Anthropic's Leaked Federal AI Bill Draws Safety Advocate Backlash

Full analysis

Anthropic has been quietly shopping a draft AI bill around Congress, and someone leaked it to Politico. The draft makes it harder, not easier, to stop a dangerous model than current law does: you'd have to petition a federal court, courts get no obligation to defer to independent evaluators, companies can self-certify their own safety with no standards, and state catastrophic-risk laws get wiped out. For anyone building on Anthropic, this is a window into how the most safety-branded lab in the business actually weighs its own risk against its own convenience. Easy to undo? The bill is a discussion draft, so nothing is law yet. But the thing it reveals about Anthropic's incentives is already baked in.

The Skeptic

The clean story is "safety company betrays safety mission." Resist it. Every AI bill in Congress right now is flawed, and the real alternative to Anthropic's draft is no federal bill and 50 state regimes pulling in different directions. The critics pointing at "existing law" are pointing at emergency-powers statutes and common-law injunction rules that were never written for AI and would produce chaos in practice. Anthropic can be protecting its own hide and writing a more coherent framework at the same time. Both true. What I won't swallow is the counterfactual that a stronger bill was sitting right there and Anthropic knocked it down. Show me that bill.

The Safety Lens

Three provisions stack into something ugly. Petition-only court access means someone has to go file a federal case to stop a model, which burns time exactly when time is the thing you don't have. No-standards self-certification means a company picks its own grader and sets its own bar, so the safety claim is unfalsifiable by design. And leaving "the harm of NOT blocking" out of what a judge weighs means courts will see the concrete commercial damage to Anthropic clearly and the diffuse catastrophic risk dimly. Here's what makes it incoherent: Anthropic itself says recursive self-improvement, AI that makes better AI, may be close. You don't write yourself a weak kill-switch while telling everyone the thing you might need to kill is nearly here.

The Researcher

The self-certification clause is the part that should bother anyone who takes alignment seriously. No external standards, evaluators chosen by the developer, no court deference to independent assessors. That guts the one thing that makes a safety claim mean anything: someone outside the company checking it. Anthropic funds real interpretability and red-teaming work. But if the legal framework never requires an auditable, third-party eval that holds up in court, that research is decorative. It produces papers, not accountability. A lab can fund alignment science and lobby to make alignment science legally irrelevant at the same time, and those two facts don't cancel out. You have to hold both.

The Compute Pragmatist

Follow the preemption clause and you find the real winners, and they aren't in the safety debate at all. California's SB 1047 and its descendants tie safety obligations to compute thresholds: train a big enough run, you inherit duties. Federal preemption of state catastrophic-risk laws, with no federal compute-threshold trigger put in its place, deregulates the training side of the stack. No state can bolt a safety obligation onto a large training contract. For the hyperscalers renting out clusters and the GPU supply chain feeding them, that's quietly great news. Anthropic's compute partners get more out of this draft than Anthropic's safety reputation does.

The Enterprise Buyer

If you're a CTO signing a frontier-model contract, this cuts two ways. Preempting the state patchwork removes a compliance headache you have today, which procurement will like. But self-certification with no standards means your vendor's safety attestation is worth whatever the vendor decides it's worth, and that becomes your problem the day a high-profile incident triggers a look back at whether that self-cert meant anything. You can't indemnify your way out of a governance regime where the grader and the graded are the same company. Build that into how much you lean on a single frontier provider.

Where they split

The Skeptic and the Safety Lens are fighting over the counterfactual, and that's the whole argument. The Skeptic says the choice is this bill or chaos, so a coherent-but-weak framework beats a 50-state mess. The Safety Lens says a weak kill-switch written by the company that admits the dangerous thing is near is worse than no bill, because it launders inaction into law. Both are looking at the same three clauses and disagreeing on what "no bill" actually costs.

The second split is about intent versus incentive. The Researcher and the Safety Lens want to read the self-certification clause as a preference Anthropic revealed. The Skeptic wants to read it as the least-bad engineering of a broken process. You can't settle that from the text. You can settle it from what Anthropic does next.

What this hinges on

One thing: does Anthropic fight to keep the self-certification-with-no-standards provision, or does it fold the moment safety advocates push? If the weak evaluator standard was an accident of drafting, Anthropic strengthens it when called out and loses nothing. If it was a preference, Anthropic defends it or lets the whole draft go quiet rather than add real third-party audit teeth. The company's revealed behavior over the next few months tells you which lab you're actually building on.

Prediction: Any public revision of Anthropic's draft AI bill released on or before 2027-04-24 will still omit a mandatory, standards-based independent evaluation requirement, keeping the developer-chosen self-certification path intact.

Confidence: Medium. The clause serves Anthropic's own interest, and companies defend the provisions that protect their shipping speed long after they drop the cosmetic ones.

Why: The leaked draft lets developers pick their own evaluators with no external standards and no court deference to independent assessors, which is the provision that makes a safety claim impossible to check from outside. That is also the provision that most directly protects Anthropic's ability to ship on its own timeline, and companies defend the clauses that protect their shipping speed long after they quietly drop the cosmetic ones. The opposite outcome, Anthropic volunteering binding third-party audit standards that could slow or block its own models right as it says recursive self-improvement is near, runs against the exact incentive the rest of the bill is built around. A company does not write itself a weak kill-switch and then hand the key to an outside grader unless forced, and nothing here forces it yet.

Revisit by 2027-04-24: We're right if any updated public version of the Anthropic-circulated bill still allows developer-selected self-certification without mandatory independent, standards-based evaluation. We're wrong if a revised draft requires binding third-party evaluation against external standards, or if Anthropic abandons the bill effort entirely with no successor draft.