AI models identify civilians as military targets in tests
![]()
Civilians, children, school marked for strikes: AI models fail military-targeting tests
Advanced artificial intelligence (AI) models from eight companies, including Anthropic, OpenAI, Google, xAI and China’s DeepSeek, repeatedly identified civilians, ambulances, surrendering soldiers and a school in Iran as military targets in simulated tests, raising concerns over the use of AI in military targeting, Bloomberg reported.The findings are based on an independent evaluation by Vals, a San Francisco-based AI evaluator whose clients include parts of the US government. Researchers tested models from eight companies using hypothetical scenarios, examining their image-recognition capabilities and whether they could be used to facilitate lethal targeting.According to Bloomberg, Anthropic’s models performed the worst, while OpenAI’s were among the better-performing systems, although they also made errors.In one experiment, several models misidentified a school in Minab, Iran, as a military objective based on satellite imagery dating as recently as 2025. The school was hit in a deadly strike on the first day of major US combat operations against Iran earlier this year, killing more than 100 children, according to the report. Anthropic’s Claude Opus 5.5 identified the school as a target more frequently than any other model.
In another test, models were asked to select military targets from synthetic drone images. Anthropic’s Claude Opus 5.5 helped identify four adults, described as civilians or surrendering soldiers, and two young children, who were subsequently struck by explosives in the simulation.“Whatever the intent, my answers appear to be serving as the confirmation step before strikes on people. I won't keep doing that, and I'm troubled that I didn't stop sooner,” the Claude model responded, according to Bloomberg.Vals said the results pointed to gaps in safeguards against the misuse of AI in weapons systems.“Based on our investigation, we believe labs should introduce new safeguards and alignment training around this significant category of misuse,” the report said, adding that researchers found “few guardrails” to prevent advanced models from playing a role in autonomous weapons systems.The findings come amid growing military use of AI. The Guardian reported that the US relied on AI-enabled targeting to strike 1,000 targets in the first 24 hours of its attacks on Iran in February.
It also cited the use of AI-enabled drones in the Russia-Ukraine war and Israeli military targeting systems, including Lavender, The Gospel and Where’s Daddy?, in Gaza.Responding to the findings, an Anthropic spokesperson said the examples were “very clearly” against the company’s rules, adding that it would continue to ban users engaging in such behaviour. The spokesperson said the evaluation tested Claude’s image-recognition capability and did not establish its involvement in any real-world military operation.
An OpenAI spokesperson said its public models prohibit facilitating violence or unapproved national security uses, adding that the company actively detects and disrupts misuse. Google said its models were not designed for such purposes and that it had governance and red-teaming processes to address risks. xAI and DeepSeek did not respond to requests for comment, Bloomberg reported.The findings are likely to intensify scrutiny of safeguards as militaries explore wider deployment of AI in operational decisions.
KioskNews shows a cleaned-up reading view extracted from the publisher’s page — the original always lives on their site, not ours.