Answers are grounded in this paper and cite their sources.
Protocol (v1.0, 2026-11-01) for a two-level systematic review of AI-generated image detection, registered before screening. Level 1 audits evaluation practice in peer-reviewed papers (2022–2026) from a pre-specified list of leading venues: how decision thresholds are chosen and whether they are independent of the test data, whether false-positive rates on real photographs are reported at a pre-fixed threshold, how the provenance and processing of real photographs are described, and whether the threshold rule applied in public code matches the rule described in the paper. The primary outcome is the proportion of papers for which it can be established, from located evidence, that a false-positive rate on real photographs of described provenance was obtained at a threshold fixed independently of the test data. Level 2 synthesises studies that evaluate detectors on real photographs whose capture or processing chain is known or controlled (in-camera processing, post-capture processing, generative editing). A reproducible re-analysis of public detector scores illustrates how conclusions change under three threshold rules. Files: protocol, fixed AI screening prompt, coding manual (v0.3) and extraction forms. Reporting follows PRISMA-P, PRISMA 2020 and PRISMA-S. 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: