Anthropic has reportedly put its lobbying weight behind the bluntest number in AI policy: fifty.
The company is pitching lawmakers a safety bill that would let a model be blocked if fifty or more people could be killed as a result of its deployment, TechCrunch reported this week. Not a benchmark score. Not a compute threshold counted in floating-point operations. Not a capability evaluation whose name has to be explained twice. A death toll, written into law before the deaths happen.
That is a stranger and more deliberate choice than it first looks.
For three years the AI regulation debate has stalled on measurement. Compute thresholds became a lobbying war over where the line falls and who has to count. Benchmark scores go stale in months and reward whoever tunes hardest to the test. Capability evaluations — can this model meaningfully help someone build a bioweapon — are the right question and the hardest thing to certify, because the honest answer is usually a probability with wide error bars, produced by a lab grading its own homework.
A harm threshold moves the burden. Instead of asking what a model can do in the abstract, the bill asks what happens if it does it. Aviation certification and pharmaceutical approval already work this way: not 'is this safe', but 'is the residual risk below a stated level', with the regulator holding the pen rather than the manufacturer.
The word doing the work is 'could'. Fifty people, assessed by whom, on what evidence, at what confidence? A pre-deployment gate requires the answer before anyone has run the model at scale; a post-incident trigger requires fifty funerals first. The gap between those two readings is where the bill will be fought, and where a well-intentioned clause becomes either an unworkable veto or a tombstone.
Then there is what 'blocked' means. A licence withheld is one thing; an order to pull weights that are already mirrored on a thousand servers is another. Regulators can stop a company from shipping. They have a much harder time stopping a file.
The politics cut both ways, and the company knows it. A lab that drafts restrictions on itself buys credibility no press release can match, and it also writes rules that a startup with twelve employees cannot afford to satisfy. Supporters call that the price of a serious law; critics call it a moat with a conscience. Both are describing the same document.
The strongest argument for the number is not technical. It is that fifty can be argued about in public. Compute thresholds are fought in comments to a rulemaking docket; a body count is fought in newspapers. If the goal is a law that survives a change of government, plain language beats precision.
What the bill cannot do is settle whether the threshold is the right one. Nobody knows the shape of the risk curve well enough to say whether fifty is prudent or absurd. But a number that is wrong and debatable can be revised. A principle that is right and unmeasurable tends to sit in a drawer.
