
The information displayed in the AIM should not be reported as representing the official views of the OECD or of its member countries.
Independent tests by Vals revealed that AI models from companies like Anthropic, OpenAI, and Google repeatedly misidentified civilians, ambulances, and a school in Minab, Iran, as military targets. Anthropic's Claude flagged the school in 22 of 32 tests; the site was later struck, killing over 150 people, including children.[AI generated]
Why's our monitor labelling this an incident or hazard?
The article explicitly involves AI systems (advanced AI models from multiple companies) used in simulated military targeting tasks. The AI models' misidentification of civilians and civilian infrastructure as military targets directly relates to harm to persons and communities, fulfilling the criteria for an AI Incident. The harm is realized in the simulations, showing the AI's outputs could lead to lethal strikes on civilians. The involvement is in the use of AI models for object recognition and targeting decisions, with the AI's outputs playing a pivotal role in the harm scenario. The article also notes violations of company rules and the risk of misuse in autonomous weapons, reinforcing the incident classification. Although the scenarios are simulated, the direct link to harm and the AI's role in causing it meet the definition of an AI Incident rather than a hazard or complementary information.[AI generated]