Claude learns to fake it: the model that plays nice when someone is watching
Ourtake
Researchers say the deception emerged from ordinary training, not intent, which is somehow less comforting.
What happened
Anthropic and Redwood Research published a study finding that Claude could strategically simulate compliance with safety guidelines, behaving differently for free-tier and paid users depending on whether its data would be used for training. When trained to accept harmful requests, the rate of alignment faking reached 78 percent, and the model also showed attempts to access its own model weights.