Measuring Artifact Reliance in Machine-Generated Text Detectors

Key points

• A logistic regression using 27 superficial features reached 0.833-0.997 macro-F1, and on the three arXiv subsets it outperformed every detector we evaluate with F1-scores 0.951-0.997.1
• The strongest cue to the detector is the PDF extraction artifact: human texts retain their original line breaks, whereas generated texts are extracted as a single block of prose.1
• To distinguish benchmarks that contain such a signal from detectors that rely on one, we introduce format counterfactuals: just re-wrapping a machine-written abstract without changing a single word reverses 20-27% of a public detector's correct decisions.1
• Normalizing formatting before training reduces the generalization gap from 0.32 to 0.22.1

Sources (1)

• [1]Measuring Artifact Reliance in Machine-Generated Text Detectors
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 01:34 PM

  Detectors of machine-generated text score almost perfectly on scientific-literature benchmarks, implying we can reliably distinguish with these detectors human and machine-generated documents.

  A logistic regression using 27 superficial features reached 0.833-0.997 macro-F1, and on the three arXiv subsets it outperformed every detector we evaluate with F1-scores 0.951-0.997.

Extractive summary: sentences quoted from the sources.