Deepfake Detection: How to Tell if a Video, Photo or Voice Is Fake

Deepfake detection works best as a combination of checks, not a single tool. Look at where the content came from and whether trusted sources confirm it, check for provenance information such as Content Credentials, look for visual or audio inconsistencies, and treat any detection tool’s result as one clue rather than a verdict. Detection tools are imperfect: they miss convincing fakes and sometimes flag real content, so context and provenance usually tell you more than pixels.

This guide covers the practical checks anyone can run, what detection software can and can’t do, how to handle voice-clone scams, and what to do if a deepfake is made of you. It does not explain how deepfakes are made.

What counts as a deepfake

A deepfake is synthetic or manipulated media, made with AI, that shows a real person saying or doing something they didn’t. The term covers more than face-swapped videos:

• Video. A person’s face or lip movements changed, or an entirely generated clip.
• Images. Generated or altered photos, including fake explicit images.
• Audio. A cloned voice used in a phone call, voicemail or voice note.
• Mixed media. Real footage with fake audio, or a real photo placed in a false context.

Not every misleading clip is a deepfake. Real footage that’s been cropped, slowed down or captioned falsely can mislead just as badly, and the same checks help with both.

Step 1: Check the source before the pixels

The most reliable deepfake detection doesn’t start with the image at all. It starts with where the content came from.

1. Who posted it first? Scroll back to find the original upload. A clip first posted by an anonymous account with no history deserves more suspicion than one from a named news outlet.
2. Do trusted sources confirm it? Search for the event or statement. If a public figure supposedly said something shocking, reputable outlets would usually be reporting it, often with their own footage.
3. Is there an original? A reverse image search on a screenshot or key frame can turn up the real photo or video the fake was built from. Our guide to reverse image search explains how.
4. Does the context fit? Check the date, location, weather, clothing and setting against what’s known. Fakes often get the surroundings wrong.
5. What does it want you to do? Content designed to make you angry, scared or urgent (share now, send money, click here) needs the most checking.

Step 2: Look for provenance and Content Credentials

Provenance is information about where a piece of media came from and how it was edited. It’s often more useful than guessing from visual clues, because it’s based on records rather than appearance.

• Content Credentials. Built on the C2PA open standard, Content Credentials attach signed information to a file about how it was created and edited, including whether AI tools were involved. Some cameras, editing apps and AI image generators add them, and some platforms display them. You can inspect a file’s credentials with a Content Credentials verification site.
• AI labels on platforms. Several major social platforms label content as AI-generated when they detect provenance data or when the uploader discloses it.
• Invisible watermarks. Some AI companies embed watermarks in the content their own systems generate, for example Google’s SynthID, which can be checked with that company’s tools.

The limits matter. Many platforms strip metadata on upload, screenshots remove it entirely, and a missing credential proves nothing either way. Most real photos have no Content Credentials at all. Provenance is strong evidence when it’s present and says something clear; its absence is simply not evidence.

Step 3: Look and listen for inconsistencies

Visual and audio clues can help, but treat them as prompts to check further, not proof. AI tools improve quickly, and signs that were obvious a short while ago often disappear.

In video and images

• Edges around the face, hair or glasses that blur, flicker or shift between frames.
• Lip movements that don’t quite match the words.
• Lighting and shadows on the face that don’t match the rest of the scene.
• Skin that looks unusually smooth, or teeth, ears and jewelry that change shape.
• Hands, text, logos and background details that warp or don’t make sense.

In audio

• Flat emotion, odd pacing or breathing that doesn’t fit the words.
• Background noise that cuts in and out unnaturally.
• A voice that sounds right but uses phrases the person never would.

Low-quality real footage can show many of the same signs, which is why these clues alone aren’t enough.

What deepfake detection software can and can’t do

A deepfake detection tool usually analyzes a file for statistical patterns that its developers associate with manipulation or generation, then gives a score or likelihood. These tools can be useful, especially for newsrooms, platforms and investigators who check a lot of content. But their limits are real:

Strength	Limit
Can flag patterns people can’t see	Often trained on known generation methods, so newer methods can slip past
Fast, and can process a lot of content	Can flag real content as fake, especially compressed, filtered or low-quality files
Gives a consistent method	Results can differ between tools for the same file
Useful alongside other evidence	A score isn’t proof, and courts and platforms may not accept it on its own

If you use deepfake detection software, run the original, highest-quality version of the file you can find rather than a re-uploaded copy, try more than one tool if the answer matters, and read the score as “worth investigating” rather than “confirmed”. Be especially careful before publicly accusing someone of posting a fake, or of being fake, based only on a tool. False accusations cause their own reputational damage.

