AI Detection

Is This Video AI? How to Spot AI Videos and Deepfakes

A step-by-step way to judge a suspicious clip: where it came from, what AI labels mean, the signs to watch for, and what to do about deepfake calls.

Updated 11 min read

On this page
  1. Is This Video AI? Start With Where It Came From
  2. Does the Video Have an AI Watermark or Label?
  3. How to Tell If a Video Is AI-Generated by Watching It
  4. How to Spot a Deepfake Video of a Real Person
  5. What to Do About a Deepfake Video Call or Cloned Voice
  6. Can an AI Video Detector Tell You for Sure?
Is This Video AI? How to Spot AI Videos and Deepfakes

Is this video AI? You can usually get close to an answer in a few minutes: find where the clip first appeared, look for a watermark or a Content Credentials label, then watch it slowly for movement and objects that don't behave. No single sign proves anything, so the goal is to collect clues, not to find one perfect tell.

This guide walks through those checks, starting with the quickest. It also covers two kinds of fake that can cost people money: deepfake videos of real people, and video calls or voice messages from someone pretending to be a boss or a relative.

Key Takeaways

  • Start by finding the original upload, because the first account to post a clip can tell you more than the clip itself.
  • A visible watermark is a strong clue, but watermarks can be cropped away, so a missing one proves nothing.
  • AI video clips are usually only a few seconds long, and objects often change shape when the camera moves.
  • In a 2021 study, people could not reliably spot deepfakes but believed they could.
  • For an urgent call or message asking for money, hang up and call back on a number you already know.

Is This Video AI? Start With Where It Came From

The fastest check isn't about pixels at all: find the original. A repost on a big news-style page usually tells you little. The account that uploaded the clip first can tell you a lot.

So scroll back to the earliest version you can find. Look at that account's other posts. Does it post the same style of dramatic clip every day? Does it say anywhere that it makes AI videos? Is the caption claiming something that no news outlet has reported?

That last question matters. If a video shows something huge, such as a public figure in an accident or a riot in a city center, check whether any news outlet has reported it. If none has, be extra careful before you share it.

Does the Video Have an AI Watermark or Label?

Some AI video makers stamp their work. In October 2025, Lead Stories checked a viral clip of a uniformed man calling himself "Sgt. Pepper" pepper-spraying a shouting protester and found that it was made with an AI video app. "This video has a watermark from Sora, an app from Open AI that generates videos from text prompts," the fact-checkers wrote, adding that the watermark was "changing locations several times" during the clip.

OpenAI has since said it would stop offering Sora: it told Engadget in March 2026 that it had "decided to discontinue Sora in the consumer app and API." Older clips can resurface, though, so it is still worth knowing what a Sora logo moving around the frame looks like.

Built-In Watermarks and Content Credentials

Other makers mark their work in other ways. Google DeepMind says videos made with its Veo model "will be marked with SynthID, our advanced technology for watermarking."

A third kind of label travels with the file itself. The Coalition for Content Provenance and Authenticity (C2PA), which is behind an open standard for this, describes Content Credentials as working "like a nutrition label for digital content, giving a peek at the content's history." When a video carries them, they can show where it came from and how it was edited.

The catch is what happens when there is no label. Full Fact, the UK fact-checking charity, warns in its guide to spotting AI content that visible watermarks "can easily be cropped away" before something is shared, and that if you pay to use an AI tool, content can often be made without a watermark. So treat a label as evidence, and its absence as nothing at all.

Three kinds of AI label on a video, and what each one tells you A visible watermark, like Sora's, is easy to spot but can be cropped away. Content Credentials work like a nutrition label showing the file's history. A built-in mark: Google says Veo videos are marked with SynthID, its watermarking technology. A missing label does not prove a video is real, because paid AI tools can often make content without a watermark. What an AI label tells you Visible watermark A logo on the picture, like Sora's. Easy to spot, but it can be cropped away. Content Credentials Like a nutrition label in the file: where it came from and how it was edited. Built-in mark Google says its Veo videos are marked with SynthID, its watermarking technology. No label? That proves nothing. Paid tools can often skip it.
The three kinds of AI label you may find on a video. Sources: Lead Stories (Sora watermark), C2PA (Content Credentials), Google DeepMind (SynthID) and Full Fact (cropped watermarks, paid tools that skip them).

