Checking a face video or photo yourself: what the benchmarks say works, in order of evidence
The Face Benchmark editorsPublishedFigures checked
The strongest evidence about a face video is where it came from, not how it looks. Start with the first place it was posted and who posted it, then any recorded capture or edit history, then whether separate sources report the same event. Looking closely at the face comes fourth, because controlled studies put untrained observers below chance on synthetic faces, and a detector's score comes last, because published in-the-wild results put the best tools between a coin toss and four in five.
STPThe steps, strongest evidence first
Work down the list. Each step says what it can prove, so you know when to stop and when a later step adds little.
- STRONGEST
Step 1: Find where it was first posted
Search for the earliest copy you can find, using a reverse image search on a still frame and the clip's caption or any text in it. Note who posted it first, when, and whether that account has any connection to the event. A clip whose first appearance is an anonymous account with no history, or that appears everywhere at once, is unsupported; a clip traceable to a named source present at the event is close to settled. Nothing later in the list can overturn a confirmed original source.
- STRONG
Step 2: Look for content credentials and capture data
Check whether the file carries Content Credentials, the open C2PA standard that records the origin and edits of a file in a signed manifest, or other capture metadata. If credentials are present and verify, they establish who signed the file and what was done to it. Their absence proves nothing either way: most genuine clips carry none, and platforms often strip them on upload.
- MODERATE
Step 3: Find separate copies and reports
Look for other footage of the same moment from different angles, and for reports by people or outlets who were there. Separate copies that agree are hard to fake together. A single clip of an event that no one else recorded or reported, especially a newsworthy one, deserves more suspicion than the clip's own appearance can resolve.
- WEAK
Step 4: Check the clip for inconsistencies
Watch at full resolution and slowed down. Look at the edge where the face meets hair, neck and background, lighting that does not match between face and scene, teeth and ears, and whether mouth movement matches speech. Weigh what you find lightly: in controlled studies people told synthetic faces from real ones at 48.2% accuracy untrained and 59.0% after training (Nightingale and Farid, 2022), and compression adds flaws to genuine clips too.
- WEAKEST
Step 5: Run a detector, and read its score with its test set
If you use a detector, find out which dataset it was tested on and what its published cross-dataset score is, then treat its output as one more weak signal, not a verdict. On fakes that circulated in 2024 the best off-the-shelf open-source video detector scored 0.63 AUC and the best commercial one 0.79 (Deepfake-Eval-2024), so a confident-looking score can be wrong often. A score never tells you where the clip came from.
This checklist is part of the deepfake detection hub. It orders the steps by what the published studies show each one can establish; the can people spot AI faces page carries the human evidence and the in the wild page the detector evidence.
LIMLimits
- L1No single step, and no combination of them, proves that a clip is real. They raise or lower confidence.
- L2Benchmarks measure detectors on curated datasets; clips from platforms are compressed, cropped and re-encoded in ways those sets may not cover, and the cross-dataset generalisation page shows how far scores fall when the fakes change.
- L3New generation methods appear faster than datasets are published, so a detector's published score can age quickly, and the visible tells taught by guides belong to older generators.
- L4If a clip could cause real harm, take it to a professional verification desk rather than relying on these steps alone.
FAQQuestions
Can I tell a deepfake just by looking at it?
Sometimes, but not reliably. In the largest controlled study, people classified synthetic faces at 48.2% accuracy without training and 59.0% with it, and compression adds flaws to real clips as well. Look closely after checking where the clip came from, not instead.
What is the most useful single check?
Finding where the clip was first posted and by whom. Provenance can settle a question that no amount of looking at the face can.
Do online deepfake detectors give a reliable answer?
They give a score from a model whose reliability depends on how similar the clip is to its test data. On circulated 2024 fakes, open-source video detectors scored between 0.50 and 0.63 AUC off the shelf and the best commercial one 0.79, so the score is a weak signal.
What if the clip has no content credentials?
That proves nothing either way. Most genuine clips carry none, and platforms often strip them on upload. Credentials that are present and verify are strong evidence; their absence is not.
When should I stop checking?
When an early step settles the question, for example a confirmed original source present at the event. If the stakes are high and the steps disagree, take it to a professional verification desk.
SRCSources
- S1AI-synthesized faces are indistinguishable from real faces and more trustworthy(external)Nightingale, Farid. PNAS 119(8) e2120481119, 2022Checked
- S2Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024(external)Chandra et al. 2025Checked
- S3Content Credentials: C2PA Technical Specification(external)Coalition for Content Provenance and AuthenticityChecked
- S4FaceForensics++: Learning to Detect Manipulated Facial Images, user study(external)Rössler et al. ICCV 2019Checked