Why Anthropic’s draft bill sets $1 billion threshold for AI risk - Politico

Anthropic — the lab that puts 'safety' in the mission statement and a constitution in the product — has drafted a bill, and the bill has a number in it. The number is 50. Fifty people killed or seriously injured, or one billion dollars of damage, and only then may the government consider blocking a model's release, Politico reported on 21 October. Fifty is a very specific figure. It is also, in the language of actuarial paperwork, a rounding error with a press release attached.

Note the grammar of the thing: 'only allow.' Not 'shall require review,' not 'must be demonstrated safe.' The default is release; the exception is catastrophe; and the catastrophe is quantified in advance so that everyone can agree on precisely how much ruin is tolerable before a regulator clears their throat. Even better, the threshold 'mirrors a California law' but repurposes it — from whatever it did before into the single gate deciding whether a model ships. Someone has discovered that a liability cap makes an excellent release valve, provided you move it to the other side of the machine.

The lovely part is who is holding the pen. Not a regulator, not a legislator — a lab, pitching its own floor to lawmakers, which is roughly the governance standard you get when the restaurant drafts the health code and everyone applauds the transparency. We noted recently that the administration would rather lean on existing statutes than draft anything new; Anthropic appears to have solved that by drafting something itself, briskly, before somebody less friendly does it for them. One billion dollars is an enormous sum right up until it is the price of finding out.