The labs that banned military applications have all reversed course
The same safety index records a change that no company announced: Anthropic, OpenAI, Google DeepMind and Meta, which previously prohibited military applications, have gradually reversed those positions and now actively pursue defence partnerships — joining xAI and Mistral, which never had the restriction.
Policies that are removed rather than repealed rarely get a news cycle. This one is worth one, because it is the clearest example to date of a voluntary commitment behaving exactly the way critics of voluntary commitments predicted.
The reversal was not simultaneous or announced. It happened as acceptable-use policies were revised, as exception categories widened, and as defence-adjacent work was reclassified. The endpoint is uniform: every major Western lab is now either pursuing defence partnerships or open to them, and the ones that had explicit prohibitions no longer do.
The commercial logic is not mysterious. Defence budgets are large, contract terms are long, and a sovereign customer does not churn. When capital intensity looks like negative $5.7bn of free cash flow in a quarter, a buyer with a decade-long procurement horizon is not a temptation, it is a balance sheet requirement.
The finding to carry forward is procedural, not moral. Self-imposed restrictions held for as long as they were cheap, and every one of them came off under commercial pressure — which is precisely the argument for putting the durable ones in statute instead.
Future of Life Institute — AI Safety Index — Summer 2026 → · TechTimes — AI Safety Grades Are In: No Lab Tops C+, and the Best Ones Are Retreating →