How to Tell If a Video Is AI-Generated by Watching It

If the source and the labels don't settle it, watch the clip itself. Start with its length. Gabriele de Seta, a researcher at the University of Bergen, told Science Norway that "most AI videos are only a few seconds long," and that longer ones "are often stitched together from several short clips." Google's own Veo page lists its benchmark videos as 8 seconds long.

Then look for these signs:

  • Movement that feels off. De Seta says most AI-generated humans and animals "still move somewhat unnaturally." Watch walking, chewing and hands picking things up.
  • Objects that change. He also notes that "objects can change shape or position when the camera moves." Keep an eye on signs, cars, furniture and people in the background as the view pans.
  • Body details. Full Fact says fingers, arms, ears, toes and teeth are often where errors appear in images and videos of people.
  • Joins between short clips. Because longer AI videos are often stitched together from short clips, notice sudden jumps to a new angle every few seconds. Real edited footage cuts too, so treat this as a weak clue.

Use these as clues, not a test you can pass. De Seta himself warns against relying too much on them, "as the technology is improving fast." His other advice is practical: trying the tools yourself makes you better at recognizing what they produce. If you want to train your eye on stills first, our post on whether you can tell which images are AI has a short quiz.

How to Spot a Deepfake Video of a Real Person

A deepfake usually means a real person's face or voice put onto footage they never made: a face swap, or a mouth moved to match words they never said.

Full Fact's older but still useful guide on how to spot deepfake videos and AI audio suggests watching "on a screen bigger than your phone's if possible." It points out that deepfakes "are most convincing when the person is looking at the camera straight on," so turns of the head are where the face may slip. It also suggests you "pause the video a few times and check whether the eyes are looking in the same direction."

Then step back from the face. Full Fact's advice is to "think about what the person in the clip is actually saying, and how they say it." Would this person really say this, in this way, in this place? Has any outlet that covers them reported it?

Why Your Gut Isn't Enough

Don't trust your gut too much here. In a 2021 experiment with 210 people, published in the science journal iScience as "Fooled twice: People cannot detect deepfakes but think they can", researchers found that "people cannot reliably detect deepfakes." Neither warnings nor cash rewards made people better at it. Participants also leaned toward calling deepfakes real, and they overestimated their own skill.

Four steps to check a suspicious video Step 1: find the original upload and the account that posted it. Step 2: look for a watermark or Content Credentials. Step 3: watch slowly on a big screen, pausing to check movement, objects and eyes. Step 4: run a check and treat the result as one clue among the others. Check a video, step by step 1 Find the original Who posted it first? Did any outlet report it? 2 Look for labels Watermark, or Content Credentials in the file. 3 Watch slowly Big screen, pause often: movement, objects, eyes. 4 Run a check Treat the score as one clue, never as proof.
A four-step routine for a suspicious clip, based on the advice of Full Fact and the researchers quoted in this guide.

What to Do About a Deepfake Video Call or Cloned Voice

Some of the costliest fakes aren't viral videos at all. In February 2024, The Standard reported on Hong Kong's first deepfake video conference scam. Citing police, it said "the speech of the 'CFO' was only a virtual video generated by the scammer through deepfake," and that the victim made 15 transfers totaling HK$200 million.

Voices can be faked the same way. The US Federal Trade Commission warns that scammers use AI voice cloning in family emergency schemes, where a "relative" calls in distress asking for money. Its advice is short: "Don't trust the voice." Call the person back and "use a phone number you know is theirs."

The FBI goes one step further in a December 2024 public warning about AI fraud. It suggests you "create a secret word or phrase with your family to verify their identity," and to verify a caller "by hanging up the phone" and calling back. It also mentions "subtle imperfections in images and videos" as a sign of fakes, though a live call gives you little time to look.

