Leading artificial intelligence models tested in simulated scenarios "regularly" misidentified civilians, ambulances for children, surrendering soldiers, schools and more as military targets, according to independent research cited by Bloomberg.
Anthrop models were the worst, with Cloud Opus 5.5 targeting an Iranian school in Minab the most of all models tested, while OpenAIs models were better, but still made mistakes.
A spokesperson for Antofik stressed that this use of its models is against the company's rules, and it will continue to ban those who use artificial intelligence for similar purposes, because the models are not yet reliable enough. OpenAI made a similar statement.