How to tell if a video is a deepfake: signs and checks

You often cannot tell a deepfake from a real video by looking, and the FBI now says so plainly. The reliable test is not the picture: it is checking where the video came from, looking for a content label, and confirming any request for money or access through a channel you already trust.

Visual and audio signs still catch the cheap fakes, so they are worth knowing. They just should not be the last line of defense.

What are the signs of a deepfake video?

The FBI's December 2024 warning on criminals using generative AI lists the imperfections to look for. Paraphrased into a checklist:

In the picture

• Faces that look indistinct or irregular
• Unrealistic teeth or eyes
• Distorted hands or feet
• Unrealistic accessories, such as glasses or jewelry
• Inaccurate shadows
• Unrealistic or unnatural movements
• Watermarks

In the sound

• Lag between the voice and the mouth
• A voice that is close to the real person but not quite matched
• Tone and word choice that do not fit how that person actually talks

A practical way to use the list: watch the clip more than once, at full size rather than in a small feed, and pause on the moments where faces and hands move. Small flaws are easier to see in a still frame than at full speed.

Why the signs are getting less reliable

The same FBI checklist comes with a warning. In a May 2025 alert about AI voice messages impersonating senior US officials, the FBI wrote that "AI-generated content has advanced to the point that it is often difficult to identify," and that real and cloned voices "can sound nearly identical."

Two things make the visual checks weaker over time:

1. Better tools leave fewer giveaways. The FBI's own wording moved from listing flaws in 2024 to warning in 2025 that AI content is often difficult to identify at all.
2. Compression hides what is left. A video that has been downloaded, re-uploaded and squeezed through a messaging app loses fine detail, and the artifacts you are looking for go with it. Real videos look rough after this too, which makes "it looks a bit off" a weak signal in both directions.

Can a deepfake detector tell me?

Not reliably. NIST, the US standards agency, reviewed detection methods in its November 2024 report on reducing risks from synthetic content. It found that detectors do best on content from the same generator they were trained on and much worse on others, especially after compression and resizing. In the cross-generator studies NIST cites, accuracy was between 50% and 70%. A coin flip is 50%.

So a free online checker that says "92% real" is one opinion, not an answer. Treat it as a hint and keep checking.

How to check a video in five minutes

These steps work on any video, real or fake, and none of them need special software.

1. Find the original upload. Search for the video's key words, or for the event it shows, and look for the first account that posted it. A clip that only exists on reposting accounts with no original is a warning sign by itself.
2. Check who would know. If it shows a public figure saying something surprising, look at that person's own verified accounts and at established news outlets. Real statements of that size get reported.
3. Look for a content label. See the next section for what to look for and what a missing label means.
4. Search for a fact-check. Searching the claim plus "fake" or "fact check" often turns up a debunk for a clip that has spread widely.
5. Ask what it wants from you. A video that pushes you to send money, click a link, buy something or share it right now is asking you to skip steps 1 to 4. That pressure is the strongest sign of all.

What content credentials and AI labels tell you

A growing number of cameras, apps and AI tools attach a signed record to a file that says how it was made and edited. The open standard behind this is called C2PA, and the records are called Content Credentials. The C2PA describes them as working "like a nutrition label for digital content."

According to the C2PA FAQ, the record is cryptographically signed, so any change to the file or its record breaks the signature and shows up as tampering. You can inspect a file's credentials at contentcredentials.org/verify.

On YouTube, two labels come from this system and from creator disclosure. Both sit in the expanded description, under How this content was made:

• Captured with a camera. YouTube's help page says this appears only when the video was recorded with tools that support C2PA version 2.1 or higher and has no edits to sound or visuals.
• An AI label. YouTube's page on disclosing AI content says creators must disclose realistic AI-generated or meaningfully altered content. YouTube may also apply the label itself when a file carries C2PA metadata or its systems detect AI. For photorealistic content the label can appear on the player too.

