Last checked: September 30, 2026
TL;DR
- Start with provenance. Look for Content Credentials, platform AI labels, or a detectable SynthID watermark before judging pixels.
- Check text, object permanence, physics, faces in motion, and audio across multiple frames, not one screenshot.
- Treat hands and edge artifacts as weak clues in 2026. Better models increasingly handle them well.
- Use an AI video detector as supporting evidence, not a final verdict. Detector accuracy varies widely outside controlled benchmarks.
To tell if a video is AI generated, check provenance data first, then look for persistent visual and audio inconsistencies across frames. Content Credentials or a detected watermark provide stronger evidence than unusual hands or faces. AI detector scores can help, but they are probabilistic signals and should not be treated as proof on their own.
Can you reliably tell if a video is AI generated?
Sometimes, but no single visual trick reliably identifies every AI video in 2026.
A useful check separates four kinds of evidence:
Provenance tells you how a file was created or edited. C2PA Content Credentials can contain cryptographically verifiable information about origin, tools, edits, and AI use.
Watermarks are signals embedded by participating generation systems. Google DeepMind's SynthID, for example, can mark AI-generated or edited media created with supported Google models.
Detectors analyze a file and estimate whether it resembles AI-generated media. They do not know the file's production history.
Human observation looks for visual, audio, physical, and contextual inconsistencies.
The strongest conclusion usually comes from combining these evidence types instead of relying on one.
The 9 signs to check
| Sign | What to look for | How reliable in 2026 |
|---|---|---|
| Content Credentials / platform AI label | Verified origin or explicit AI disclosure | High when present |
| SynthID | Google AI watermark detected by supported tools | High when detected |
| Text and logos | Letters or packaging changing between frames | Medium |
| Object permanence | Objects appearing, disappearing, or transforming | Medium |
| Physics | Unnatural water, cloth, collisions, crowds, gravity | Medium |
| Faces in motion | Teeth, ears, jewelry, blinking, identity drift | Medium |
| Audio | Lip-sync errors, strange room tone, breath, loops | Medium |
| Hands and edges | Fingers, hair, limbs, object boundaries | Fading |
| Source and corroboration | Original uploader, other footage, authoritative sources | Medium for AI origin, high for event verification |
1. Check for Content Credentials or an AI label
The strongest first check is whether the file or platform already tells you how the video was made. Reliability: High when present.
YouTube now shows a “Made with AI” disclosure for meaningfully altered or synthetically generated video when creators disclose it or YouTube receives qualifying signals. Meta similarly uses an “AI info” label for detected or disclosed AI content.
Content Credentials go further. They can show creation and edit history rather than only a binary label.
A label is useful evidence, but it does not tell you whether the underlying claim is true. For the deeper distinction, see why AI labels don't build trust.
2. Check for a SynthID watermark with Google's tools
A detected SynthID watermark is strong evidence that all or part of the video was created or edited using supported Google AI. Reliability: High when detected.
Signed-in Gemini users can ask whether an uploaded video is AI generated. Gemini checks for SynthID and, where available, Content Credentials. Google says Gemini can currently recognize SynthID from Google AI tools, even though other companies have begun adopting the technology.
If detected, Gemini can also identify portions of a video where the watermark appears.
3. Look at text, logos, and packaging across frames
Do not only check whether text looks correct in one frame. Watch whether it stays correct while the camera or object moves. Reliability: Medium.
A package may read “ORBIT” at the beginning and become “OR8IT” after it rotates. Tiny logo geometry may also shift between frames.
Current generators are much better at text than earlier systems, so one correct label proves little. Persistent changes over time are more useful than a single misspelled word.
4. Watch object permanence
Look for objects that appear, disappear, merge, or quietly change identity during motion. Reliability: Medium.
Watch background furniture, jewelry, utensils, vehicle parts, handles, reflections, and items passing behind people.
A coffee cup may lose its handle after an occlusion. One person in a crowd may briefly merge with another. These are temporal consistency failures, and they are easier to notice when you replay a short section frame by frame.
5. Check physics
Watch whether objects continue to behave consistently under gravity, contact, collision, liquid flow, and deformation. Reliability: Medium.
Water may splash in the wrong direction. Cloth may pass through an arm. A dropped object may change momentum unnaturally.
Sora, Veo, Kling, Seedance, and other current models have improved substantially, so unrealistic physics is not guaranteed. Use it as one clue, especially during complex or sustained interactions.
6. Look at faces in motion
A convincing still face can still drift when the subject talks, turns, or becomes partly hidden. Reliability: Medium.
Watch teeth, ears, earrings, glasses, hairlines, eye shape, and facial proportions during head turns.
Lip movement may also briefly stop matching speech. The useful question is not “Does this face look strange?” It is “Does the same identity remain geometrically stable throughout the movement?”
7. Listen to the audio
Audio can reveal inconsistencies that the image hides. Reliability: Medium.
