To check if a video is AI generated, work through four kinds of evidence in order: the platform's AI label, the file's Content Credentials, an invisible watermark check such as Google's SynthID Detector, and finally a classifier score plus your own source checks. A positive label, credential or watermark is strong evidence that AI was used. A negative result from any of them proves almost nothing, because most generators add no signal and every platform reencodes uploads.
This matters more this week than it did a month ago. On October 29, 2026, Google updated its SynthID Detector announcement to list Apple alongside OpenAI, NVIDIA and Kakao as partners whose content the free portal can check (Google). Apple's iOS 27.2, released October 26, added SynthID watermarking for images made or edited with its generative tools (Apple), and on the same day as Google's update, the standards body ETSI published TS 104 275, a specification for stopping AI systems from producing harmful deepfakes in the first place (ETSI). None of that gives you a single button that says "real" or "fake". This guide shows what each check can and cannot tell you about a video.
How to check if a video is AI generated, step by step
- Read the label on the platform. YouTube, TikTok, Instagram and Facebook all show AI labels, and some are applied automatically. Details are in the next section.
- Find the earliest copy you can. Look for the first upload, the original account and the original file. Every repost, screen recording and download from a social app can strip metadata and weaken a watermark.
- Inspect the Content Credentials. Drop the file into Content Credentials Verify. It accepts MP4, MOV and AVI along with image and audio formats (Content Credentials). A valid credential shows which app or camera signed the file and whether AI tools were used.
- Run a watermark check. Upload the file to the SynthID Detector at synthid.com. For video, it reports the segments where it found a watermark, which helps when only part of a clip was generated (SynthID Detector).
- Get a classifier opinion. A detector such as Hive estimates the chance that a clip is synthetic even when it carries no label. Treat the score as a lead, not a verdict.
- Check the source and the context. Reverse search a few keyframes, find who posted it first, and look for coverage from people who were there. For a clip of someone asking for money or access, call them back on a number you already know.
The order is deliberate. Steps 1 to 4 look for signals someone put there on purpose, which are reliable when present. Step 5 guesses from patterns, which is far less reliable on video, as the research section shows. Step 6 is the only one that works on everything.
Platform AI labels on YouTube, TikTok and Meta
Platform labels are the fastest signal, and the main platforms now apply some of them automatically.
YouTube. Creators must disclose realistic AI content, such as making a real person appear to say something they did not say, altering footage of a real event, or generating a realistic scene that never happened (YouTube Help). Since a May 27, 2026 update, the label for photorealistic AI content sits directly below the player on long videos and as an overlay on Shorts, and YouTube applies a label automatically when its systems detect significant photorealistic AI use that the creator did not disclose (YouTube Blog). Labels for content made with YouTube's own AI tools, such as Veo or Dream Screen, and for files whose C2PA metadata says they are fully generative cannot be removed by the creator. Labels for animated or lightly edited content appear in the expanded description instead, so open it before you decide a video is unlabeled.
TikTok. TikTok shows a "creator labeled as AI generated" label when the poster discloses it, and an automatic "AI generated" label for content made with its AI effects or uploaded with Content Credentials attached. Once an automatic label is applied, the creator cannot remove it (TikTok Support).
Facebook and Instagram. Meta displays an "AI info" label for content it detects was generated by an AI tool, says whether the label comes from industry signals or self disclosure, and requires people to disclose organic posts with photorealistic video or realistic audio that was digitally created or altered (Meta).
The catch is the same on every platform. A label is good evidence that AI was involved. No label means only that nobody disclosed it and no signal survived the upload.
Which checker can read which video
Watermarks and credentials only exist if the tool that made the video added them, and each checker reads only some of them. This table shows the current picture for the main sources of AI video, from official documentation checked October 31, 2026.
| Where the video came from | Signal it carries | Best place to check |
|---|---|---|
| Google AI tools (Veo, Gemini, YouTube tools) | SynthID watermark | SynthID Detector or the Gemini app |
| OpenAI's Sora (now discontinued) | Content Credentials only | Content Credentials Verify |
| NVIDIA and Kakao tools | SynthID watermark | SynthID Detector |
| Apple's generative tools | SynthID, for images only so far | SynthID Detector (images) |
| Any other generator | Often nothing | Classifier plus source checks |
Two rows deserve a note. OpenAI's help page lists Content Credentials as the only provenance signal for video and says Sora is discontinued, while images and audio from ChatGPT and its API carry SynthID (OpenAI Help). OpenAI's own verification tool accepts only image and audio formats, so it cannot check a video file (OpenAI). For an old Sora clip, Content Credentials are your only built in signal, and they vanish if the file was reencoded.
Apple's iOS 27.2 notes describe SynthID support for images generated or modified in Photos and Image Playground. They say nothing about video. Also note that the SynthID Detector FAQ still names only NVIDIA, OpenAI and Kakao as adopters, while Google's newer announcement adds Apple; we go with the newer source.
Using SynthID Detector and the Gemini app on video
The SynthID Detector portal covers Google and its partners at once, so it is the broadest watermark check for video. Upload the highest quality file you have: Google says heavy compression and editing reduce detection, and that the detector can, in very rare cases, trigger on content with no watermark. A "not detected" result can mean the clip came from a model without SynthID, was made before a partner adopted it, or was edited enough to degrade the signal.
