How to Tell if a Video Is AI-Generated

Before you touch the pause button

A short clip lands in a group chat. It shows a politician, a fire, a celebrity or somebody who looks a lot like your boss. The natural reaction is to stare at the face. Resist it for a minute. First ask where the clip came from. Who posted it, when did they post it, and do they explain how they obtained it?

This sounds almost too basic, but it catches a surprising amount of bad information. Real footage is constantly reposted with a new date or location. Old satire loses its label. A cropped interview loses the question that made the answer make sense. None of those problems requires generative AI, and no deepfake detector can fix them for you.

• Copy a distinctive phrase from the caption and search it in quotation marks.
• Look for a longer upload, not just the most popular one.
• Check whether local news, eyewitnesses or official sources show the same event.
• Save the post, account name and timestamp before anything is deleted or edited.

Watch once, then watch differently

Play the whole video at normal speed first. Do not hunt for glitches yet; work out what it is actually claiming. On the second viewing, turn off the sound. On the third, slow down the moments when a face turns, a hand passes in front of it, or somebody moves from shadow into bright light.

Those transitions are useful because the system has to keep identity, lighting and motion consistent at the same time. You may notice an earring vanish, teeth change shape, skin briefly stick to the background or a reflection lag behind. What matters is repetition. One ugly frame is often just compression or motion blur. The same failure appearing several times is harder to dismiss.

The soundtrack may be a separate problem

Picture and sound do not have to share the same origin. A genuine press-conference video can carry cloned speech; an authentic voice can be placed over an invented scene. Listen through headphones if you can. Do consonants meet the lips? Does the room tone jump between sentences? Are breaths missing in places where a speaker would normally need one?

For a public figure, the quickest answer is often elsewhere. Find the broadcaster's full recording, the venue feed or the speaker's verified channel. Comparing the disputed clip with the original is more persuasive than arguing over whether a mouth movement looks slightly unnatural.

Where a detector earns its place

A detector can pick up patterns that are not obvious on screen, which makes it a useful second opinion. Give it the best copy you can find. A screen recording of a compressed social-media preview has already thrown away much of the information the model might use.

Then read the result narrowly. A suspicious score means the file deserves more checking. A low score means that this detector did not find strong synthetic signals in this copy. Neither result tells you whether the caption is true. And an inconclusive answer is just that: not a secret way of saying 'real.'

A five-minute review

Write the claim in one sentence, trace the earliest useful upload, reverse-search two clear frames, inspect unusual transitions and then run a detector. Compare all the findings. If they disagree, leave the conclusion open.

That last part matters. If a clip could affect someone's safety, job, money or reputation, five minutes is only triage. Keep the original file, document what you checked and seek a qualified forensic review. Being unsure for longer is better than being confidently wrong in public.