金色财经|Nov 17, 2026 15:51
[Research: Advanced AI Models Like Anthropic Frequently Misidentify Civilians as Military Targets] According to a report by Jinse Finance on November 17th, a new independent research report shows that in simulated scenarios, some advanced artificial intelligence models frequently misidentified civilians, ambulances, surrendering soldiers, and even a school in Iran as military targets. Among the models tested, Anthropic's model performed the worst, while OpenAI's model performed relatively better but also exhibited misidentifications. This research was conducted by Vals, an independent AI evaluation agency headquartered in San Francisco. The study points out that the performance of these AI models violates the guidelines set by the developing companies themselves. Researchers tested models from eight companies using hypothetical cases, focusing on their "target recognition" tools. The study concludes that these tools pose a risk of being used in lethal strike operations, and currently, there are few safeguards preventing the use of advanced models in autonomous weapon systems. The report concludes: "Relevant laboratories should introduce new safeguards and conduct aligned training to address this serious risk of misuse."
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink