Sarah Heck, Anthropic’s head of public policy, said on September 16, 2026, that advanced-AI companies should not be left to assess their own safety without government involvement and outside evaluation. The position puts Anthropic on one side of a widening policy argument: who should check frontier AI—the companies building it, independent evaluators, government, or some combination of all three?
The dispute is separate from the legal status of the FRONTIER Act, H.R. 9925. The bill was introduced on July 23, 2026, and proposes a federal oversight framework, but its provisions remain proposals rather than current law.
Sarah Heck’s position: companies should not grade their own homework
Heck said Anthropic supports a government role in creating an appropriate regulatory framework for advanced AI. She also backed third-party evaluators in principle, arguing that the companies developing powerful systems should not be the only parties checking their safety work.
That position does not amount to an endorsement of every provision in H.R. 9925. Anthropic was described as working with lawmakers on the right framework, while Heck stopped short of immediately supporting the bill’s specific mandatory outside-evaluator requirement.
Anthropic CEO Dario Amodei has expressed a similar preference for public involvement. In a September 2026 interview, he said government and the public should have a stake in AI oversight and described some form of “joint governance” as a long-term possibility. He distinguished that arrangement from handing Anthropic over to the government.
What the FRONTIER Act would do
H.R. 9925 would establish graduated obligations for large and very large frontier-AI developers. Under the bill’s proposed definition, a frontier model would be a foundation model trained with more than 10²⁶ integer or floating-point operations, including specified later modifications.
The proposal would cover a large developer with more than $50 million in gross revenue and at least $1 billion in AI-related development expenditures over the preceding 36 months. A very large developer would cross the higher thresholds of more than $5 billion in gross revenue and at least $10 billion in AI-related development expenditures over that period.
For covered companies, the bill proposes:
- A published frontier-AI framework addressing risk thresholds, assessments, deployment decisions, cybersecurity, incidents and governance.
- An annual independent compliance audit for large developers, with a redacted report and summary made public.
- Assessments by licensed independent verification organizations at least every six months for very large developers.
- Reports on critical safety incidents, generally within 72 hours after the developer has enough facts to reasonably believe an incident occurred.
- A 24-hour notification to law enforcement when an incident involves imminent death or serious injury.
- Temporary Commerce Department orders restricting development, deployment or internal use after a written finding of imminent catastrophic risk, subject to the bill’s procedures.
The proposed penalties would reach $1 million per violation for several covered obligations and $10 million per violation for violating an emergency order.
None of those mechanisms currently applies as federal law through H.R. 9925. The bill is listed as introduced and referred to House committees.
Two competing models for AI safety
The core disagreement is not simply regulation versus no regulation. It concerns who gets authority to evaluate safety, whether laboratories should coordinate their pace, and how much power government should have when a system is considered dangerous.
| Policy question | Anthropic’s reported position | Meta’s reported position |
| Who should assess safety? | Companies should not be the only assessors; Anthropic supports third-party evaluators and a government role. | Each laboratory should take responsibility for ensuring its own models are safe. |
| Should development slow down collectively? | Dario Amodei proposed pacing frontier-AI development so safety measures can keep up. | Mark Zuckerberg opposed a coordinated industry-wide slowdown and emphasized laboratory-by-laboratory responsibility. |
| What should government do? | Work with companies and lawmakers on a regulatory framework with public oversight. | Allow each laboratory to take its own safety actions rather than relying on a coordinated industry-wide pause. |
Zuckerberg has argued that every laboratory has both the responsibility and the incentive to move at the pace required to train its models safely. Meta’s position also includes taking action on individual products: Zuckerberg cited the company’s delay of its Muse AI agent while it worked on safety and security.
The result is a policy split over coordination, not a documented agreement among the industry’s major companies. The positions described here do not amount to a binding industry-wide slowdown or safety pact.
Amodei’s proposal adds a question about pace
The debate over outside oversight follows a September 12 proposal from Amodei calling for a slower pace of frontier-model development so safety measures can keep up. The proposal included ongoing access for third-party evaluators, coordination among democratic countries and broader international coordination.
That idea has drawn different responses inside the AI industry. The reported positions include support for greater pacing from OpenAI CEO Sam Altman and Elon Musk, while Jensen Huang opposed an industry-wide slowdown and supported safety pauses for individual products. Zuckerberg also rejected a coordinated slowdown.
For Anthropic, the practical message is that frontier developers should not be the sole judges of their own safety. For Meta, the responsibility remains with each laboratory, which can delay or adjust its own work when necessary. H.R. 9925 would move that argument into a formal legal structure if Congress enacted it, including audits, recurring independent verification, incident reporting and emergency authority.