The lesson for calls is different from viral clips. You don't need to decide whether the face is real. You only need a second channel you control: a known number, a code word, or a colleague you can ask in person.

Can an AI Video Detector Tell You for Sure?

No tool can be relied on alone. A detector gives you a probability, and you should read it alongside everything above. Even professional fact-checkers are cautious: Full Fact says of online AI detection tools that "so far we have not found these tools to be reliable."

What a good check can add is detail you can't get by eye. The IsItAI.com AI video detector scores one frame every 2 seconds, so you see where in a clip AI appears instead of one yes or no.

It reads signed Content Credentials first when a file has them, offers an optional face swap check, and gives a hint when the soundtrack may be AI-generated music. Its verdicts read "Likely AI-generated," "Possibly AI-generated" or "Likely human-made," never "proven."

For a wider look at text, images and audio too, see our guide on how to detect AI-generated content. And if the clip is really a still photo that looks retouched, our post on how to tell if a photo was edited with AI covers the checks for stills.

Check a photo yourself

Upload it to the detector. Investigate online shows where it has appeared before.

Check an image

Frequently Asked Questions

How can you tell if a video is AI-generated?

Find the original upload and check the account that posted it. Look for a watermark or Content Credentials. Then watch slowly: AI clips are usually a few seconds long, people and animals often move unnaturally, and objects can change shape when the camera moves. No single sign is proof, so combine several clues before deciding.

Does a video without a watermark mean it is real?

No. Visible watermarks can be cropped away before a clip is shared, and Full Fact notes that if you pay to use an AI tool, content can often be made without a watermark. A missing label tells you nothing, so look at where the video came from and how it behaves instead.

Can people spot deepfake videos by eye?

Not reliably. In a 2021 experiment with 210 people, published in iScience, participants could not reliably tell deepfakes from real videos. Warnings and cash rewards did not help, people tended to call fakes authentic, and they overestimated their own skill. That is why checking the source matters more than trusting your gut.

What should I do if a video call or voice message asks for money?

Don't act on the call itself. Hang up and call the person back on a number you already know is theirs, as the FTC and FBI advise. The FBI also suggests agreeing a secret word or phrase with your family in advance. A second channel you control works even when the face or voice looks and sounds right.

Sources

  1. Full Fact, "A fact checker's guide to spotting if something is AI", 11 April 2025, retrieved 2026-11-03, fullfact.org
  2. Lead Stories, "Fact Check: Video Showing 'Sergeant Pepper' Spraying A Shouting Protester In The Face Is NOT Real, Made With AI Video App", 8 October 2025, retrieved 2026-11-03, leadstories.com
  3. Engadget, "OpenAI is shutting down its Sora video generation app", 24 March 2026, retrieved 2026-11-03, engadget.com
  4. Google DeepMind, "Veo", retrieved 2026-11-03, deepmind.google
and 7 more
  1. Coalition for Content Provenance and Authenticity (C2PA), home page, retrieved 2026-11-03, c2pa.org
  2. Science Norway, "Researcher: How to spot AI-generated videos", 30 October 2025, retrieved 2026-11-03, sciencenorway.no
  3. Full Fact, "How to spot deepfake videos and AI audio", 20 December 2023, retrieved 2026-11-03, fullfact.org
  4. Köbis, Doležalová and Soraperra, "Fooled twice: People cannot detect deepfakes but think they can", iScience, November 2021, retrieved 2026-11-03, pmc.ncbi.nlm.nih.gov
  5. The Standard, "HK's first deepfake video conference scam involving HK$200 million", 4 February 2024, retrieved 2026-11-03, thestandard.com.hk
  6. US Federal Trade Commission, "Scammers use AI to enhance their family emergency schemes", 20 March 2023, retrieved 2026-11-03, consumer.ftc.gov
  7. FBI Internet Crime Complaint Center, "Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud", 3 December 2024, retrieved 2026-11-03, ic3.gov

Was this video made by AI?

We score one frame every 2 seconds and show you where AI appears in the video.