Listen for lip sync that slowly drifts, breaths that disappear between phrases, room tone that changes unexpectedly, repeated music textures, or environmental sounds that do not respond to what happens on screen.
Native AI audio is improving quickly, so these clues are weaker than they were. Check whether sound and image remain causally connected.
8. Check hands, hair, and edges around overlapping objects
Hands and edge artifacts are now weak evidence on their own. Reliability: Fading.
Older advice often starts with finger counting. That is increasingly outdated.
As of September 2026, hands are no longer a reliable sign in top models.
Instead, inspect difficult interactions: fingers gripping small objects, hair crossing a face, a hand passing behind a cup, or limbs crossing clothing. Temporal edge failures are more informative than whether one hand contains five normal-looking fingers.
9. Check the source, not just the pixels
Ask who posted the video, where it first appeared, and whether independent evidence shows the same event. Reliability: Medium for AI origin, high for verifying the event itself.
Search for the earliest upload. Check the person's or organization's official channels. Look for other camera angles, credible news footage, livestreams, photographs, or eyewitness material.
A real-looking file from an unknown repost account is weaker evidence than an imperfect-looking clip whose origin can be independently established.
How do you check Content Credentials on a video?
Content Credentials are based on C2PA, the Coalition for Content Provenance and Authenticity's open provenance standard. They are cryptographically signed and can record information about how a file originated and changed.
To check a video:
- Get the highest-quality or original file available. Reposts may remove ordinary metadata.
- Look for the
crContent Credentials icon on a supporting platform. - Upload the file to a Content Credentials inspection tool. The official Verify service supports formats including MP4, MOV, and AVI.
- Review the origin and history. Look for the recording app or device, AI use, editing actions, signer, and available previous versions.
- Treat the credential as provenance, not proof of truth. A signed AI video can still depict an invented event.
Adobe also provides a free Inspect tool that can reveal who created supported content, how it was created, and whether generative AI was involved.
Do AI video detectors work?
They can provide useful signals, but detector scores are not reliable enough to serve as proof by themselves.
An AI video detector usually analyzes visual or temporal patterns and estimates whether a clip resembles examples generated by AI. Performance can fall when the model, compression level, editing process, or source material differs from its training data.
The gap can be large. A 2026 study tested 95 AI deepfake detectors and found that their accuracy on a lower-quality, real-world-style video benchmark fell to 53.7%, close to chance, while human participants reached 78.4%. Other benchmarks have found substantially higher results under different conditions, which is exactly why one accuracy number should not be generalized to every detector or every type of AI video.
Use detector results like this:
“This detector found signals consistent with AI generation.”
Not:
“This detector proves the video is fake.”
False positives matter too. A real compressed, edited, low-light, or unusual video can look statistically unfamiliar to a detector.
Why does “no watermark” not mean “not AI”?
The absence of a watermark or Content Credential is not evidence that a video is human-made.
This asymmetry is critical.
Google explicitly says that if Gemini does not detect SynthID, the video could still have been created by another AI system. Even supported watermarks can sometimes become undetectable after enough alteration.
Content Credentials are also not universal. Many creation tools do not add them. Some platforms strip embedded metadata during processing. Durable implementations may recover credentials through fingerprinting or watermarking, but that does not apply to every file.
Positive provenance evidence can be strong. Missing provenance proves nothing.
What five-step verification workflow can you reuse?
For journalists, moderators, marketers, teachers, or anyone repeatedly asking how to check if a video is AI generated online, use the same order each time:
- Check provenance first. Look for platform labels, Content Credentials, and supported watermarks.
- Check the source. Find the earliest upload and look for independent evidence of the event.
- Inspect the video over time. Review text, objects, physics, faces, and audio across frames.
- Use detectors only as secondary evidence. Run more than one if the decision matters.
- Record the conclusion and uncertainty. If the signals disagree, classify the video as unverified rather than forcing a yes/no answer.
This order prevents a weak visual artifact from outweighing stronger provenance or source evidence.
What should you do if you think a video is fake?
Do not amplify the clip while you are still checking it.
Look for the original source and the depicted person's official channels. Search for independent footage of the same event. If the video violates a platform's rules, use that platform's reporting process.
If the content depicts you without authorization, use the platform's privacy, impersonation, likeness, or synthetic-media reporting options and preserve the URL, screenshots, upload date, and a copy of the media where appropriate.
What does this mean if you make AI video?
Make realistic AI content easier to verify, not harder.
Follow relevant platform disclosures. Keep your original files and production records. Preserve Content Credentials when your workflow supports them.
For publishing requirements, review YouTube's AI disclosure rules. For commercial work, review rights and licensing for AI video, which covers input rights, likeness, voice, music, trademarks, and recordkeeping.
AKOOL also documents consent, human review, automated moderation, and user-control practices on AKOOL Trust and Safety.
If you create AI video, do it transparently. The AKOOL AI video generator gives you access to multiple current video-generation models, but disclosure and verification remain part of the publishing workflow.
