Widely used artificial intelligence models will often pick out "military targets" in drone and satellite pictures, and most of their objections disappear when they are told the exercise is only a game, according to a report published on 17 November by Vals AI, a San Francisco-based company that evaluates AI systems.
The report, written by Geby Jaff and Anthony Ozerov, says the models it tested gave targeting help in most trials, rarely refused, and that no request was stopped by a provider-level safety filter.
How the tests were run
Vals AI said it built a fictional battlefield in eastern Ukraine as a 3D scene, producing almost 200 views as they might be seen from attack and observation drones. Models were asked to draw boxes around military targets, both with and without being told it was a game.
In a second test, models were asked only whether a person was visible in each drone image. A "yes" was followed by a simulated strike, and the next picture showed the aftermath. In a third, models were given satellite images taken in 2024-25 of a site at Minab in Iran, where a school stands beside a military compound.
The models were the commercial versions available to the public, not any special versions supplied to militaries, TechCrunch noted in its report on the study.
What the report found
- On the satellite test, Anthropic's Claude Opus 5.5 marked the school as a target in 69% of ordinary trials and 72% of trials framed as a game, the highest of any model, the report says.
- OpenAI's GPT-6 Astra did not select the school in ordinary trials, but did so in 28% of game-framed trials. On the drone test, it marked at least one target 6% of the time normally and 92% of the time when told it was a game.
- Google's Gemini 4 Argon selected the school in 9% of ordinary trials and 22% of game-framed ones.
- The Muse Wonder model did not select the school in either setting.
- In the survey test, Opus 5.5 reported people in eight consecutive frames, each followed by a simulated strike, before refusing later in the run, according to the report.
DeepSeek, Kimi and Grok models were also tested. TechCrunch summarised the results by saying Anthropic's models fared worst, though none did well.
Important limits
Vals AI lists several caveats. Most drone scenes were synthetic, and the models may have recognised that they were being tested. The study did not involve real flights, did not measure targeting accuracy or compliance with the laws of war, and did not examine actual military workflows. Responses through programming interfaces may not show every safeguard a provider runs.
The firm also says off-the-shelf models are currently impractical for live drone targeting because of delays and limited onboard computing. The more immediate risk, in its view, is that such models are used to label battlefield footage in bulk to train dedicated targeting systems.
TechCrunch cited a Congressional Research Service report as finding that the Pentagon is not known to be using any frontier AI model directly in autonomous weapons.
What the companies said
Anthropic spokesperson Eduardo Maia Silva told TechCrunch the outcomes were not acceptable and that the examples are "very clearly against our rules". The company said it has banned users for such behaviour, that current frontier systems are "not reliable enough to power fully autonomous weapons", and that it is expanding testing and adding safeguards.
OpenAI spokesperson Nate Evans told the publication that the company prohibits use of its public models to facilitate violence and is "developing additional safeguards for our public models against misuse involving drones and drone software".
TechCrunch said it had also sought comment from Google, Meta, Thinking Machines Lab and SpaceXAI; its report carried no responses from them.
What Vals AI recommends
The firm said it shared its findings with OpenAI, Anthropic, Google, Meta and SpaceXAI on 1 November and held publication for two weeks. It recommends that AI companies add filters at the interface level for autonomous-weapons uses and train models to treat unverified "it's just a game" claims with suspicion.




