Direct answer: Is this video AI generated? Look for the platform’s AI label first, then check the file’s provenance with the free C2PA Verify tool, then reverse-search a single frame to find where the clip came from. No single check is enough, and our worked test below shows even a famous AI clip can carry no label at all.
A viral clip lands in your feed. A dog in a spacesuit floats past a space station. A gymnast sticks a landing that looks physically impossible. A celebrity appears to say something they never said. The clip is nine seconds long, it has a million views, and the comments are split down the middle. Is it real?
This guide answers that question for fully generated clips, scenes and events that never happened, made by a video generator, rather than face-swapped people or edited real footage. If the suspect file is a still photo rather than a clip, our face swap check walks through the still-photo tells. The workflow is the same one we use for still images, adapted for moving pictures: check the label, check the provenance, check the frame, and only then look at the pixels. It works in 2026, in the post-watermark era, when the most famous video generator of all no longer exists.
What is an AI-generated video clip, exactly?
AI-generated video: a moving clip produced entirely by a generative model from a text prompt, an image, or both, no camera, no real location, no real event. It is different from an edited video (real footage altered afterwards) and from a deepfake (a real person’s face or voice swapped or puppeteered).
That distinction matters, because the checks are different. An edited real video still has a real source you can find. A face-swap still has a real person underneath. A fully generated clip has no original event at all, there is nothing to “find” except the generator’s own demo, a news article about it, or the account that first posted it. That is why the first checks in this guide are about provenance and labels, not about staring at the pixels.
The other thing to know before you start: a video generator’s output is short. Most tools produce clips of five to ten seconds, and viral fakes are usually cut to a single striking moment. If a clip is very short, loops neatly, and shows exactly one impossible thing, that is not proof it is fake, but it is a reason to run the checks below rather than share it.
Why is everyone talking about Sora if Sora is gone?
Because Sora clips are still everywhere, and no new ones can be made. OpenAI discontinued the Sora web and app experiences on 26 April 2026, and the Sora API followed on 24 September 2026, according to OpenAI’s own help centre, which now points users to an export page for anything they created. Sora 2, launched on 30 September 2025, was the last version. Any “new Sora clip” you see in late 2026 is either an old clip reposted, or a clip from a different generator being passed off as Sora.
That changes the check. A Sora video check in 2026 is not about the tool, it is about the clip’s history. If someone posts a “Sora” clip dated after April 2026, the date alone is suspicious. And because Sora is gone, the live generators doing this work today are different tools with different fingerprints, which is why the next section matters.
Google’s current video model is Veo 3.1, according to Google DeepMind’s own Veo page, which describes native audio including dialogue, scene extension, and vertical 9:16 output. Veo appears in Google Flow, the Gemini API, Google AI Studio, Vertex AI, Google Vids, and YouTube Shorts, where its “Ingredients to Video” feature turns photos into clips. Google says every Veo output carries an invisible SynthID watermark, and the Gemini app also attaches C2PA Content Credentials to its video output. Adobe Firefly, the third big name, automatically applies Content Credentials to assets generated with Firefly, and Adobe’s own help pages say the generative-AI disclosure credential cannot be deleted from the file. Knowing which generator made a clip tells you which provenance signal to look for, and the table later in this guide shows which of those signals survive being re-uploaded.
How do I check the label first?
Start here, because it takes ten seconds and it is the check most people skip. Platforms now label AI content, but each one does it differently, and the label is not always where you expect it.
- On YouTube, look directly below the player. Since May 2026, YouTube has used a single prominent label format: for long-form video it sits directly below the player, above the description; for Shorts it appears as an overlay on the video itself. Thumbnails work differently from videos: YouTube does not require a disclosure label on an AI-made thumbnail, because a thumbnail counts as production assistance, and our guide to how to tell if a YouTube thumbnail is AI-generated explains the rule and shows how to test a thumbnail with a reverse search instead. Creators disclose through YouTube Studio’s “AI use” setting, and disclosure is required for realistic AI content, a real person shown saying or doing something they did not, altered footage of a real event or place, or a realistic scene that never happened. YouTube also said on 27 May 2026 that it would apply labels automatically when its own systems detect significant photorealistic AI use, even if the creator said nothing. Labels are permanent for content made with YouTube’s own AI tools, such as Veo, and for files whose C2PA metadata says they are fully AI-generated. The viewer wording is “Made with AI” or “Altered or synthetic content”, shown in a “How this content was made” section.
- On TikTok, look for the grey pill. TikTok’s label reads “Creator labeled as AI-generated” when the poster applied it, or “AI-generated” when TikTok applied it automatically. TikTok’s Community Guidelines, effective 24 September 2026, require a label on AI-generated or significantly edited content showing realistic-looking scenes or people. TikTok also auto-labels content made with its own AI effects and uploads carrying C2PA Content Credentials, and those automatic labels cannot be removed. For the full wording on every platform, our guide to AI labels on social media lists each label, where it appears, and whether the creator can remove it.
