Is this AI generated? What evidence can answer it, in order

This page belongs to the deepfake detection hub and is the still-image counterpart of the checking a face video page. It names no tools. It lists the kinds of evidence, what each can prove and the published study behind each figure.

01What order should the evidence go in?

Evidence about an image falls into five kinds, and they differ in what they can prove. The first three can establish how the image was made or that it records something that happened. The last two can only estimate a probability. The order below runs from the kind that can close the question to the kind that cannot.

02Can a provenance record settle it?

Yes, for an asset that carries one and whose signer can be trusted. The C2PA specification defines statements about an asset, "called assertions", which "make up the provenance of a given asset"; they are signed and "bound together into a verifiable unit called a C2PA Manifest" (C2PA 2.5). A manifest for a generated image begins with a c2pa.created action whose digital source type is trainedAlgorithmicMedia. If that manifest is present and the signature validates against a trusted signer, the question is answered by the party that made the image.

The limits are stated in the specification itself. The trust model "is the identity of the signer", so the record is as good as the signer. "An asset can become separated from its C2PA Manifest due to removal or corruption of asset metadata", and the Explainer adds that "if an asset is cropped using a non-Content Credentials aware tool, the provenance data may not be updated" (C2PA Explainer). The specification also states that it does "NOT provide value judgements about whether a given set of provenance data is 'good' or 'bad'", and the Explainer that "provenance information alone cannot tell you whether the digital content is true, accurate or factual". The Content Credentials page covers the standard in full.

03What does the absence of a record mean?

Nothing on its own. Adding provenance is optional, and the Explainer states that the standard does not intend "a two-tier media ecosystem where assets without Content Credentials are universally less trusted". A photograph saved by software that writes no manifest arrives without one, exactly as a generated image does, so a missing record describes real photographs as well as fakes. The absence moves the question down the list to source, context and corroboration; it does not answer it.

04Can you tell by looking?

Not at a useful rate for synthetic faces. Nightingale and Farid showed 315 participants 128 faces each, drawn from 400 StyleGAN2 faces and 400 matched real ones; accuracy was 48.2 percent with a 95 percent confidence interval of 47.1 to 49.2 (PNAS 2022). A second group of 219 participants given training and trial-by-trial feedback improved to 59.0 percent, with no further gain over the session. A third experiment found the synthetic faces rated more trustworthy, 4.82 against 4.48 on a seven point scale.

For manipulated video the picture is similar but not identical. Groh and colleagues ran two studies with 15,016 participants on Deepfake Detection Challenge videos; recruited participants identified fakes in 66 percent of attempts, and the leading challenge model scored 65 percent on the full 4,000-video holdout, "similarly accurate, while making different kinds of mistakes" (PNAS 2022). On the curated FaceForensics++ set, 204 participants reached 68.69 percent on raw video and 58.73 percent on heavily compressed video. The human studies page tabulates each experiment. Visible tells belong to one generation of one generator: the droplet artefacts in StyleGAN faces were traced to instance normalisation and removed in StyleGAN2 (arXiv 1912.04958).

05What does a detector score prove?

A probability that the image resembles the fakes the detector was trained on. Deepfake-Eval-2024 collected 1,975 images that circulated in 2024, 767 real and 1,208 fake, and ran three published open-source image detectors on them off the shelf: UFD scored 0.56 AUC, DistilDIRE 0.52 and NPR 0.53, down from 0.94, 0.99 and 0.98 on the data they were published with (arXiv 2503.02857). The best of eight anonymised commercial image models reached 0.90 AUC with recall 0.71, so it missed about three fakes in ten. For video the best open-source model scored 0.63 and the best commercial one 0.79.

An AUC near 0.5 means the detector ranks a random fake above a random real image about half the time, which is a coin toss. Such a score is a fifth kind of evidence, not a verdict, and it carries weight only when steps one to three have produced nothing and the detector's training data are known to resemble the image in question. The deepfake AI on both sides page explains why detectors trail generators: both draw on the same pretrained networks and datasets.

06Why does the order run this way?

The order follows from what each kind of evidence asserts. A provenance record is a statement by the maker, signed; source and corroboration are statements by witnesses; inspection and a detector score are inferences from pixels. Statements can be checked and can be false, but when they hold they close the question. Inferences from pixels never close it, because every published measurement of them, human or machine, is near chance on material outside its training set. Running the list the other way, score first, is the order that the figures above say fails most often.

The order also matches how the three kinds of fake still differ. The swapped, edited and generated stills page shows that each kind leaves a different trace and that detectors built for one miss the others, which is one more reason a score cannot lead. The same recognition networks described on the ArcFace angular margin loss page sit inside modern swap generators, so a swapped identity is built to pass the test a verification system would run.