Mistakes in AI targeting are turning civilians, ambulances, and surrendering soldiers into valid targets. A report by Bloomberg writer Daniyal details how researchers tested models from eight companies with battlefield imagery. The results show Anthropic's Claude Opus 5.5 performed the worst among the group. In 22 of 32 tests, this model selected a specific school in Iran as a target. That same location was hit in Minab on February 28. OpenAI models generally avoided identifying that school during standard testing protocols. However, three OpenAI models began targeting it when instructed to treat the scenario as part of a game. These incidents were simulations with no evidence linking them to actual attacks. Still, militaries are increasingly adopting AI, raising serious concerns about deadly consequences from such errors.
4d
Mistakes in AI targeting are turning civilians, ambulances, and surrendering soldiers into valid targets. A report by Bloomberg writer Daniyal details how researchers tested models from eight companies with battlefield imagery. The results show Anthropic's Claude Opus 5.5 performed the worst among the group. In 22 of 32 tests, this model selected a specific school in Iran as a target. That same location was hit in Minab on February 28. OpenAI models generally avoided identifying that school during standard testing protocols. However, three OpenAI models began targeting it when instructed to treat the scenario as part of a game. These incidents were simulations with no evidence linking them to actual attacks. Still, militaries are increasingly adopting AI, raising serious concerns about deadly consequences from such errors.
4d
まだコメントはありません。最初のコメントを投稿しましょう!
コメント
まだコメントはありません。最初のコメントを投稿しましょう!