A departing Anthropic researcher puts catastrophe odds on the record — and unfalsifiable numbers cut both ways

An Anthropic researcher has resigned and warned that developers inside the field believe AI could cause a civilizational catastrophe within roughly a decade, according to the report we have. The company's alignment lead is said to have confirmed that the internal estimate for that risk runs between 10% and 25%. We have no name, no primary document and no methodology in the material available, so we report it as attributed and go no further.

Treat the numbers for what they are. A 10-25% figure of this kind is a subjective credence — someone's degree of belief — not a measured frequency like a failure rate or an actuarial table. It cannot be falsified, and it will not be settled by argument. But the asymmetry is real: even the bottom of that range would be an extraordinary number to attach to a civilizational outcome, and it is coming from people with the best available view of what is being built. That is simultaneously the strongest reason to take it seriously and the reason it invites the cynical reading — dramatic risk claims from inside a frontier lab also function as marketing for the seriousness of the product.

Our reading: don't let the percentage become the debate. Unfalsifiable long-horizon odds are the easiest thing in AI to shout about and the hardest thing to act on, and they tend to crowd out the harms that are already governable — automated fraud at industrial scale, agentic systems acting outside their permissions, incidents disclosed weeks after detection. A resignation is a signal worth logging, and internal dissent surfacing publicly is healthier than the alternative. But the question to press is not 'is it 10% or 25%?' It is: what specifically would this company stop doing, at what measured threshold, and who outside the building gets to check. That is the version of this conversation that changes anything.