AI Text Detectors Amplify Typicality, Not Human-AI Boundaries

A new study on arXiv (2605.21653) reveals that specialized AI text detectors struggle to tell apart human writing from AI-generated text, instead focusing on a standard similarity measure. By using raw encoders without specific task guidance, researchers found that projecting the differences between AI and human text achieved performance levels that matched or surpassed fine-tuned models, with AUROC scores of 0.806, 0.944, and 0.834 across three different architectures, reaching 86-106% of the best fine-tuned results. Interestingly, complete fine-tuning of RoBERTa-base lowered discrimination effectiveness for fluent-formal texts. For non-native ESL writing, the performance dropped significantly, yielding AUROC scores between 0.06 and 0.20. A fixed probe with 24 examples performed comparably to fine-tuning (0.900 vs. 0.895). A closed-form Jacobian predictor was able to accurately parameterize adjustments, significantly improving ELECTRA-CE performance.