’t want this public yet. That could mean the talks are early, or that they’re sensitive about antitrust optics. Three dominant players agreeing on rules together draws scrutiny fast.
- The standards fight is a market-share fight. Whoever writes the tests defines what “safe enough to ship” means. Smaller labs and open-weight developers would have to meet a bar they didn’t set. That’s a real competitive lever dressed up as governance.
Why Practitioners Should Care
If you build on these models, the standards these labs agree on become your constraints. Expect them to show up as:
- Stricter model cards and capability disclosures you’ll need to read before deployment
- Shared evaluation categories (bio, cyber, autonomy) that shape what gets released and when
- Possible staged rollouts where new capabilities gate behind safety sign-offs
- Procurement checklists from enterprise buyers that reference whatever the body publishes
None of that is bad. Consistent rules beat guessing which lab’s framework applies to your use case. But it does mean the labs, not your regulator, will write the first draft.
What Stands Out
What stands out to me is who’s missing from The Information’s report. No Meta. No xAI. No Chinese labs. A standards body that only covers three companies isn’t an industry standard. It’s a club. Whether it grows into something broader, or stays a coordination channel among the top three, decides whether this matters for the whole field or just for them.
The other open question is teeth. Voluntary frameworks have a habit of bending when a competitor ships first. If this body can’t enforce anything, it’s a press release with a logo. If it can, it’s the closest thing to a frontier AI regulator that currently exists.
Next Watch
Watch for a formal announcement, a named executive director, or a published evaluation suite. Any of those turns a quiet conversation into a real institution. Until then, treat this as an early signal: the labs know rules are coming, and they’d rather write them than read them.
The Information has the full account of the discussions.