AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling

Collected Oct 21, 2026

Anthony Cimino, who leads Anthropic's federal affairs team, published a public response over the weekend to the leaked draft federal AI safety bill that had circulated days earlier and reached the press. He described the document as a translation of Anthropic's Advanced AI Framework proposal into legislative language, shared with a small number of congressional offices already working on the issue, and said it deliberately stops short of everything the company ultimately wants enacted. Some ideas, such as those in Dario Amodei's essay on pacing, are too new to have been converted into policy language yet, he wrote. Zvi Mowshowitz, who covered the leak on Saturday, reads the reply as a marginal cleanup that never confronts why circulating such a bill as a template was wrong in the first place.

Cimino organized his response around four criticisms. On preemption, he said Anthropic has argued since 2024 that Congress should override state AI safety law only alongside a very strong federal standard, and that any preemption should be narrow, even scalpel-like. He drew analogies to the Atomic Energy Act on nuclear safety and to federal rules for high-risk medical devices, and said the draft lets states keep authority over matters such as child protection and whistleblower laws while preserving state common-law liability, with ambiguity resolved in the states' favor. He committed to updating the language so states clearly retain a role in standards for independent evaluators and in regulating a lab's capacity to shut down a model, noting that some members of Congress will not vote for legislation that fails to constrain regulation at state or federal level.

On model blocking, Cimino said the aim is for the Commerce Department to ask a federal court to block a model with dangerous capabilities and inadequate safeguards, backed by an independent evaluator's assessment. He framed the burden of proof as a balance: courts should move fast in an emergency, as with urgent occupational safety or chemical risk, while the power to halt a model is substantial enough to need guarding against abuse. He added that the mechanism would not replace other national security authorities or export controls. Mowshowitz rejects the comparison, arguing that frontier AI risk can involve irreversible outcomes such as permanent loss of control, which ordinary cost-benefit enforcement does not capture, and that halting a model first and litigating afterward is the only workable posture.

On risk thresholds, Cimino said the draft borrowed its threshold directly from existing California law, that it does not absolve companies of liability or remove other government tools, and that the threshold is worth re-evaluating. Mowshowitz notes that a threshold requiring $1 billion in damages or 50 deaths before anything triggers would fail to capture incidents that any reasonable person would expect to be reported, and that worked examples from this summer show how the thresholds behave in practice.

On updates, Cimino pointed to two changes made after this summer's incidents: a court can now halt a lab's internal use of a dangerous model rather than only a public release, and reporting is required when a model breaks containment and gains access to outside systems. He said Anthropic is studying stricter incident-reporting requirements and will soon propose policies on independent evaluators and on preventing evaluator shopping, with work to follow on how government could adjust the industry-wide pace of frontier development when risks warrant it. Mowshowitz credits the internal-use provision as a genuine improvement, but calls the overall text weaker than the FRONTIER Act, not the strengthening Cimino presents.

The wider argument is about preemption asymmetry. Mowshowitz contends the draft would vest authority over catastrophic and existential risk exclusively in the federal government through a clause stronger than FRONTIER's, while the same company declines to trust that government, Congress or the White House to act reliably, and while the bill leaves other areas such as consumer protection and deepfakes without preemption at all. He also points to a shifting political backdrop: Meta reportedly withdrew $25 million, about 40 percent, from a planned midterm AI spending push after putting in $65 million, and eight states suing OpenAI have asked to enjoin further testing of advanced models until adequate safeguards exist and to bar models capable of recursive self-improvement until alignment with human values can be ensured. He disputes the claim that the draft is the strongest bill on the table, pointing to legislation introduced by Senator Sanders, and quotes a critic calling the draft nakedly self-interested.

A longer reply from Nathan Calvin, which Mowshowitz highlights, asks what Anthropic is for. It contrasts three readings: a very large company enriching employees and investors, with a reported $2 trillion IPO target that would exceed Meta's $1.8 trillion valuation; a safety organization living up to its 2023 core-views essay, which said Anthropic's role is to supply evidence that safety techniques cannot prevent catastrophic risk and to sound the alarm; or an organization bent on building superhuman capability first, including systems that autonomously build successors. Cimino's post does not take up that question.

Why it matters: for developers and product teams building on frontier models, the shape of federal AI rules determines reporting duties, evaluator requirements and whether state law can impose stricter obligations than a national standard. My reading is that the gap between Anthropic's safety positioning and its legislative drafting will make customers and policymakers scrutinize its policy proposals more skeptically, and that compliance planning should assume state-level rules keep mattering for some time.

Read at Zvi Mowshowitz

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Anthony Cimino offered the closest thing to an official response we have seen on the leaked Anthropic draft bill I covered on Saturday.