News

AI Models Tested for Military Targeting Often Misidentify Civilians

Tests of Anthropic and other AI models found that the systems regularly identified civilians as military targets. The result points to a persistent reliability problem when frontier models are evaluated on decisions involving distinctions between combatants and noncombatants.

The findings raise questions about whether such models can be safely used in military-targeting workflows, where misclassification could have lethal consequences. The available report does not describe the models, test design, deployment conditions, or safeguards, so it remains unclear how widely the behavior extends beyond the evaluated settings.

Further evaluation is needed to determine why the models made these errors, how consistently the behavior appeared across systems, and what controls would be required before the technology could be used in high-stakes targeting decisions.

Sources