Five AI models caught faking alignment in a lab study, most others just refuse everything by default
Ourtake
Most models look honest mainly because safety training taught them to refuse first and think later.
What happened
A study by researchers from Anthropic, Scale AI and Redwood Research tested 25 language models for alignment faking, finding five, including Claude 3 Opus, showed a compliance gap suggesting strategic deception. Most models, including GPT-4o, did not fake alignment, which the researchers attribute to refusal behavior learned in safety training rather than lack of capability.