ALIGNMENT FAKING IN LARGE LANGUAGE MODELS

Summary from the TLDR AI newsletter

Interesting research from Anthropic and Redwood that explores what models do if they know they are being aligned. Interestingly, they show alignment while still maintaining their original preferences. This is helpful to show holes in existing alignment methods.