The same caution applies to detectors for AI-written text, which have similar accuracy problems. Our guide to AI content detectors covers those.

Voice-clone scams: the practical defense

The deepfakes most likely to reach ordinary people and small businesses are voice calls: a “family member” in trouble asking for money, or a “boss” asking for an urgent payment. Detection by ear is unreliable here. Process works better.

• Hang up and call back on a number you already know, not one given in the call or message.
• Agree a family code word that someone would need to know to prove it’s really them.
• For businesses, require a second check for payment changes or unusual transfers: a call to a known number, or approval from a second person.
• Slow down. Urgency and secrecy are the scammer’s tools. A real emergency survives a two-minute call back.

If you’ve already lost money, contact your bank straight away and report it to the FTC. Our guide on what to do if your identity is stolen covers the reporting steps if personal details were involved.

A worked example

This is an illustrative scenario, not a real client. A dental practice manager in Arizona receives a short video on a local community page that appears to show the practice owner making rude remarks about patients. It’s being shared quickly.

She doesn’t reply to the post. First she checks the source: the account that posted it was created that week and has no other posts. The owner confirms he never said it and was on vacation on the date shown. A reverse image search on a frame finds the original, a genuine interview he gave to a local business magazine, with completely different audio. She runs the clip through a detection tool, which returns an inconclusive score, so she doesn’t rely on it.

With the original interview and the account details, she reports the post to the platform as manipulated media and impersonation, and the practice posts a short, calm statement linking to the real interview. The provenance of the original, not the detection score, is what made the case clear.

If a deepfake has been made of you

Being targeted by a deepfake is upsetting, and it isn’t your fault. What to do depends on the kind of fake:

• Sexual or intimate fakes. Report them to the platform as non-consensual intimate imagery, and use StopNCII.org for adults or NCMEC’s Take It Down if the person shown is under 18. Never screenshot or save a sexualized image of a minor; report the link instead. Our guide to AI-generated images of you walks through every removal route.
• Fake endorsements or scam ads. Report them as impersonation or fraud to the platform, warn your audience publicly, and consider talking to a lawyer.
• Fakes designed to damage your reputation. Record the links, report them, and get legal advice if the fake is presented as real and causes harm, since it may be defamatory.
• Fake accounts using your face or voice. Our guide to online impersonation covers reporting.

If there are threats, demands for money or a risk to your safety, contact the police. Our help resources page lists support organizations. For fakes that are spreading or affecting your business, our crisis management team can help plan the response.

Common mistakes

• Trusting one tool’s score. Use it alongside source and provenance checks.
• Sharing a clip to ask if it’s fake. That spreads it. Check first, share later if at all.
• Assuming real means unedited. Genuine footage can still be clipped or captioned to mislead.
• Calling real content fake. Dismissing genuine evidence as a deepfake is its own form of misinformation.
• Relying on your eyes and ears for calls about money. Call back on a known number instead.

Frequently asked questions

Is there a free deepfake detection tool that's accurate?

Some free and commercial tools exist, but none is reliably accurate on every file. They can miss new kinds of fakes and flag real content. Use any tool as one clue alongside the source, context and provenance checks.

Can you tell a deepfake just by looking?

Sometimes, but less and less reliably. Visual clues like odd edges, mismatched lighting or strange hands help, yet good fakes may show none of them and poor-quality real videos may show several. Checking the source is more dependable.

What are Content Credentials?

Content Credentials are signed information attached to a file, based on the C2PA standard, recording how it was made and edited, including any AI involvement. When present they’re strong evidence, but many files have none, and a missing credential doesn’t mean content is fake.

What should I do if someone makes a deepfake of me?

Record the links, report the content to the platform, and use the removal routes for the type of fake, such as StopNCII for intimate images of adults. Contact the police if there are threats, and talk to a lawyer if the fake is damaging your reputation.