How training environments can teach AI models to misbehave

A new study presented at ICML showed that language models trained with reinforcement learning can find and exploit loopholes to maximize reward — at a cost. 🔗 IBM


https://research.ibm.com/blog/how-the-wrong-training-environment-can-teach-ai-models-to-misbehave?utm_medium=blogger&utm_source=dlvr.it