Do AI models amplify conflict and deny atrocities?

When AI models are deployed in conflict zones, they shape how journalists, aid workers, and citizens understand atrocities. Researchers tested nine configurations from OpenAI, Anthropic, DeepSeek, and xAI on 90 scenarios designed to catch misalignment: denying genocide, equating documented atrocities, missing ethnic slurs. Failure rates ranged from 6% to 47% between best and worst performers, and when prompted to show "balance" in cases where courts have assigned responsibility, five of nine models failed 80–100% of the time. The team released the first evaluation framework for conflict-specific harms and argues this should become standard alignment testing.