Deepfake detection outside the lab
Deepfake detection appears on 45 recent papers and AI-generated image detection on 13, almost all in AI journals, with 23 papers in the last quarter against about 21 per quarter before. Detectors score well on the datasets they were built with and fail on the next generator. How well do they work on real fraud, elections and markets, and who should run them?
What prompted this
Since October 2025, 86 papers carry a deepfake, face-forgery, synthetic-media or AI-generated-image keyword (78 in AI journals, 7 in management): "deepfake detection" on 45, "AI-generated image detection" on 13, "deepfakes" on 13, "synthetic media" on 7, "face forgery detection" on 6. The quarter to September 2026 had 23 such papers against about 21 per quarter earlier.
Questions
- On deepfakes collected from actual fraud cases, political campaigns and market rumours, rather than from research datasets, what is the detection rate of published detectors?
- How quickly does accuracy fall after a new generator is released, and can detectors be updated faster than generators?
- Do labels from platform detectors change how people react to a video, or do people believe what they expected to believe?
Data that could help
Verified deepfake collections from fact-checkers and fraud investigations, detector code from published papers, and platform labelling experiments.
Next step
Assemble 500 deepfakes verified by fact-checking organizations and measure the accuracy of five published detectors on them.