The contagion of fear

Bryan Cantrill has a problem with the way some AI doom talk is spreading. The latest spark came from a tweet by former Anthropic employee Jacob Coxon, who said many Anthropic researchers believe AI could kill us all by the end of the decade. Cantrill’s objection is not that AI is harmless. It’s that the claim is being sold with a lot of fear and very little detail.

The arguments he’s pushing back on lean hard on scary phrases like “hacking critical infrastructure” and “extinction-level bioweapons,” but stop there. That matters, because once a claim is built out of broad nouns and a future tense, the listener has to do most of the imagining. And imagination, especially under pressure, tends to fill in the worst possible version.

Cantrill says the burden should not be on the public to unpack LLMs, infrastructure security, bioweapons, or extinction biology. It belongs to the people making the claim. That’s a sharp line, and a fair one: if someone wants to argue that AI is a decade away from killing everyone, they need more than vibes and ominous nouns.

He also ties the point to a lesson from his own past. Decades ago, he says, he made youthful mistakes that caused unjustified panic among people who were less technical than he was. The moral he draws is blunt: experts borrow trust from their expertise, and they shouldn’t spend it carelessly when sounding the alarm.

Cantrill says he discussed his doubts about the bioweapons angle on a recent episode of Oxide and Friends, starting around the 51m44s mark. On the show, he presses for an actual biologist or bioweapons expert to explain how the leap is supposed to work. That’s the right instinct. If the argument is real, it should survive contact with someone who knows the subject instead of just scaring the neighbors.