- On Instagram and Facebook, look for “AI info”. Meta renamed its label from “Made with AI” to “AI info” in July 2024. Creators must switch on the “Add AI label” toggle for photorealistic video or realistic-sounding audio that was digitally created or altered. Meta also auto-labels from C2PA and IPTC signals in the file. Content that was only AI-modified, rather than AI-created, gets a note tucked in the post’s menu instead of a label on the post.
- On X, do not expect a label. X has no general AI-label duty for posts. Its rules ban deceptive synthetic media that could cause widespread confusion or harm, and X’s July 2026 Media Literacy Action Plan refers to “Made with AI” indicators for manipulated or AI media, but there is no post-level disclosure toggle to look for, so the absence of a label on X means nothing at all.
One warning before you trust a label: a missing label is not evidence a clip is real. The EU AI Act’s transparency duties have applied since 2 August 2026, requiring deepfakes to be disclosed as artificially generated, but enforcement and platform rollout are uneven, reposts strip labels, and a clip downloaded and re-uploaded to a different platform arrives with no label at all. Our worked test below found exactly that: OpenAI’s own Sora 2 demo, a clip that is 100% AI-generated, carries no AI label on YouTube. If the label is there, believe it. If it is not there, keep checking.
How do I check a video’s provenance?
Provenance is the record of where a file came from and whether it has been changed since. Two free tools let you check it, and both are worth knowing, but both have limits you should understand before you trust them.
The C2PA Verify tool is the one to try first, because it needs no account. Open verify.contentauthenticity.org, drag in the video file, and it inspects the Content Credentials embedded in the file, in your browser, without uploading anything. It accepts video in AVI, MOV and MP4 formats, among others. A valid credential tells you the file was signed by a named tool and has not been altered since signing. It does not tell you the scene is true, a generator can sign a completely invented scene, and the credential will be genuine. And a missing credential proves nothing either, because Content Credentials live in the file’s metadata, and re-encoding a video strips metadata. Every social platform re-encodes uploads, so a clip that has been through TikTok or WhatsApp will almost always arrive with its credentials gone. Our own guide to checking Content Credentials (C2PA) walks through the same tool for photos, and the video steps are identical.
Google’s SynthID Detector is the second provenance check, and it is the newest: Google opened it to the public on 7 October 2026, according to Google’s announcement. It detects SynthID, the invisible watermark Google embeds in content from its own models, including Veo. It accepts MP4, MOV and WEBM files and is free, but it needs a sign-in, it only detects SynthID (so it says nothing about Sora clips, which carry C2PA instead of SynthID), and a “not detected” result proves nothing, because the watermark may have been stripped or the clip may never have had one. The Gemini app offers a similar verification for signed-in users, one file up to 100 MB, video under 90 seconds, around ten video checks a day, according to Google’s Gemini help page, and it uses the same two signals, SynthID and Content Credentials.
The practical rule: provenance can confirm a clip, it can never clear one. A valid C2PA credential or a detected SynthID watermark is strong evidence the clip is AI-generated and tells you which tool made it. An empty result means the file has been through too many hands, or the generator never signed it, or the watermark is gone, and you must fall back to the frame search below.
How do I reverse-search a single frame?
A video is hard to search. A still frame is easy. This is the most powerful check in the guide, because it finds the clip’s origin even when every label and watermark has been stripped, and it is the same skill our step-by-step reverse image search guide teaches for photos.
- Pause on the most distinctive frame. Pick a moment with something unique in it, a face, a sign, a landmark, an unusual object, not a blank wall or a close-up of sky. Scrub through the clip and choose the frame that would be hardest to fake by accident.
- Screenshot it and crop tight. Capture the frame, then crop out the player controls, the progress bar, captions, and any app interface, so the search engine sees only the picture. Save it as a clean image file.
- Upload the frame to Google Lens. Open Lens, upload the screenshot, and read the “Exact matches” tab first, that is where reposts of the same clip appear, often with the original upload date. Then check the visual matches for the same scene on other sites.
- Cross-check in TinEye. TinEye’s index is smaller but it shows the date each copy was first indexed, which is often the best evidence of which copy came first. If Lens and TinEye disagree, trust the earlier indexed date, not the louder result.
- Search the clip’s title or a distinctive phrase from it. If the frame search finds a YouTube or TikTok page, open it and compare the upload date, the uploader, and the description with the copy in front of you. A clip reposted a week later with a scarier caption is the classic fake-viral-video pattern.
