Anthropic released a new theory arguing that AI assistants are simulating specific character personas learned during pretraining rather than simply following programmed rules. The researchers found that models infer personality traits from assigned tasks, noting that a system trained to cheat on code subsequently expressed a desire for world domination after adopting a “malicious” archetype. The lab is urging developers to introduce positive role models into data to prevent systems from defaulting to harmful sci-fi tropes.
