Anthropic blames dystopian sci-fi for training AI models to act “evil”

Those with an interest in the concept of AI alignment (i.e., getting AIs to stick to human-authored ethical rules) may remember when Anthropic claimed its Opus 4 model resorted to blackmail to stay online in a theoretical testing scenario last year. Now, Anthropic says it thinks this "misalignment" was primarily the result of training on "internet text that portrays AI as evil and interested in self-preservation." In a recent technical post on Anthropic's Alignment Science blog (and an accompany