What a missing label means

Very little. Plenty of real videos carry no credentials at all, because support for them is still spreading across cameras and apps. And NIST notes that metadata "often is" stripped when files are shared, for example through social media. The C2PA FAQ also says a manifest can be separated from the file.

Treat labels as evidence only when they are there. A valid "Captured with a camera" credential is good evidence. An AI label is good evidence the other way. No label tells you nothing, and NIST adds a caution worth keeping: provenance "may contribute to trustworthiness but does not guarantee it." A real clip can still be cut out of context.

Deepfake video call scams: how they work and how to stop them

This is where deepfakes cost people money. The FBI says criminals "generate videos for real time video chats with alleged company executives, law enforcement, or other authority figures," and make videos to "prove" an online contact is a real person. Voice clones are used the same way on phone calls and voice messages.

The pattern is nearly always the same: a familiar face or voice, an urgent reason, and a request that cannot wait. Pay this invoice. Move this money. Read me the code we just sent you. Install this app so we can talk somewhere more secure.

Do not try to win by spotting glitches on a live call. Win by moving the conversation somewhere the scammer cannot follow:

1. End the call. Politeness is what the scam relies on. Say you will call back.
2. Call back on a number you already had. The FBI's advice is to look up the contact yourself and call it directly, not to use the number or link in the message.
3. Verify new numbers and accounts the old way. If someone you know contacts you from a new number or app, the FBI says to confirm it "through a previously confirmed platform or trusted source."
4. Use a family code word. The FBI recommends agreeing on a secret word or phrase with family members. A clone cannot know it.
5. At work, require a second channel for money. Any payment or bank-detail change requested on a call gets confirmed by a separate call or in person, no matter how senior the person on screen looks.
6. Never read out a sign-in code. The FBI warns that scammers talk people into sharing two-factor codes. No real colleague, bank or agency needs one. Our guide to two-factor authentication explains which methods resist this.

Many of these scams start or end with a link in an email. If one arrives after a strange call, check it with our guide on how to spot a phishing email, and if you already clicked, follow what to do if you clicked a phishing link.

If you think you were targeted

• Stop all payments and contact your bank on the number printed on your card.
• Change the password on any account you discussed or logged in to, and check its sign-in methods.
• Report it to the FBI at ic3.gov. The FBI asks for contact details used by the scammer, transaction details, and a description of how they reached you.
• Tell the person who was impersonated. They may not know their face or voice is being used.

To make yourself a harder target, the FBI suggests limiting how much of your image and voice is public: set social accounts to private and keep followers to people you know.

Frequently asked questions

Can you detect a deepfake with your eyes?

Sometimes. Low-effort fakes still show the signs the FBI lists, such as distorted faces, odd accessories and voice lag. Well-made ones often show none of them, which is why checking the source matters more than inspecting pixels.

Is there a free tool to check if a video is AI generated?

There are many, and none is reliable enough to trust alone. NIST found detector accuracy between 50% and 70% in cross-generator studies. Checking Content Credentials is more useful, because a valid credential is signed rather than guessed.

How do I know if a video on YouTube is AI?

Open the expanded description and look for the How this content was made section. YouTube shows an AI label there for disclosed or detected AI content, and a "Captured with a camera" label for verified camera footage with no edits.

Can someone deepfake me on a live video call?

The FBI says criminals already use real-time deepfake video on calls, posing as executives, police and other authority figures. That is why the defense is a callback on a number you already had, not a closer look at the screen.

Does a deepfake need a lot of my photos and voice?

The FBI's advice to limit public images and voice recordings suggests less material makes impersonation harder. It does not give a minimum, so the safe assumption is that any public video of you could be used.

Make a leaked address easy to trace

Deepfake scams work best when the attacker already knows your contact details. Secria Mail lets you give each site its own alias, so if one leaks or starts receiving targeted phishing, you can see which site it came from and stop it by deleting that alias. Tracking pixels are removed automatically and senders never see your IP address. Sign-up is a name, the address you want and a password, with no phone number. See how Secria Mail works or create a free account.