Anthropic Drafts AI Regulatory Bill Critics Call Dangerously Weak

Full analysis

Anthropic, the lab whose CEO writes essays about AI ending humanity, is quietly circulating a draft federal AI bill that gives regulators less power than they have right now. The draft reportedly sits below the June 2024 FRONTIER Act on rulemaking authority, emergency shutdown, and auditor independence, and it wipes out state AI laws like California's SB 53 through preemption. For anyone building with or buying from frontier labs, the question is whether a weak federal ceiling that kills the state patchwork helps you or sets you up for a worse fight later.

How hard is this to undo? Hard. Federal preemption that erases state law is close to permanent. Once Congress passes a ceiling, the only way up is another act of Congress, which almost never happens absent a disaster. A bill passed in a lame-duck session cements the floor for years.

What's actually being decided: Not "should AI be regulated" but "who holds the one lever that can stop a frontier model before it ships, and how high is the bar set." The draft sets that bar at imminent risk of 50+ deaths or $1B+ in damages, with no deference to regulators' judgment.

What sets the deadline: The lame-duck congressional session. Zvi Mowshowitz's worry is timing: pass this now, before SB 53 and other state laws harden, and you lock in the weak version.

The Skeptic. Anthropic wrote the safety sermon and is now shopping the loophole. That gap is not an accident, so stop asking whether it's hypocritical and ask what it reveals. The cleanest explanation: Anthropic's policy team fears executive-branch action more than it fears a bad model. A hostile administration that can shut down a training run on short notice is an existential business risk. Fifty-death thresholds and no regulatory deference read as a company buying insurance against its own government. The safety branding keeps funding and recruiting alive. The bill protects the balance sheet. Both can be true at once, and the people who wrote each one may never have been in the same room.

The Safety Lens. A low federal ceiling plus preemption of the states is the worst available structure. The "prove 50 deaths first" standard institutionalizes the one thing that makes catastrophic AI risk hard to govern: you cannot run the counterfactual, so imminent-harm evidence arrives after it's useless. Auditor capture without government licensing is the pre-2008 ratings-agency setup, where the thing being rated picks and pays the rater. And stripping emergency shutdown authority that survives a court challenge removes the mechanism you'd want most at the worst moment. This makes the US posture structurally worse than passing nothing. The optimistic assumption buried inside it, that a better bill comes later, is exactly what preemption forecloses.

The Researcher. The 50-death, $1B threshold sounds studied. It is an arbitrary anchor dressed as rigor. Frontier risk is low-probability, hard to attribute, and non-linear, which is precisely the profile this bar cannot catch, because by the time harm is "imminent" and measurable, intervention is moot. No regulatory deference means even a well-founded expert judgment gets second-guessed in court against a near-unfalsifiable standard. The auditor gap compounds it. Without licensing, third-party evaluation becomes a market where labs shop for the friendliest grader, and the external check that researchers rely on to verify a lab's safety claims quietly disappears.

The Enterprise Buyer. Here is the uncomfortable part: this bill is attractive to the CTO signing the contract. One federal rule beats fifty state regimes your compliance team can't track. SB 53 and its cousins are a genuine cost, and preemption makes them go away. But buyers deploying agents into healthcare, finance, and anything touching a regulated workflow should want the opposite of what this bill offers. Credible, licensed audit infrastructure is what lets you tell your own regulator and your own board that the model you bought was independently checked. Strip that out and every incident lands on the buyer with no paper trail to hide behind. Cheap compliance now, uninsurable liability later.

The Builder. The 90-day read is that nothing changes on Tuesday and most labs quietly like it. No new eval gates, no pre-deployment checks with teeth, no shutdown switch hanging over a launch. The second-order read is uglier. Pass something this weak and the first serious AI incident triggers emergency legislation written in a panic, and panic laws are always worse than negotiated ones. A thin federal ceiling with no audit backbone is a loaded gun pointed at every product roadmap in the industry. You are trading a known compliance headache today for an unknown overcorrection you don't control.

Where the council splits. The Enterprise Buyer and the Safety Lens want opposite things from the same clause: preemption is a gift to the compliance budget and a disaster for anyone who needs an independent audit trail. The second fault line is motive. The Skeptic says the bill is a coherent expression of Anthropic's actual risk model, that current systems aren't frontier-dangerous so weak thresholds are internally consistent. The Safety Lens and Researcher say the thresholds are incoherent on their own terms regardless of intent. Both can hold: the bill can be sincere and still be bad policy.

What it hinges on. Three things. Does preemption actually make it in, because that's what turns a weak bill from harmless into net-negative. Does any audit-licensing requirement survive. And does the lame-duck window stay open long enough to pass anything at all. The council leans hard that this specific draft, as described, would leave US oversight weaker than the status quo. The disagreement is entirely about why, not whether.

For buyers, the thing to pin down before leaning on any vendor's safety claims: ask whether their third-party audit comes from an evaluator the vendor selected and paid, or an independent one. Under this bill, it will be the former, and that changes what the audit is worth.

Prediction: The version of a federal AI bill that moves through Congress with Anthropic's backing by the end of the 119th Congress's first session (January 3, 2027) will retain broad preemption of state AI laws while omitting any government-licensed independent auditor requirement.

Confidence: Medium. Preemption is the clause labs want most, and auditor licensing is the one they'll trade away first.

Why: The leaked draft already pairs broad preemption with no auditor licensing, and Zvi Mowshowitz, Nathan Calvin, and others identify preemption as the feature Anthropic is actively selling to lawmakers, because killing the SB 53 patchwork is the concrete win every frontier lab wants. Independent auditor licensing is the clause with no industry constituency pushing for it and an obvious cost to the labs being audited, so it is the first thing cut in negotiation. The opposite outcome, a bill that keeps licensing and drops preemption, would require the labs to abandon their main ask while accepting the oversight they drafted the bill to avoid, which runs against the incentive driving the whole effort.

Revisit by 2027-01-15: We're right if any federal AI bill advanced with Anthropic's public support through the end of the first session contains state-law preemption and lacks a government-licensed auditor mandate. We're wrong if such a bill either drops preemption or adds government licensing of independent auditors, or if no such bill advances at all.

The lame-duck clock is the live risk to this call. If nothing moves before the session ends, the prediction fails on "no bill advanced," and that's a fair loss. But the structure of the draft, cheap wins kept, costly checks dropped, is the shape these fights almost always take.