To tell if a video is AI-generated, trace its source, examine the original file where possible, review motion across time, and compare any detection result with provenance and independent evidence. A strange frame or mismatched lip movement is a reason to investigate—not proof of AI generation.
Video adds a complication that photographs do not have: you must check how the scene changes over time. It may also combine genuine footage, generated inserts, an altered face, subtitles, and replacement audio. A single label can hide those differences.
1. Separate the video from the claim
Write down exactly what you are checking. Is the question whether a person said particular words, whether an event happened at a claimed location, or whether the entire scene was generated?
These questions lead to different checks. A genuine clip can be misleading because it is old, shortened, dubbed, or paired with a false caption. A video can also use AI for one element without being fully synthetic.
For a purported public statement, look for a full recording from an identifiable source. For footage of an event, look for other recordings, reporting, and contextual details. Do not let an AI label substitute for checking the underlying story.
2. Preserve the source and obtain a better copy
Save the post URL, account, caption, access time, and any creator disclosure. Request the original or a higher-quality version if you can. A screen recording of a repost is a weaker starting point for technical inspection than the underlying file.
Keep the original untouched and review a working copy. Record the clip’s duration and whether it appears edited. Note any missing beginning, ending, or sound. Do not assume a short circulating excerpt contains the context needed to interpret it.
Be careful with sensitive footage. Before uploading, consider consent, confidentiality, and the service’s data-handling terms. A public source investigation may answer the question without exposing another person’s private media.
3. Search representative frames and the surrounding context
Choose several clear frames from different moments: a wide view, a distinctive location, and a frame containing recognizable signage or objects. Reverse-search these individually. A blurry transition frame is usually a poor search starting point.
Try both the full frame and a meaningful crop. Search visible phrases and names separately. Compare dates and captions on matching pages, then follow those leads toward an original upload or longer version.
WITNESS’s video-verification resources provide a broader foundation for verifying eyewitness footage. The central task is establishing what a recording supports, not merely deciding whether it looks convincing.
Google’s image-context guidance can help with frames, but an indexed match is only a lead. No match does not mean the video was generated. A private, recent, or heavily altered recording may not have an accessible earlier copy.
4. Watch normally, then examine changes over time
First watch the whole clip at normal speed, with sound if appropriate. Note timestamps worth revisiting. Then slow down those sections and step through adjacent frames. A single paused frame can exaggerate an ordinary transition or blur; the surrounding sequence matters.
Look for consistency in identity, object shape, contact, and movement. Does a person’s clothing change without a cut? Does an object disappear while unobstructed? Does a hand pass through something it appears to hold? These are investigation prompts, not standalone tests.
Check whether an apparent anomaly has an ordinary explanation, such as occlusion, a camera move, a cut, a reflection, or a low-quality copy. Also remember that coherent motion does not establish authenticity: a convincing synthetic sequence may have no obvious visual defect.
| Area to inspect | Question to ask | Important limitation |
|---|---|---|
| Faces and clothing | Are features consistent across adjacent frames? | Blur, cuts, and occlusion complicate comparison |
| Hands and objects | Does contact remain physically coherent? | One unclear frame is weak evidence |
| Background | Do signs and structures persist through camera motion? | Perspective and editing can change appearances |
| Reflections and shadows | Do changes fit the scene and movement? | Complex lighting can be unintuitive |
| Suspicious interval | Does the issue persist in a better copy? | A repost may introduce its own artifacts |
Keep timestamped notes. “The cup changes shape between these two moments” is more useful than “The whole video feels fake.”
5. Treat audio as a separate evidence track
Listen for abrupt changes in background sound, edits around important words, and whether the audio belongs to the same apparent setting. Compare a claimed speech with an official full recording or transcript when available.
Lip-sync mismatch does not prove voice cloning or an AI face edit. Dubbing, synchronization errors, and conventional editing are alternative explanations. Likewise, natural-sounding speech is not proof that the named speaker recorded it.
Do not infer a voice-cloning verdict from a visual detector. AIFakeScan’s Basic audio inspection does not classify cloned or synthetic speech, and Basic Scan does not include specialized face-swap classification. Those limits matter especially when evaluating a clip centered on a person’s identity.
6. Inspect provenance without confusing it with truth
If the file has Content Credentials, inspect the validation status, signer, and recorded actions. An AI-related declaration can be useful evidence about the recorded production history. Determine whether it describes generation, an edit, or another action before summarizing it.
The C2PA explainer explains how provenance is bound to media. A valid history does not independently establish that the caption is accurate or that the depicted event happened as claimed. Missing credentials are not evidence of AI generation.
Watermarks are a separate mechanism. Google DeepMind describes SynthID watermarking for text and video, but a vendor-specific verification method is not a universal scanner for all synthetic media. AIFakeScan does not promise access to every invisible-watermark verifier.
7. Understand what a video detector actually checked
Before interpreting a result, ask whether the service reviewed every frame, sampled frames, inspected metadata, or checked available provenance. These approaches have different coverage. A sample can miss a brief insert, and a timestamp may identify an inspected interval rather than an exact edit boundary.
The AIFakeScan video detector should be used with the published methodology. Beta calibration and evaluation remain incomplete. Passive-only results stay Inconclusive; there is no validated accuracy percentage or measured AI probability to use as a verdict. Deep Scan is not publicly available.
If the file cannot be assessed, record that failure rather than treating it as a negative result. If two tools disagree, keep their observations separate and investigate the reason instead of averaging scores.
8. Report the narrowest conclusion the evidence supports
Imagine a clip labeled as a breaking event. Reverse-searching a frame finds a longer upload from an earlier year. You may now have evidence against the current date claim. That does not, by itself, show the footage is AI-generated.
Alternatively, a creator might explicitly disclose a synthetic production and provide matching provenance. You can report that attributed evidence while still distinguishing it from an independent assessment of every frame.
Use a short verification record: source URL, file version, claim, timestamps examined, findings, and unresolved questions. “Unverified” is a legitimate outcome. For consequential accusations, obtain original material and qualified review before acting.
Frequently asked questions
Can one screenshot prove a video is AI-generated?
Usually not. A screenshot removes motion and audio context and may capture an unrepresentative moment. It can help locate an earlier source, but assess the sequence and original file too.
Is unnatural blinking a reliable deepfake test?
No single behavior should decide the case. Look for corroborating evidence across the source, sequence, and production history rather than relying on a fixed checklist of visual tells.
Can a video be genuine but still misleading?
Yes. The date, location, caption, or surrounding context may be wrong even when the recording itself is genuine. Verify those claims separately.
Does a detector check every frame and the speaker’s voice?
Not necessarily. Read the service’s stated coverage. Frame sampling, face-swap classification, and synthetic-speech detection are different capabilities; do not assume one includes the others.
Your next step
Save the source, select representative frames, and document the claim before running a scan. Use the AI video detector for supported checks, then combine its evidence with your source investigation. An honest unresolved result is more useful than an unsupported confident label.
Sources
- Verifying Eyewitness Video — WITNESS
- Learn more about an image — Google
- C2PA and Content Credentials Explainer — C2PA
- Watermarking AI-generated text and video with SynthID — Google DeepMind