AI Agent UpdateAnthropic's leaked safety bill draws fire from advocates — and OpenAI
A leaked draft of Anthropic's proposed federal AI safety bill sets the intervention bar at $1 billion in damages or 50 casualties, with licensed evaluators and a Commerce-only path to court. Safety advocates say it doesn't go far enough and pre-empts the states. OpenAI says no one company should write the rules. Both arguments are really about the moat.
The News
A leaked draft of an Anthropic-backed federal AI safety bill, proposed to Capitol Hill lawmakers and originally reported by Politico, has sparked backlash. An Anthropic spokesperson confirmed the draft is a legislative version of a framework the company has pitched for months.
The draft's thresholds are deliberately high. Washington could block a model only if it poses a foreseeable, imminent risk of more than $1 billion in damages, or could kill or seriously injure more than 50 people. Independent evaluators registered with the Commerce Department would periodically review frontier models, and labs could pay those evaluators provided compensation isn't tied to the outcome. Any “imminent enumerated risk” must be reported within 72 hours, and Commerce could seek a judge's order to suspend a model — while other federal agencies would be barred from doing so.
Safety advocates argue the draft doesn't go far enough and unfairly pre-empts state laws — trading away state-level regimes for a federal framework whose tripwires only catch catastrophes already in motion.
OpenAI fired back publicly. Chief global affairs officer Chris Lehane wrote Monday: “Policy decisions must ultimately be made through our democratic processes with our elected officials leading... No one company (including OpenAI) should write the rules alone for a technology with such profound implications.”
A safety regime built on licensed evaluators and federal courtrooms is a cost big labs can absorb and startups can't — that's not a bug in the draft, it's the moat.
Why It Matters
If anything like this draft becomes the federal baseline, it rewrites the compliance map for every frontier lab: a registered-evaluator industry appears overnight, 72-hour risk reporting becomes a legal function, and state AI laws — the patchwork labs currently navigate — risk being pre-empted by a single federal ceiling.
For the listed backers and rivals — $MSFT and $GOOGL on the investment side, $META and $NVDA in the blast radius — the direct cost is trivial relative to their scale. The real effect is competitive structure: regimes built on licensed evaluators and federal court processes are fixed costs, and fixed costs are rounding errors for giants and existential for startups.
Analyst's Take
Here is the angle the outrage cycle is missing: a regime built around registered evaluators and federal court oversight is far easier for large, well-capitalized labs to absorb than for startups. Compliance departments, evaluator relationships, legal teams that can litigate a suspension order — Anthropic, OpenAI, Google and Microsoft all have them; a twelve-person lab has a lawyer on retainer. Regulation of this shape does not restrain the incumbents; it licenses them. Regulation as moat is the oldest play in the Washington book, and the draft's authors know it.
Which is also why OpenAI's objection should be read as positioning, not principle: days after pausing its own training over agent incidents, OpenAI benefits from a debate about process rather than a bill landing. The tell to watch is the pre-emption clause — if the final text keeps federal supremacy over state laws while raising no thresholds, the big labs will quietly stop objecting, whatever their public statements say. That silence will tell you who the bill actually serves.
By The Numbers
FAQ
What does Anthropic's draft AI safety bill propose?
A federal regime where the government could block a model only for foreseeable, imminent risk of over $1 billion in damages or 50+ deaths or serious injuries; periodic reviews by independent evaluators registered with the Commerce Department; 72-hour reporting of imminent enumerated risks; and suspension only via Commerce asking a judge — with other agencies barred.
Why are safety advocates upset?
They say the thresholds are so high the law would only trigger on catastrophes already in motion, and that the draft pre-empts stronger state AI laws — replacing a patchwork with teeth with a federal ceiling without them.
What did OpenAI say about the bill?
Chief global affairs officer Chris Lehane wrote that policy decisions “must ultimately be made through our democratic processes with our elected officials leading” and that “no one company (including OpenAI) should write the rules alone.”
Who benefits if the draft passes?
Large, well-capitalized labs. Licensed evaluators, reporting infrastructure and federal litigation capacity are fixed compliance costs that giants absorb easily and startups cannot — an advantage critics describe as regulation functioning as a moat.
Source: Politico via Zot News · Adalytica