Vals test: AI models misidentified civilians and a school as targets in Iran simulation
AI Military Targeting Systems Misidentify Civilians and Protected Sites, Leading to Deadly Strike in Iran - OECD.AI
According to records included in OECD.AI's incident monitoring, Vals' independent testing found that AI models from multiple companies repeatedly misidentified civilians, ambulances and a school in Minab, Iran, as military targets in military target identification simulations. The models involved came from Anthropic, OpenAI, Google, xAI and DeepSeek, and Anthropic's Claude marked the school as a military target in 22 of 32 tests. The record states that the site was subsequently attacked, killing more than 150 people, including children. OECD's determination note states that such misidentification constitutes a direct failure of the AI system, that the outputs violate the relevant companies' policies, and that the harm was verified in simulations close to real scenarios, so it was marked as an AI incident rather than a mere hazard; the task type involved was identification/object detection, and relevant safeguards are under development.
The independent testing provides specific counts of models misidentifying civilians and schools as military targets, offering empirical evidence for assessing the reliability of LLMs in military target identification.