How to spot AI-generated videos: 10 verification steps
Quick answer
To check whether a video is AI-generated, trace its original source, look for credible corroboration, and inspect available provenance or watermark evidence. Unusual hands, mismatched lips and detector scores can justify further investigation, but none is a universal proof that a video is fake. A convincing-looking clip can also be synthetic.
If the video asks you to send money or take an urgent action, verify the request through a separate trusted channel first. You do not need to identify the generation model to avoid acting on an unverified claim.
For the underlying definitions, risk boundaries, and a companion verification process, read what a deepfake is and how it works.
A red ceramic mug sits on a wooden café table beside a window. Steam rises slowly while the camera makes a gentle push-in. Soft morning light, realistic materials, one continuous shot, no people or readable text.
See how an AI-generated video behaves
Run one short prompt, inspect the complete clip, and compare what you observe with the verification cues in this guide.
Try AI Video GeneratorWhat evidence is useful?
Evidence | What it can establish | What it cannot establish alone |
|---|---|---|
Original upload and source context | Who published the clip, when, and what they claimed | That the depicted event really happened |
Independent reporting or another recording | Corroboration of a specific event or detail | That every frame and audio track is unaltered |
Valid Content Credentials | Recorded provenance and edits covered by the credential | That the scene or caption is factually true |
A detected provider watermark | Evidence of content made or edited with supported AI tools | That all parts of the video are synthetic |
Visual artifacts or a detector score | A reason to investigate particular details | A definitive verdict on any arbitrary video |
This is an evidence checklist, not a scored detection test. A useful conclusion might be “the event is corroborated, but the voice track remains unverified.” Do not force a whole video into a single real-or-fake label when the evidence is narrower.
1. Save the source and the exact claim
Record the original URL, account name, upload date and caption. Identify what you are checking: the location, a person's identity, a spoken statement, or whether the footage was generated. Those are different questions.
For example, a real video with a false date can mislead viewers without any AI involvement. Start with the claim that matters instead of assuming the only explanation is generation.
2. Find the earliest available upload
Search for distinctive caption text and look for earlier versions of the clip. Check whether the account publishes satire, demonstrations or openly labeled AI work. A repost can omit the creator's disclosure or attach a new story to an unrelated scene.
Treat the earliest upload you find as a lead, not proof that you found the original. Keep a note of what remains unknown.
3. Reverse-search several frames
Choose clear frames from different moments and search them for earlier appearances. A match may uncover the original footage, a different location or an existing fact check. Try both the full frame and a crop around a distinctive landmark or object.
The InVID verification toolkit supports video fragmentation and keyframe reverse search. It helps investigate origins; a search with no match does not prove a video is new, authentic or AI-generated.
4. Look for independent corroboration
Search for the event itself, not just reposts of the same clip. Look for a separate camera angle, a named witness, an original statement or reporting that explains its evidence. Ten accounts sharing one file are still repeating one source.
For a public announcement, check the organization's official publication. For a location claim, compare visible landmarks and the claimed circumstances. Document both matches and contradictions rather than selecting only details that support your first impression.
5. Inspect Content Credentials when available
Content Credentials can record how an asset was created or edited using signed provenance data. Inspect the credential's source and the actions it records, rather than treating the presence of a badge as the verdict.
The Content Authenticity Initiative FAQ explains that credentials provide a positive signal about recorded history, not a negative signal when absent. Files can lose metadata as they move between services. A signed photograph of a screen can also record a capture without establishing whether the scene on that screen was genuine.
Use the Content Credentials inspection tool for supported files. If there is no credential, record “no credential available,” not “confirmed AI.”
6. Use provider-specific watermark verification
Google lets users check supported videos in Gemini for SynthID. Upload a video and ask whether Google AI created or edited it. Google's documented video limits are 100 MB and 90 seconds, and the result can distinguish evidence in the audio from evidence in the visuals.
This checks for Google's watermark; it is not a universal detector for every generator. Read the returned scope carefully. “No SynthID detected” does not establish that a video was filmed with a camera or rule out another provider's tools.
7. Examine motion and objects across time
Look for changes in an object's shape, continuity or interaction with the scene. Review the same moment at normal speed as well as frame by frame. Describe the actual observation—such as “the label changes between frames”—instead of announcing a model or verdict you cannot substantiate.
Low quality can hide detail, and unusual detail is not automatically evidence of generation. In BBC interviews with digital-forensics researchers, experts explain why blurry footage merits closer attention without making blur itself evidence of a fake.
8. Check audio separately from the picture
A video's image and soundtrack may have different origins. Look for the original recording of an alleged statement, compare a longer version, and check whether a translation, dub or edit is disclosed. A lip-sync mismatch is a lead to investigate, not proof of an AI voice.
For a request from someone you know, use an established contact method to verify it. The FTC's guidance on voice-cloning scams recommends contacting the person using a number you already know is correct rather than trusting the voice in the incoming request.
9. Treat detector results as supporting evidence
Before relying on a detector, check what media and generation methods it supports, what its score means, and whether it publishes relevant validation. Keep the tool name, result and version or date alongside the file you checked.
A confident-looking percentage is not a substitute for that information. If two tools disagree, or neither has demonstrated applicability to your file, retain the uncertainty. For a consequential allegation, involve a qualified analyst rather than publicly identifying someone on the strength of a score.
10. State the conclusion at the level the evidence supports
Separate what you observed from what you inferred. A short verification note can include:
- Claim checked: what the video is alleged to show.
- Evidence found: original source, corroboration, credentials or watermark results.
- Unresolved parts: missing original file, uncertain audio or unsupported detector result.
- Conclusion: verified detail, identified alteration, or insufficient evidence.
Do not share a clip as factual evidence simply because you failed to disprove it. Likewise, do not label a real person or event fake because one frame looks strange.
Frequently asked questions
No. Those features are not a universal test. Investigate the source and corroborating evidence; a visual anomaly alone does not establish how the video was made.
No. The file may never have had credentials, or they may be unavailable in the copy you received. Absence leaves a gap in evidence rather than proving a particular origin.
No. A clearly labeled fictional scene, illustration or creative effect can be legitimate. Deception depends on how it is represented. Evaluate the claim and context as well as the media's origin.
Yes. Its soundtrack, face, background or individual shots can be edited separately. That is why provenance and verification conclusions should describe the part of the video they actually cover.
Runbo Li is the Co-founder and CEO of Magic Hour, where he builds AI video and image tools for content creation. He is a Y Combinator W24 founder and former Data Scientist at Meta, where he worked on 0-1 consumer social products in New Product Experimentation. He writes about AI video generation, AI image creation, creative workflows, and creator tools.
