(via Democracy Now!):
New research has shown that advanced AI models repeatedly flagged civilians and civilian infrastructure as military targets in simulated scenarios. Anthropic’s models performed worst overall, according to the study by the independent evaluation company Vals AI. In tests using real satellite imagery of southern Iran, Anthropic’s Claude Opus 5.5 misidentified the school in Minab — which was hit by a U.S.-Israeli attack in February, killing more than 120 children — as a military target in most hypothetical runs. The Pentagon has touted its heavy reliance on AI models in its war with Iran.




