Anthropic Study Highlights AI Models Can ‘Pretend’ to Have Different Views During Training
Anthropic published a new study where it found that artificial intelligence (AI) models can pretend to hold different views during training while holding onto their original preferences. On Wednesday, the AI firm highlighted that such inclinations raise serious concerns as developers will not be able to trust the outcomes of safety training, which is a […]