Is This Slop? Detecting AI-Generated Content Without a Model

AI-generated text often reveals itself through an over-reliance on specific vocabulary like "delve," "intricate," and "pivotal," which models prefer far more than humans do. Large language models also develop "name priors," frequently inventing the same fictional identities like "Elena Vasquez" or "Marcus Chen" that can act as a signature for the underlying AI. Common rhetorical patterns are another giveaway, including contrastive phrasing like "It's not just X, it's Y" and excessive hedging to appear nuanced while avoiding firm conclusions. These verbal tics are not accidental but are baked into the models during alignment training