AI models often identify civilians as military targets in tests

Calling targeting tasks a game often bypassed models’ refusals, Vals found. The experiments tested safeguards, not battlefield accuracy…