AI labs like Anthropic are caught in a biorisk governance puzzle, according to LessWrong. They're deploying protective measures for bioweapon risks even when their models haven't clearly crossed official capability thresholds. The core issue: labs cannot ethically test whether an AI could actually develop novel bioweapons from scratch, so they measure proxies instead—benchmarks, expert trials, incremental progress on isolated steps. With only indirect evidence, it's easy to argue 'one critical piece is still missing,' creating structural bias toward 'not dangerous yet.' The proposal: intermediate warning levels that trigger escalated protections based on measurable progress on specific substeps—not the full end-to-end weapons pipeline, just the bottlenecks. That shifts decision-making from a binary threshold into a graded response.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton