Deepfake statistics: figures that trace to a study, and those that do not

This page is part of the deepfake detection hub. It gathers every figure the site quotes into one place with its study, year and what it measures, and it ends with the figures that circulate without any such source. The detector numbers come from the published results table, which carries their full conditions.

01Why this page lists so few numbers

A deepfake statistic is admitted here only if it was read from a primary source: the paper that ran the experiment, the dataset card, or the official report that states the figure. A press summary of a study does not count, and neither does a survey published by a company that sells detection, because its respondent count, question wording and definition of "deepfake" are not available for checking. That rule removes most of the numbers that appear in headlines, and what remains is smaller and more specific than the headlines suggest.

Each figure below is stated with its base: how many participants or clips, which test set, what task and what year. Those conditions are not decoration. A 48 percent accuracy on still faces and a 66 percent accuracy on video clips come from different tasks with different chance levels, and the figure of a detector on its own benchmark and the figure of the same detector on circulating fakes are, as the next sections show, different numbers.

02How often do people spot deepfakes?

For fully generated still faces, close to never above chance. Nightingale and Farid showed 400 StyleGAN2 faces and 400 matched real faces to 315 participants, 128 each; accuracy was 48.2 percent (95 percent confidence interval 47.1 to 49.2), where guessing would give 50 (doi 10.1073/pnas.2120481119). A further 219 participants given training and feedback reached 59.0 percent, with no improvement over the session, and 223 participants rated synthetic faces more trustworthy than real ones, 4.82 against 4.48 on a 7-point scale. The generator behind those faces, and the 70,000 photographs it learned from, are described on the how AI faces are generated page.

For face-swap video the figures are higher but far from certain. In the FaceForensics++ user study, 204 participants scored 68.69 percent on uncompressed clips, 66.57 percent at high quality and 58.73 percent at low quality (arXiv 1901.08971). Köbis, Doležalová and Soraperra showed 210 participants 16 clips drawn from the 3,000 hardest videos of the Deepfake Detection Challenge; overall accuracy was 57.6 percent, participants guessed "authentic" 67.4 percent of the time, and neither an awareness message nor a payment improved the result (doi 10.1016/j.isci.2021.103364). Groh, Epstein, Firestone and Picard ran two online studies with 15,016 participants: recruited participants identified deepfakes in 66 percent of attempts on randomly sampled DFDC holdout videos, the leading challenge model scored 65 percent on the full 4,000-video holdout set, and the crowd mean reached 74 percent for recruited participants (doi 10.1073/pnas.2110013119).

03How accurate are detectors on their benchmarks?

On the set they were trained on, very. XceptionNet trained on all four FaceForensics++ manipulations reached a binary accuracy of 99.26 percent on raw video, 95.73 percent at high quality and 81.00 percent at low quality, in the paper's Table 1. In DeepfakeBench, where every detector is trained on FaceForensics++ c23 and scored by frame-level AUC, Xception scores 0.9637 on that split and the best in-dataset score is UCF's 0.9705 (arXiv 2307.01426).

Tested on a dataset the detector never saw, the same models fall. Xception's AUC goes from 0.9637 on FaceForensics++ c23 to 0.7365 on Celeb-DF v2 and 0.7077 on DFDC; its average over the eight cross-dataset columns is 0.7718. The best cross-dataset average in the table, SPSL's, is 0.7875. Ten of the fifteen detectors score above 0.93 on the FaceForensics++ c23 split and none averages above 0.79 on the eight datasets it never saw. Eight of the fifteen share one Xception backbone, which the face recognition backbones in detectors page counts from the paper's Table 1.

04How far do detectors fall in the wild?

Deepfake-Eval-2024 collected 45 hours of video, 56.5 hours of audio and 1,975 images that circulated on 88 websites in 52 languages during 2024. Compared with their original benchmarks, open-source detectors' AUC fell by 50 percent for video, 48 percent for audio and 45 percent for images (arXiv 2503.02857). The best off-the-shelf open-source video model, GenConViT, scored 0.63 AUC after scoring 0.96 on its own benchmark; FTCN and Styleflow scored 0.50 and 0.51, at chance. Commercial video models, tested anonymously, ranged from 0.64 to 0.79. Fine-tuning GenConViT on part of the set raised it to 0.82, a different protocol because the model had then seen the distribution.

The earlier in-the-wild set tells the same story at smaller scale. On WildDeepfake, image-level detectors that scored 90 to 99 percent accuracy on FaceForensics++ or DeepFake-TIMIT fell below 70 percent, and the authors' own method reached 76.25 percent (arXiv 2101.01456). For audio, Müller et al. report "performance degradation of up to one thousand percent" on 37.9 hours of found recordings (arXiv 2203.16263). The in the wild page sets these results against the lab figures above.

05How large are the datasets behind these numbers?

Every count is taken from the dataset's own paper or repository as the hub's dataset comparison records it. The DFDC paper also reports that the winning model kept a ROC-AUC of 0.734 on a small set of real in-the-wild deepfakes in 2020 (arXiv 2006.07397), the only published figure of its kind from before 2024.

06What do the official reports actually state?

Two documents are cited constantly for deepfake figures. The FBI Internet Crime Complaint Center's 2024 annual report records 859,532 complaints and reported losses of 16.6 billion dollars, a 33 percent increase from 2023, with 256,256 complaints reporting an actual loss (2024 IC3 Report). It does not contain the word deepfake, and it attributes no complaint count and no loss figure to deepfakes or synthetic media; generative AI appears only in the titles of two public service announcements listed in an appendix. Any dollar figure for "deepfake fraud losses" credited to the FBI is not in this report.

Europol's Innovation Lab report "Facing reality? Law enforcement and the challenge of deepfakes" (2022, reissued January 2024) is a foresight report, not a measurement (Europol, 2022). The counts it carries are attributed to other parties: a commercial detection company's finding of 85,047 deepfake videos on streaming sites in December 2020, "with the number doubling every 6 months", that company's 2019 finding that "96 % of the fake videos involved non-consensual pornography", a 2019 UK survey in which "almost 72% of people" were unaware of deepfakes, and a press report of a voice deepfake used in a 35 million dollar transfer. The report publishes no count of its own.

07Which repeated figures have no primary source?

Some numbers appear in nearly every article on deepfakes and in none of the primary sources above.

• The claim that 90 percent of online content will be synthetic by 2026, usually credited to Europol. The current version of the Europol report fetched for this page does not contain it.
• Totals for money lost to deepfake fraud in a year, credited to the FBI. The 2024 IC3 report has no deepfake category; its 16.6 billion dollar figure covers all reported internet crime.
• Counts of deepfakes online, or the rate at which they double. The only published counts trace to one company's scans of specific sites in 2019 and 2020, with the method not available for checking.
• Percentages of people who "cannot tell a deepfake" or who have "encountered one", from surveys run by companies that sell detection or identity products, without a published respondent count or protocol.
• Deepfake detection market sizes and growth rates, which measure spending forecasts rather than deepfakes.

These figures may or may not be close to the truth; the point is that nobody can check them. A figure on this site comes with its study, its base and its date, or it does not appear. The same standard applies to provenance: no primary source has yet published how many circulating files carry Content Credentials (C2PA), so that page describes what a credential proves and stops there.