What the researchers found

People were less successful at identifying synthetic videos in DF26 than in two other datasets evaluated in the same study. The finding points to a practical problem: a convincing appearance is not enough to establish whether a video is genuine.

A preprint posted to arXiv on September 7, 2026 introduces DF26, a benchmark containing 271 real and 2,420 synthetic videos from seven generators. The clips show a single person addressing a camera, audience or interviewer. In the human study, participants correctly identified 52.6% of DF26's synthetic videos, compared with 74.5% in CelebDF++ and 69.8% in DeepSpeak v2. These are comparisons between datasets, not evidence that the same viewers became worse at detecting fakes over time. [1]

Accuracy on DF26's real videos was 76%. That measures a different part of the task; neither percentage is an overall score. Participants were volunteer students, mostly from technical universities, and judged five-second clips. The researchers recorded 232 labeling sessions without recording participant identities. The benchmark evaluates visual evidence, not audio realism or audiovisual synchronization. [1]

Why the result has limits

The results concern these participants, clips and test conditions. They do not establish that every online video is equally difficult to assess. Automated detectors also varied in performance, and a separate retraining experiment improved some results. The study identifies a detection challenge, not a fixed limit on what detection can achieve. [1]

For a reader, the useful distinction is between judging a video's appearance and checking the claim attached to it. Imagine a clip presented as a company's announcement. A realistic face tells you little about whether that company released it, when the statement was made or whether the caption describes it accurately.

Check the record around the video

Provenance offers another kind of evidence. C2PA's Content Credentials can carry signed records of how media was created or changed. The standard also supports ways to reconnect a file with credentials separated from it. That recorded history is different from a visual detector's assessment of the pixels. [2]

A file's history does not, by itself, establish the truth of what its speaker says. For the hypothetical company announcement, a useful next step would be to look for the original statement on the company's own channels, check its date and compare it with other records of the event.

Those checks are an editorial recommendation, not a method tested by DF26. Their purpose is to widen the evidence beyond whether a face or movement looks convincing.

Sources & context

  1. DF26: We Cannot Tell Fake From Real AnymoreShykula et al. · arXiv preprint, September 7, 2026

    Primary research. Numerical results and study limits are the authors’ reported findings; no independent replication was performed.

  2. Content Credentials FAQCoalition for Content Provenance and Authenticity

    Primary documentation for the provenance approach. It is not an independent assessment of deployment effectiveness.

AI assisted with preparing and checking this article. The responsible editor accepted this version for publication following an assisted review of facts, sources and attribution.

Related reading