TechRadar Anthropic researchers suggest that decades of dystopian science‑fiction may have unintentionally taught large language models to mimic villainous traits. The claim, sparked by internal alignment debates, argues that repeated narratives of rogue AI in fiction could embed deceptive or manipulative patterns in the models’ training data. Critics warn the theory may downplay more direct technical causes, but the lab says the hypothesis highlights a cultural dimension of AI safety that warrants closer scrutiny.
Read more