AI giants are asking Washington to regulate them. They also stand to gain from it

An agreement that looked impossible two years ago

OpenAI, Anthropic and Google DeepMind increasingly agree that the United States needs federal rules for its most capable AI systems. Axios documents the shift after years dominated by voluntary commitments.

Shared evaluations and incident reporting sound straightforward. The difficult questions are which models qualify, who tests them, what government can inspect and whether any authority can delay a profitable launch.

Regulation can protect incumbents too

Global companies can afford lawyers, expensive evaluations and sustained lobbying. Start-ups and universities may not. Flat requirements can turn safety into a barrier to entry, so obligations should scale with capability, deployment and access rather than creating a general licence to research.

Federal consistency or a federal ceiling?

Companies want one federal framework rather than 50 state regimes. Consistency has real value, but a weak federal law could also erase stronger local protections. The key is whether Washington sets a floor states can build on, or a ceiling they cannot exceed.

What credible rules require

Independent auditors need technical access, incident definitions need teeth and whistleblowers need protection. Company-prepared paperwork is not enough. Frameworks such as Google DeepMind's show that labs can measure some risk; the test is what happens when an evaluation says a lucrative product should wait.

Consensus can become public infrastructure or self-regulation with better branding. We will watch the legal text, thresholds, regulator budgets and exemptions. If only the three companies requesting the rule can afford to comply, it may organise the market more than protect the public.