Whose Threshold? A Calibration Audit of AI-Text Detectors Applied to ICLR Peer Reviews

Materials for a study that estimates the prevalence of AI-generated and AI-rewritten text in ICLR peer reviews after correcting for detector error, and that audits what happens when four AI-text detectors are used at their shipped default thresholds instead of at calibrated operating points. No review text is redistributed. The deposit contains review identifiers, per-review detector scores, derived tables, figures and the full pipeline. Anyone with OpenReview API access can reconstruct the exact corpora from the identifiers using the retrieval scripts included here.