AI Detection

Sign in or create an account

How to Tell if a Video Is AI-Generated

Updated

You usually cannot tell by looking. Current research finds that people judge synthetic media at close to chance, and the visual flaws that gave older fakes away are fading as generators improve. What works better is checking where the video came from: its source, its context, any provenance data attached to it, and whether independent reporting backs it up. Visual and audio signals still help, but as reasons to look harder, not as proof.

This site checks text, not video, so this guide stays informational. It covers what each kind of evidence can and cannot show, and a workflow for weighing them.

Key takeaways

  • No single visual or audio flaw proves a video is AI-generated, and a clean-looking video proves nothing either.
  • Start with source, context and provenance. Look at pixels last.
  • Content Credentials and watermarks are strong evidence when present, but their absence is not evidence of anything.
  • Detection tools return a probability from a model, tested under narrow conditions. Treat the output as one input.
  • Many “tells” from the face-swap era, such as blinking or odd hands, are unreliable on current generators.

Can you tell by looking?

Mostly, no. A study published in Communications of the ACM tested 1,276 participants on images, video, audio and audiovisual clips and found mean performance close to the 50% chance level. Prior knowledge of synthetic media helped little, and the authors concluded that relying on human perception alone is no longer viable.

Researchers who study the problem are careful about this. MIT’s Detect Fakes project lists points to watch, including skin texture, shadows around the eyes, glare on glasses, facial hair, blinking and lip movements, while stating that there is no single tell-tale sign. Those points were built mainly around face-manipulation deepfakes. Fully generated video is a different problem, and generators change quickly. Berkeley’s Hany Farid describes current models as passing through the uncanny valley.

So treat any list of visual tells, including the one below, as a source of questions. A flaw is a reason to investigate. Its absence is not a reason to trust.

What visual and audio signals are worth checking?

Check these in roughly this order, and note that each can be produced by ordinary cameras, compression and editing as well.

Frame inconsistencies. Step through the clip slowly. Look for objects that change shape, count or position between frames, text on signs or clothing that shifts or turns to gibberish, and background details that appear or vanish when the camera moves. These were once common, and some current tools still produce them, especially in busy scenes or longer clips.

Facial motion. Watch the jaw, teeth, tongue and the skin around the mouth while the person speaks, and see whether the face holds its structure when the head turns sharply or something passes in front of it. Blink rate was a popular cue for early face-swap fakes. Do not rely on it now; real people blink irregularly and current generators are not tied to the old patterns.

Lip sync. Mismatch between mouth shapes and speech is a classic sign of a dubbed or voice-cloned clip, and MIT lists lip movement as a point to watch. But modern generators can produce speech and mouth movement together. Google introduced Veo 3 in 2025 as the first major video model with native audio, so good sync no longer clears a clip.

Lighting and physics. Do shadows agree with the visible light sources? Do reflections in glass, water or glasses move correctly? Do liquids, cloth and falling objects behave plausibly across the whole shot? Generators have improved here, so a physics slip is a lead, not a verdict.

Temporal consistency. Look at the same person or object across cuts and across the length of the clip. Clothing details, jewellery, scars, a logo or a number can drift. Many AI clips are short single shots, so the absence of any cut, or a clip assembled from a series of very short shots, is worth noting as context only.

Audio. Listen for speech with unusual rhythm or flat emotion, room sound that does not match the space, ambient noise that loops or stops abruptly, and voices that stay identical in acoustics as the camera moves. Audio is often a separate layer and may be a clone of a real voice laid over real or generated footage. Because voices can now be cloned from short samples, a convincing voice is weak evidence either way.

Old tells that no longer work as rules. Malformed hands, six fingers, melting backgrounds and very uniform “waxy” skin were all common in earlier tools. They can still appear, but newer generators often avoid them, and real footage can show odd hands through motion blur and compression. Do not teach yourself that a clip is real because the hands look fine.

Which checks beat looking closely?

Source. Who posted it first? Does the account have a history, a real name, a location and earlier footage that matches? An anonymous account that posts a dramatic clip with no original context is a weak source whether or not the video is synthetic. Look for the earliest upload, not the most-shared copy.

Context. A real event leaves traces: other angles, local reporting, official statements, witnesses, matching weather and location details. A shocking clip that no one else filmed, from a place full of phones, deserves suspicion. Search for the event by description and for the clip’s claim in news and fact-checking sites.

Reverse search on key frames. Pull several frames and run them through image search to find earlier or original versions, or evidence the footage has been recycled from another event. Journalists use the InVID-WeVerify browser extension, which breaks videos into keyframes and sends them to several reverse-search engines. Reverse search finds reuse and misattribution. It will not necessarily find a brand-new generated clip, because there may be no earlier version to match.

Creator disclosure and labels. Ask the poster, and check the description and any platform label. A platform label may exist, but a missing one means little: in a Washington Post test of eight social apps in October 2025, only YouTube added any indication that an uploaded AI video was not real, and it appeared inside the description and read “Altered or synthetic content”. Platform behaviour may have changed since then.

What do content credentials and watermarks tell you?

Provenance data is the strongest evidence when it exists, and weak when it does not.

