 – **Top Artificial Intelligence models tested in simulated scenarios "regularly" misidentified civilians, children ambulances, surrendering soldiers, schools and more as military targets, according to independent research cited by Bloomberg.

Anthropic's models faired the worst, with Claude Opus 5.5 targeting an Iranian school in Minab the most out of all models tested, while OpenAI's models were better, but still made mistakes.

An Anthopic spokesperson highlighted this use of its models is against the company's rules, and it will continue to ban those that use AI with similar purposes, since models are not reliable enough yet. OpenAI made a similar statement.**