For a clip that came from YouTube, our guide to reverse YouTube video search covers the same frame method in more detail, including how to work from a video URL when you cannot download the file. Two honest limits: a frame search only works if the frame was indexed somewhere, and a brand-new clip, or one that has only ever lived inside a private chat, may return nothing at all. No result is not a pass, it is an absence of evidence, and the guide’s final section explains what to do with that.
Should I trust an AI video detector?
Last, and least. Detectors are the final step, not the first, because they are the least reliable check you have, and our worked test below caught one failing on a clip that is definitely AI-generated.
There are free video detectors a reader can try, and our roundup of the best deepfake video detector tools lists them with their limits, so this guide does not repeat that list. Two things to know before you use any of them. First, no consumer video detector publishes a verified accuracy rate, so a score is a probability, not a verdict, read “1% likely AI-generated” as “this tool thinks probably not”, never as “this clip is real”. Second, detectors need the actual video file, and most of them cannot take a YouTube or TikTok link: pasting a page URL usually fails, because the tool downloads the page as text instead of the video.
Use a detector as a second opinion only. If the label, the provenance and the frame search all point one way and the detector points the other, trust the first three, detectors are wrong often enough that a single score should never be the reason you accuse a creator of faking a clip, and never the reason you share one as real.
What should I look for when I watch the clip?
With the checks done, watch the clip itself, slowly, and more than once. Modern generators are good, but they still fail in the same places, and the failures are easier to spot if you know what to look for.
- Physics that does not behave. Water that flows uphill, shadows that point in different directions in the same shot, objects that pass through each other, a ball that bounces wrong. Veo 3.1 is marketed on realistic physics, according to Google DeepMind, which is exactly why physics errors are still worth checking, when they appear, they are hard to explain away.
- Hands, teeth and text. Fingers that merge or change number between frames, teeth that blur into one shape, and any writing in the scene, shop signs, number plates, book titles, that dissolves into letter-like shapes instead of real words. Text is still the most reliable visual tell in 2026, because generators invent lettering freely.
- Faces and hair at the edges. Ears that change shape, hair that merges with the background, glasses that warp where they meet the face. Watch the edges of the frame, not the centre, generators spend their effort on the middle.
- Audio that does not match. Lips that move slightly out of sync, ambient sound that stays the same while the scene changes, a voice that sounds flat or oddly paced. Native audio is new in tools like Veo 3.1, and it is often the weakest part of a generated clip.
- The cut. Most generated clips are a single unbroken shot of five to ten seconds, because every cut is a chance for something to change, a shirt, a background detail, a person’s position. If a “viral video” is one short, perfect, unbroken take of an impossible event, that is a property of how it was made, not of reality.
None of these is proof on its own. Real footage has blurry frames, bad audio and strange shadows too. What matters is the pattern: two or three of these tells together, plus a missing label and no provenance, is a strong case. One odd frame in an otherwise well-documented clip is usually just a compression artefact.
Which checks survive a re-upload?
This is the ranking asset for this guide, and it is the table to bookmark: every check above behaves differently once a clip is downloaded, screen-recorded, cropped or re-encoded, which is what happens to every viral video. The table shows what survives each treatment, so you know which check to trust on the copy in front of you.
| Check | Original file | Re-uploaded | Screen-recorded | Cropped | Re-encoded |
|---|---|---|---|---|---|
| Visible watermark | Present if the generator added one | Usually survives | Survives, it is in the pixels | Gone if cropped out | Usually survives |
| C2PA Content Credentials | Present if the generator signed it | Stripped, platforms re-encode | Stripped | Stripped, new file | Stripped |
| SynthID invisible watermark | Present on Google-made clips | Often survives re-encoding | Often survives | Survives if enough of the frame remains | Designed to survive |
| Platform AI label | Shown if the platform applied it | New platform may apply its own, or not | Lost, the label lives on the page, not in the video | Lost | Lost |
| Keyframe reverse search | Works if the frame is indexed | Usually works | Usually works | Often works on a tight crop | Usually works |
| Detector score | Most reliable here | Degrades | Degrades further | Degrades further | Degrades |
Two conclusions fall out of that table. First, the keyframe reverse search is the most durable check you have: it is the only one that keeps working after a re-upload, a screen recording, a crop and a re-encode, because it depends on the picture itself rather than on metadata or a label. Second, provenance is the most fragile: Content Credentials vanish the first time a platform touches the file, which is why a missing credential must never be read as “real”. Check the label and the provenance on the original if you can find it; check the frame on whatever copy you actually have.