Content Credentials (C2PA). These are cryptographically signed records of a file’s origin and edit history. The Content Authenticity Initiative says they can be removed during editing or sharing and that they are not intended to indicate whether content is “real”. The C2PA’s own FAQ likewise says manifests can be separated from the asset. So a valid credential that names an AI tool is meaningful, while no credential is meaningless. In the Washington Post test above, none of the eight platforms kept the credentials on the uploaded video or let users see them.

Visible watermarks. Some tools stamp a visible mark on exported clips. OpenAI’s Sora 2 system card described a visible moving watermark plus C2PA metadata and internal detection tools, and stated that there is no single solution to provenance. Farid said removal of the watermark after launch was “predictable”. OpenAI announced in March 2026 that it was shutting down Sora, but clips made earlier still circulate. A visible mark suggests the origin; a cropped or removed one tells you nothing.

Invisible watermarks. Google’s SynthID embeds an imperceptible watermark in output from its models. The Gemini app lets you upload a video and ask whether Google AI created or edited it, with limits of 100 MB, under 90 seconds and a daily quota. Google states the check only recognizes content from Google AI tools: if nothing is found, the video could still be AI-generated by another system. A positive result is informative. A negative one is not.

What can independent detection tools tell you?

Some free and commercial services analyze a file with trained classifiers and return a likelihood. One example is the University at Buffalo’s DeepFake-O-Meter, which runs several research detection methods on an uploaded file. Commercial vendors sell similar services, with their own claims.

The honest limit is generalization. The Deepfake-Eval-2024 benchmark collected deepfakes that circulated in 2024 and found that open-source detectors’ AUC fell by about 50% for video compared with earlier academic benchmarks. Commercial and fine-tuned models did better but did not reach the accuracy of human forensic analysts. In practice that means a score from one tool, on one file, tested against generators it may not have seen, is not decisive. Compression, re-recording a screen and cropping also change what a detector sees. If two tools disagree, take that as uncertainty. For the mechanics, see how AI video detection works.

The same discipline we apply to text applies here. A detection result is a statistical judgment about resemblance to training data, not a finding about authorship. Our guide on whether AI detectors can be wrong explains that idea for text, and it carries over.

What is a practical verification workflow?

Work from the cheapest, strongest checks to the weakest.

  1. Decide what you need to know. “Is this clip from a real event?” and “did a generator make these pixels?” are different questions, and the first is often more useful and easier to answer.
  2. Find the earliest source. Trace the clip to the first poster and read their history. Save the URL and download the best-quality copy.
  3. Check for provenance. Look for a visible watermark, a platform label, and any Content Credentials via a verifier. Run the file through SynthID verification if a Google origin is plausible. Record the result either way.
  4. Search for context. Look for other footage, news coverage and fact-checks of the same event, and reverse-search several key frames.
  5. Inspect the media. Step through frames, then listen with headphones. Note anomalies, then ask if ordinary compression or editing could explain each.
  6. Run one or two independent detectors. Record the tool, date and score, and treat it as a lead.
  7. Weigh and state the confidence. Say “I could not confirm this clip’s origin” or “provenance data shows an AI tool made it”. Avoid declaring a fake or a real one on pixel inspection alone.

For the difference between a manipulated real recording and a fully generated clip, see deepfake vs AI-generated video. The same provenance-first approach applies to stills; see how to tell if an image is AI-generated.

Signal table: what each signal can and cannot show

Signal What it can show What it cannot show
Frame glitches (shifting objects, garbled text) A generator or heavy edit may be involved Real footage is fine if compression or motion blur caused them; clean frames prove nothing
Facial motion and lip sync A face swap or dub may have altered the mouth Modern models sync speech and lips; blink rate is unreliable
Lighting and physics A scene may be composited or generated Cameras, lenses and filters also break these rules
Temporal consistency Details drifting across a clip can point to generation Short, single-shot clips offer little to compare
Audio A cloned or pasted voice, or mismatched room sound Natural-sounding generated audio is now common
Source and account history Whether the clip has a credible origin Credible accounts can share fakes by mistake
Context and reverse search Reused or misattributed footage; whether the event happened A new generated clip may have no earlier version
Content Credentials A signed record of origin and edits Absence proves nothing; they are easily stripped on upload
Visible watermark The tool the clip came from Can be cropped out or removed
SynthID check Google AI generated or edited part of it Anything not made with Google tools
Independent detector score A model-based likelihood, a lead Certainty; scores shift with compression, new generators and file quality

What should you do with an uncertain result?

Say what you know and what you do not. If the clip matters, because it could affect an election, a legal matter, a financial decision or someone’s reputation, do not share it as authentic or fake on a hunch. Ask a professional newsroom verification team or a recognized fact-checker. Our methodology page describes how we handle uncertainty on text, which is a useful model: report a finding with its limits.

If you also have written material tied to the video, such as a transcript, caption or script, and want to review it, run it through our AI-generated text detector and review the flagged passages rather than relying on the overall score.

Sources and check dates

All checked 5 October 2026. Vendor and platform claims are the vendor’s own.

AI text detector Check any text for AI-writing patterns. Methodology What we analyse, what a result means and its limits. Accuracy What we can and cannot claim about accuracy.