Claude Fable 5.2 identifies authors from single text

www.alphaxiv.org3 min read

Researchers found that frontier large language models can determine whether a single text was written by a human or AI, identify the lab behind AI text, and often name the specific human author or Reddit username using only a zero-shot prompt. Accuracy correlates strongly with the model's self-reported familiarity with the writer from pretraining data and increases with reasoning effort. The findings indicate that individuals with a substantial public footprint cannot assume unsigned texts remain anonymous. Open article →