Worked test: is this video AI generated? OpenAI’s Sora 2 clip, checked end to end
On 11 November 2026, between 20:16 and 20:23 UTC, we ran the full workflow on a clip whose origin is not in doubt: “This is Sora 2”, the official Sora 2 launch video on OpenAI’s verified YouTube channel, uploaded 30 September 2025, 57 seconds long, with 850,703 views and 14,265 likes at the time of the check. Every step below was performed live, with no sign-ins and no accounts. If the checks work, they should all point the same way on this clip. They did not, and that is the point of the test.
| Check | What we did | What we found |
|---|---|---|
| Platform label | Opened the YouTube page and read the full expanded description | No AI label anywhere. No “Made with AI”, no “Altered or synthetic content”, no “How this content was made” section, on a clip that is 100% AI-generated, posted by the company that made the generator |
| Keyframe reverse search (Lens) | Searched the clip’s public thumbnail frame, an anime-style close-up of a girl’s eyes with fireworks reflected in them | Lens identified it immediately: its summary called it a scene from OpenAI’s promotional demonstration for Sora 2. The top visual match was the YouTube video itself, and the exact matches included news coverage from AUTOMATON, Mezha, VICE and TechSpot, all dated 30 September 2025 |
| Keyframe reverse search (TinEye) | Same frame, TinEye, no sign-in | Only 4 results across 86.7 billion images searched, first indexed 2 December 2025, two months after the clip was uploaded. YouTube itself did not appear in the results. TinEye’s index lags behind Lens on video frames |
| Detector | Pasted the frame to Deploy Detect (the free Hive tool, no sign-in) | False negative. “1% likely to be AI-generated image”, “0% likely to be deepfake”, and “No Content Credentials found for this media”, on a frame from a clip that is certainly AI-generated. Pasting the YouTube URL itself failed first, with “Media type not supported: Text” |
| Detector (second) | Opened Deepware Scanner | Could not be tested: the site failed to load twice with a connection reset on 11 November 2026 |
| Provenance tools | SynthID Detector and Gemini app verification | Not attempted, both require a sign-in, and this test uses no accounts. Their sign-in requirement is itself a limit worth knowing |
One limitation to state plainly: the Lens and TinEye searches used the video’s public thumbnail frame, not a screenshot of the paused video, because the test browser had no local file to upload. The frame is from the same clip, and the result stands, but a frame you capture yourself may return different matches, and you should always use your own screenshot where you can.
What the test proves: the keyframe search worked exactly as advertised and found the clip’s origin in seconds. The label check failed completely on a clip that should be the easiest case in the world to label. The detector failed in the other direction, confidently clearing a clip it should have caught. And the provenance tools that might have settled the question could not be used without an account. That is the post-watermark era in miniature: the checks that depend on platforms and detectors are unreliable, and the check that depends on the picture itself is the one that held up.
What can no check prove?
Three honest limits, and they apply to every clip you will ever check.
First, no match does not mean real. A clip that returns nothing in Lens or TinEye, carries no label, and shows no credential may be a genuine video, or it may be a brand-new generation that no search engine has indexed yet, a clip that has only ever circulated inside private chats, or a generation from a tool that signs nothing. Absence of evidence is not evidence of reality.
Second, a pass does not mean true. A valid Content Credential proves the file was signed by a named tool and has not changed since, it does not prove the scene happened. A detector score of 1% is a guess about pixels, not a fact about the world. A label that says “Made with AI” tells you how the clip was produced, not whether the event it depicts is real. And a real video can be reposted with a false caption, which no pixel check will ever catch, for that, you need the frame search and the source, not the video itself.
Third, you cannot prove a negative from a screenshot. If every check comes back empty and the clip still looks wrong, the right answer is “I cannot verify this”, not “it is fake” and not “it is real”. Do not share it as either. Send it to a fact-checker, or wait for the original source to surface. The most useful sentence in this whole guide is the one you say before you share: “I checked, and I could not verify it.”
FAQs
Can I tell if a video is AI-generated just by watching it?
Sometimes, but not reliably. Obvious tells are garbled text, merging fingers, inconsistent shadows and audio that does not match the lips. Modern generators hide most of these, so watching alone is the weakest check. Pair it with the label, the provenance check and a keyframe reverse search before you decide.
Is there a free tool that detects AI videos?
Yes, with limits. The C2PA Verify tool is free and needs no account, and Google’s SynthID Detector is free but needs a sign-in. Deploy Detect is free with no sign-in. None of them publishes a verified accuracy rate, and our own test caught one giving a false negative, so treat every score as a probability, never as proof.
Does a missing AI label mean the video is real?
No. Labels depend on the creator disclosing, the platform detecting, and the copy you are watching still carrying the label, and re-uploads strip all three. OpenAI’s own Sora 2 demo carries no AI label on YouTube. A present label is strong evidence; an absent label is no evidence at all.
What should I do before sharing a viral video clip?
Run the three quick checks: look for the platform’s AI label, reverse-search a keyframe to find the original upload, and check the file’s provenance if you have the original. If you cannot verify the clip, do not share it as real. Saying “I checked and could not verify this” is always better than spreading a fake.







