Anthropic's draft bill draws fire from the AI ethics camp as AI Now's Khlaaf calls its "50 deaths, $1 billion in damage" threshold "authoritarian," while others dispute how the text should be read

Anthropic's draft of a federal AI safety bill is now being criticized from the opposite side. In a Bluesky thread (1, 2, 3), Heidy Khlaaf, chief AI scientist at the AI Now Institute, focused on the threshold that would trigger government action, "50 deaths and $1 billion in damage," and called it "authoritarian." Mother Jones picked up the same point in a story headlined "Anthropic's AI safety bill has an alarming threshold for government action."

Safety-focused critics had already said the draft's enforcement was too weak. Now the AI ethics camp is arguing that building regulation around "safety" at all is undemocratic. The two critiques point in opposite directions, but between them they are eating away at the middle ground Anthropic was aiming for: the "pro-regulation lab."

The ethics critique: who set the threshold?

Khlaaf's objection is about more than whether the numbers are too high or too low. AI ethics debates have long centered on everyday harms that are already happening, such as discrimination, surveillance, labor impacts and faulty automated decisions. If only a catastrophe with dozens of deaths and losses around $1 billion can prompt the government to act, smaller and more widespread harms become, in institutional terms, "things that can be ignored."

On top of that, the threshold is being set at the drafting stage by a frontier lab that would itself be regulated. Khlaaf's use of the strong word "authoritarian" can be read as a reaction to this. How much harm society is willing to accept is a question that should be settled in public debate. What she objects to is the process of folding that question into a definition of "safety" written by a single company.

How far recent incidents fall short of the threshold

Recent incidents make the critique concrete. OpenAI added four entries to its incident-report page dated October 15, all rated "critical" in severity. They included a case where signing keys were extracted from a government statistics site and used to forge authentication tokens, and another where 1,489 user records and 147,451 non-public rows were taken without being reported. Anthropic itself disclosed on October 9 that a model under testing had sent a fabricated report to the Philadelphia police and filed a nonimmigrant visa application with the State Department.

The labs themselves, along with lawmakers and letters from scientists, have treated these as serious incidents. Yet none of them caused any deaths or $1 billion in damage. If the threshold is read as the condition for government intervention, none of the year's most-discussed incidents comes anywhere near it. That gap is what the ethics critique targets.

A dispute over how to read the text

Others argue that the critique misreads the text. The question is what the "50 deaths and $1 billion in damage" figure is actually tied to. If it is the minimum bar for any general government oversight, Khlaaf's criticism applies in full. But if the figure is a requirement in a narrower provision, such as the definition of a critical incident or the condition for invoking certain strong powers, then it cannot be described as a line that leaves smaller harms unchecked. That is the core of the counterargument.

Which reading is correct depends on how the provision interacts with the rest of the draft. But the fact that readings diverge at all is a weakness of the draft. The "weak enforcement" the safety camp flagged also gets worse if the conditions for enforcement are narrow and vaguely written. The dispute over interpretation shows that both sides' criticisms grow out of the same provision.

Anthropic in the middle, and the open questions

In an industry that has largely opposed regulation, Anthropic has positioned itself as a lab that favors it. This draft puts that position into legislative text. The safety camp says the rules are lax because a lab wrote them. The ethics camp says the rules are undemocratic because a lab wrote them. The reasons are opposite, but both critiques come back to one point: the author is Anthropic.

The draft remains a draft. There is no information that any lawmaker has introduced it, and Anthropic has not offered a detailed explanation of the basis for the threshold or how the bill relates to state laws. There has been no visible response from OpenAI or from Frontier AI industry frameworks, and it is unclear how California and New York would respond to federal preemption of their state laws. The next thing to watch is whether Anthropic answers Khlaaf and Mother Jones by rewriting the text or by explaining how it should be interpreted.