Source-reported · Independent verification status is not recorded for this article; claims remain attributed to their sources.
Study Finds Leading AI Models Mislabel Civilians and Schools as Targets in War‑Game Tests
A Bloomberg‑reported study by San Francisco firm Vals, which works with U.S. government agencies, found that major AI models in simulated conflicts repeatedly mistook civilians, ambulances, surrendering soldiers and even a school in Iranian Minab for military targets.
Fresh testing of artificial intelligence in combat simulations has exposed serious flaws in how current systems distinguish civilians and protected sites from legitimate targets.
Bloomberg reported on a study by Vals, a San Francisco‑based company whose clients include American government agencies. In simulated military scenarios, all leading AI models assessed in the research regularly misidentified civilians and civilian objects as valid targets.
According to the summary, the models in these war‑game‑style tests sometimes classified ambulances, surrendering soldiers and even a school in the Iranian city of Minab as military objectives. That behavior appeared across multiple top‑tier models from major developers, including Anthropic.
The reporting does not specify which exact models or versions were tested, nor does it say that any of the systems are currently deployed in real‑world weapons. But the scenarios were designed to mimic battlefield conditions where rapid target identification matters.
The findings arrive at a time when defense agencies are exploring ways to use AI to support targeting, surveillance and decision‑making. Mislabeling ambulances and schools as combat targets, or failing to recognize attempts to surrender, would directly undermine legal and ethical requirements to distinguish between combatants and civilians.
Vals’ work suggests that, even in controlled simulations and with state‑of‑the‑art models, AI can still make systematic errors that would be catastrophic if translated into live operations. The fact that the tests are being conducted and reported, however, indicates that at least some parts of the defense and research community are probing these weaknesses before any wider deployment.
What happens next will depend on how militaries and policymakers react to such studies. Tighter limits on AI roles in targeting chains, more demanding testing standards, and clearer rules on autonomous systems are among the steps that could be shaped by this kind of evidence.