Story
Oct 20, 2026
Anthropic’s Safety Bill Faces a Familiar AI Fight: Who Sets the Rules?
Anthropic’s draft federal AI safety framework would impose reporting and testing obligations while limiting intervention to catastrophic threats. Supporters call it a practical national standard; critics say labs would retain too much control over their own oversight.
Anthropic has quietly brought a draft federal AI-safety bill to Capitol Hill, entering a crowded debate over how aggressively Washington should police powerful models — and who should write the standards in the first place.
The proposal would require frontier labs to build safety frameworks, publish transparency reports, protect whistleblowers and report critical incidents to the Commerce Department. It would allow the government to seek a court order blocking a model only when there is a foreseeable, imminent risk of more than $1 billion in damage or at least 50 deaths or serious injuries, tied to biological weapons, loss of developer control or cyberattacks. “These are sensible policies we believe Congress should act on,” an Anthropic spokesperson said.1
Its enforcement design is deliberately narrow. Registered independent evaluators would periodically assess labs and models, and must alert Commerce within 72 hours if they find an imminent specified risk. But the evaluators can be paid by the companies they examine, provided payment is not contingent on the result; other federal agencies would be barred from suspending a model’s development, deployment or internal use. The draft would also override some state rules on disclosures, evaluations and critical-incident reporting.1
That balance has drawn opposition from a different direction. One account of the leaked proposal argues that, unlike the FAA-style approach Anthropic CEO Dario Amodei previously invoked, “the labs get to write the standards that they must then comply with.” Safety advocates want government, not industry, to set them.2
David Sacks, meanwhile, warned that the bill’s reach could grow well beyond the elite developers it purports to target. Some provisions “could be reasonable if limited to true frontier companies,” he wrote, but as drafted it could “eventually sweep in most of Silicon Valley,” hurting startups.
3
The clash comes as Congress weighs rival bipartisan AI-safety bills and as lawmakers dispute whether a national framework should displace tougher state laws. Anthropic’s draft attempts to occupy the middle: mandatory procedures for labs, but government intervention reserved for catastrophe. Its critics see that middle ground as industry self-regulation in legislative clothing.