The Gemini app is the other option. Signed in at gemini.google.com, you can upload one file and ask "Is this AI generated?". Google's help page sets clear limits for video: files of 100 MB or less, clips shorter than 90 seconds, and a quota of about 10 video checks (up to 5 minutes total) in any rolling 24 hour window. Gemini recognizes SynthID only in content from Google's own AI tools, though it also reads Content Credentials, and for video it states which parts carried a watermark (Gemini Help). For a long video, trim the section you care about before you upload it.
Our SynthID Detector guide covers the portal in more depth, including how the watermark is embedded and how it can be degraded.
How accurate are AI video detectors?
Classifiers are the only tools that work on a video with no label at all, so their accuracy matters. The best independent evidence is Deepfake-Eval-2024, a benchmark of deepfakes actually circulating online in 2024, flagged by journalists and social media users. Its video set is 45.1 hours across 2,036 clips (paper).
The results are sobering:
- Three open source video detectors scored an AUC of 0.50 to 0.63 on the real world clips, against 0.87 to 0.96 on the academic datasets they were built for. An AUC of 0.5 is a coin flip.
- The best of the six commercial video models tested reached 0.78 accuracy and 0.79 AUC. Vendors in the study included Hive, Reality Defender, Sensity AI and others, but results were anonymized, so you cannot tell which product scored what.
- Models struggled most with partial edits. Accuracy fell 31 percent on videos where only some faces were manipulated, and 17.4 percent on videos where objects, places or bodies were altered rather than faces.
- Off the shelf detectors did 21.3 percent worse on videos that appeared to be diffusion generated, the kind tools like Sora produce.
The researchers' own human labelers disagreed on 6.6 percent of a checked sample of videos, which shows how hard the task is even for trained analysts. For more on how detection benchmarks are built and gamed, see our guide to the RAID benchmark for AI text detectors, and our explainer on whether AI detectors are accurate.
What this means in practice: use Hive's free Chrome extension, which Hive says checks video as you browse with no fees or login (Hive), or an enterprise tool such as Reality Defender as one input. A high score is a reason to dig further. A low score clears nothing.
Visual and audio clues worth checking
Your eyes are the weakest tool on this list, but some clues still help. Google's Gemini help page suggests looking for garbled text, errors in hands, teeth or background details, inconsistent shadows and lighting, and patterns that repeat unnaturally. The Deepfake-Eval-2024 team leaned on one cue above all for talking head clips: in real footage, lips and mouth movement are crisp and match the sound exactly, while AI lip sync tends to be roughly in time but soft or unnatural.
A few more checks that cost nothing:
- Pause on frames with text. Signs, labels and screens are still hard for many generators to keep consistent across frames.
- Watch the edges. Hair, glasses, hands crossing a face and objects passing behind each other are where blending fails.
- Listen separately. Background music can hide synthetic speech. The study found audio detectors missed more fakes when music was added.
- Check the physics. Water, smoke, crowds and reflections that behave oddly are a reason to keep looking.
Treat each of these as a reason to investigate, never as proof. Current generators fix many classic tells, and real footage can look strange after heavy compression.
Why a missing watermark proves nothing
Almost every check above has the same weakness: it can confirm AI use, but it cannot confirm that a video is real. Content Credentials disappear when a file is screenshotted, converted or reencoded. Watermarks survive more, but only exist if the generator added one, and Google itself says someone who tries many transformations may find one that defeats detection. Videos from generators outside these programs, including open weight models people run on their own machines, may carry no signal at all.
ETSI's new specification is aimed at that gap from the other end. TS 104 275 asks AI developers and operators for risk assessments before deployment, human oversight, content filtering, AI content labeling, ongoing monitoring, and test scenarios with pass and fail criteria. It is a set of requirements for builders, not a detector for viewers, so it will not change what you can check today. If you build or test generation systems, our guide to LLM red teaming tools covers how to probe a model for this kind of misuse.
For the wider picture across images, video and audio, read our deepfake detection guide, which compares Hive, SynthID and C2PA in more detail. If the question is about text rather than video, see does ChatGPT have a watermark.
FAQ
Is there a free AI video detector?
Yes, several, each with limits. SynthID Detector and the Gemini app check for Google and partner watermarks for free after you sign in. Content Credentials Verify reads provenance data from MP4, MOV and AVI files for free. Hive offers a free Chrome extension that gives a classifier score. None of them can prove a video is real.
How do I check if a YouTube video is AI generated?
Look directly below the player on long videos, or for the overlay on Shorts, and open the expanded description, where labels for animated or lightly edited AI content appear. YouTube adds some labels automatically. If there is no label, fall back on source checks: who uploaded it first, and whether trusted outlets confirm the event.
Can SynthID detect Sora videos?
Do not count on it. OpenAI's help page lists Content Credentials, not SynthID, as the provenance signal for video, and says Sora is discontinued. Try Content Credentials Verify on the original file instead. If the clip was reposted, the credential is probably gone.
Can I check a video by pasting a link?
Not with SynthID Detector or Content Credentials Verify, which both work on uploaded files. You need the file itself, ideally the original. Browser tools such as Hive's extension can scan videos on the page you are viewing.
Does "no watermark detected" mean a video is real?
No. It only means the checker found no signal it recognizes. The video could come from a generator that adds no watermark, or the signal could have been stripped or degraded by reencoding and editing. Combine the result with source and context checks.
