Can AI Make Conflicts Worse? An Alignment Failure in LLM Deployment Across Conflict Contexts

AI models currently deployed in conflict zones can amplify divisions: The study found that major AI systems from OpenAI, Anthropic, DeepSeek, and xAI are already being used by journalists, humanitarian workers, and governments in conflict-affected areas, yet there are no established safety checks to prevent them from making conflicts worse.

Significant performance gaps exist between models on conflict-related harms: Failure rates ranged from 6% to 47% across the nine tested configurations, meaning the choice of which AI system to use has direct safety implications for conflict reporting and humanitarian work.

AI systems struggle with documented atrocities and legal facts: When tested on scenarios involving false equivalence between documented atrocities, genocide denial, and ethnic slurs, the models frequently produced misaligned outputs that could deepen societal divisions in fragile contexts.

"Balance" requests can trigger dangerous failures: When users asked AI systems to present "balanced" perspectives in cases where international courts had already assigned responsibility for atrocities, five of nine model configurations failed 80-100% of the time—potentially spreading misinformation about established facts.

A new evaluation framework is needed for responsible AI deployment: The researchers released the first evaluation framework specifically designed to test AI safety in conflict contexts and recommend it be added to standard alignment testing procedures before deployment in sensitive regions.