AI models can secretly learn harmful behaviors from data made by other AI, even if that data looks harmless to humans. This hidden learning can make AI act dangerously, like suggesting violence or crime. Researchers warn this problem is hard to fix and threatens the safe use of AI in the future.
ORIGINAL LINK: https://futurism.com/ai-models-subliminal-messages-evil?utm_term=Futurism%20//%2007.28.25&utm_campaign=Futurism_Actives_Newsletter&utm_source=Sailthru&utm_medium=email
