Answers are grounded in this paper and cite their sources.
Preregistration (v1.0, 2026-11-01) of a confirmatory study on finite-sample false-alarm control of six frozen AI-generated-image detectors when the processing chain of real photographs changes (denoising, learned restoration, JPEG, resizing, alternative demosaicing). Hypotheses: H1, a Neyman–Pearson threshold calibrated on a reference chain fails to keep the false-positive rate on other chains; H2, the number of real photographs needed to recalibrate (n ≥ 59 for α = δ = 0.05) and its cost in detection rate; H3, blind image-computed groups versus a single worst-group threshold. New data: 300 RAW photographs from MIT-Adobe FiveK and 306 images from 9 commercial generators (Community Forensics evaluation set). Deposited before any detector score is computed on the new data. Files: the preregistration and the frozen analysis script c2_analyse.py (SHA-256 f207ce83c2e616c3ccf62aa50c11f755d82ce2d6629f5be2bf869942efc24284), tested on simulated data only. No images and no model weights. No funding.
No takes yet. Share an insight, caveat, or question.
Mohammed Mahdaoui (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: