Deepfake detection tips: which ones the controlled studies support

This page belongs to the deepfake detection hub and sits beside the signs of a deepfake page, which lists what people can see by manipulation type. Here the question is turned round: for each tip that circulates, what did a study measure when people tried it?

01The tips, and what measured them

The table pairs each common tip with the study condition closest to it. A tip is only marked as supported where a published measurement shows it moved accuracy; the figure is the accuracy people reached while following it.

Three of the eight tips have a measurement above 60 percent. Two of those are not visual tips at all: pooling judgements and reading a signed manifest. The third, watching uncompressed frames, falls back to 58.73 percent once the video is compressed the way platforms compress it.

02Why do visual tips score so close to chance?

The tells that the tips describe were real, and they dated quickly. The Which Face Is Real learn page, written by the game's authors at the University of Washington, lists water-splotch blobs, background problems, asymmetric glasses and earrings, hair that behaves like paint and extra teeth, with the caveat that "We can't expect future algorithms to have this problem". The same page records that the site's one hard rule, that StyleGAN could not produce two images of the same fake person, lasted three months after the February 2019 launch before a paper animated a face from a single image. A tell is a property of one generator's output, and generators are replaced faster than the advice about them.

The PNAS study is the direct test of the tells. Its second experiment gave 219 participants training on exactly these cues plus feedback after every answer, and they reached 59.0 percent; the authors write that "there was no improvement in accuracy over time" within the session (PNAS 119(8)). A tip that is taught, practised and corrected 128 times in a row still leaves the viewer wrong two times in five.

03Does watching the video rather than a still help?

Video gives a viewer motion, and the FaceForensics++ user study shows what motion is worth. Its 204 participants scored 68.69 percent on raw frames, 66.57 percent on high-quality compressed video and 58.73 percent on low-quality video (arXiv 1901.08971). That is the best figure for an individual anywhere on this page, and it depends on the viewer seeing the uncompressed frames. The same study, broken down by manipulation type, puts the two reenactment methods at or below chance even on raw frames.

The Köbis study used the harder clips: 16 videos of about 10 seconds from the 3,000 DFDC videos that were hardest for the challenge classifiers. Its 210 participants scored 57.6 percent and guessed "authentic" 67.4 percent of the time, and the authors report that "neither raising awareness nor financial incentives increase detection accuracy" (iScience 24(11)). The same participants "overestimate their own detection abilities", which is the finding behind the tip about instinct: the synthetic faces in the PNAS study were rated more trustworthy than the real ones, 4.82 against 4.48.

04Which tips have evidence behind them?

Pooling judgements does. In Groh et al., recruited individuals identified DFDC holdout videos correctly in 66 percent of attempts, while the crowd mean reached 74 percent for recruited participants, 80 percent for nonrecruited participants and 86 percent for the 1,879 nonrecruited participants who had seen at least 10 videos, against 80 percent for the leading challenge model on the same 50 clips (PNAS 119(1)). Showing participants the model's prediction raised recruited accuracy from 66 to 73 percent, but a confident wrong prediction pulled accuracy on one clip from 56 to 34 percent. A second opinion helps; a second opinion from a machine helps only when the machine is right.

Reading a signed manifest is a different kind of evidence. A C2PA Content Credential is a set of assertions bound into a manifest and signed; the specification states that "The basis of making trust decisions in C2PA, our trust model, is the identity of the signer" (C2PA 2.5). It does not rely on a viewer's eye at all, which is why its limits are different too: the manifest can be stripped with the file's metadata, and the sibling page on whether Content Credentials survive upload sets out what the platforms say they do with it.

Running a detector sits between the two. On fakes that circulated in 2024, the best open-source video detector off the shelf scored 0.63 AUC and two others scored 0.50 and 0.51, while the best anonymous commercial model scored 0.79 (arXiv 2503.02857). The same paper names the curated training sets, including the ForgeryNet dataset with its 15 manipulation methods, as using "outdated manipulation techniques" compared with what circulates. A detector score is evidence, and the published figures say how much.

05What should a checklist look like?

The order follows the evidence. Provenance first, because a valid Content Credential or a traceable original settles the question without perception. Pooling second, because many judgements beat one. A detector score third, read with its published in-the-wild figure. Visual inspection last, because every controlled study puts a lone viewer between 48 and 69 percent, and the artefacts the tips describe belong to the generators of 2019. The step-by-step version is the checking a face video page; the generator side of the story, including why GAN and diffusion faces leave different artefacts, explains why a fixed list of tells cannot last.

No single tip, visual or technical, raises a lone viewer's accuracy on synthetic faces far above chance in the published studies; the gains come from provenance and from pooling.