Blog Media literacy

How to spot a deepfake: a 10-minute check for video, images and voice

Spot a deepfake by checking the source first, then faces, audio and provenance labels. A 10-minute routine, free tools and what US law says.

Illustration of a smartphone showing a video frame with a pixelated half face under a magnifying glass, next to a checklist, an audio waveform and a padlock shield

Short answer: to spot a deepfake, check where the clip came from before you study the pixels. Find the original uploader, look for the same event from a second trustworthy outlet, run a reverse image search, and only then inspect the face, the lighting and the voice. Visual tells still help, but the newest generators hide most of them, so the source check catches more fakes than any single glitch.

A deepfake is a video, image or audio clip in which a real person’s face, body or voice has been generated or swapped by software so that they appear to say or do something they never did. The detection tools fact-checkers use are themselves computer vision models trained on large sets of real and synthetic frames. Ask any computer vision development company that builds such a detector and you will hear the same caveat researchers give: a model that catches yesterday’s generator can miss tomorrow’s. That is why people still need a routine of their own.

This guide gives you that routine: the clues in faces, images and voices that still work, the free tools worth knowing, how provenance labels such as Content Credentials work, and what US law now says about synthetic media. If you mostly worry about jokes that travel without their label, start with our guide on how to tell satire from fake news.

Key takeaways. Source first, pixels second. No single visual clue proves a fake, and the old “they never blink” rule no longer works. Free tools (reverse image search, provenance checkers, detectors) give hints, not verdicts. When a clip pushes you to act fast, pay or share, slow down: urgency is the most reliable warning sign of all.

What is a deepfake, and why is it harder to spot now?

A deepfake is synthetic media made with machine learning, usually to put words or actions in a real person’s mouth. The term covers four common formats:

  • Face swaps, where one person’s face is mapped onto another person’s body in a video.
  • Lip-sync fakes, where a real video is kept but the mouth is re-animated to match new audio.
  • Voice clones, where software imitates a specific person’s voice from samples of their speech.
  • Fully generated images and video, made from a text prompt with no original footage at all.

Spotting them got harder for three reasons:

  • Better generators. The warped teeth and melting earrings of early fakes are now rare.
  • Cheap tools. Consumer apps mean fakes are no longer limited to skilled hobbyists.
  • Small screens. Most people watch video on a phone, scrolling fast with the sound off or low, which hides exactly the details that used to give fakes away.

Research backs this up. In a 2022 study in the Proceedings of the National Academy of Sciences, Matt Groh and colleagues found that ordinary viewers and a leading detection model were similarly accurate overall, but they made different mistakes. Neither was reliable alone. That finding is the best argument for a checklist that combines context, close looking and tools.

How do you check a suspicious clip in 10 minutes?

Run these steps in order and stop as soon as one gives you a clear answer. Most fakes fall at step two or three, long before you need to zoom into anyone’s face.

The seven steps, in order

  1. Pause and notice your reaction (30 seconds). If the clip makes you furious, delighted or scared, that is the moment to slow down. Content built to provoke is the content most worth checking.
  2. Find the original uploader (2 minutes). Who posted it first? A reposted clip with no source, a cropped watermark or a brand-new account is a red flag. Search the platform for the earliest copy.
  3. Look for independent confirmation (2 minutes). If a public figure really said something shocking, several established outlets will be reporting it. Search the claim, not the clip. Fact-checkers call this lateral reading: leave the post and check what other sources say about it.
  4. Reverse search a frame (2 minutes). Take a screenshot of a clear frame and run it through Google Lens or TinEye. You may find the original, unaltered video or the same image from an older, unrelated event.
  5. Check provenance labels (1 minute). Look for a Content Credentials icon, a platform “AI-generated” label or a SynthID result (see below). A label can confirm a fake. A missing label proves nothing.
  6. Inspect the media closely (2 minutes). Only now look at faces, hands, lighting, text and audio, using the clue lists in the next sections.
  7. Decide and act (30 seconds). If you are still unsure, don’t share it. If it targets you or someone you know, report it (see “What to do if you think you’ve found a deepfake”).

Which free tool fits each step?

Each step checks something different, so each catches a different kind of fake:

StepWhat you are checkingFree toolCatches
SourceWho posted it first and whyPlatform search, account historyReposts, impostor accounts
ConfirmationWhether trusted outlets report itNews search, fact-checking sitesInvented events and quotes
Reverse searchEarlier or original versionsGoogle Lens, TinEyeRecycled or altered footage
ProvenanceSigned history or AI watermarkContent Credentials verify, SynthIDLabeled AI output
Close lookVisual and audio artifactsYour eyes, ears and a big screenLower-quality fakes

What visual clues still give away a deepfake video?

The clues below still catch many fakes, especially cheap ones, but treat each as a reason to keep checking rather than proof. The MIT Media Lab’s Detect Fakes project, which tested how people judge manipulated video, suggests paying attention to the same parts of the face.

Face, skin and edges

  • Skin that is too smooth or too old. Does the skin of the cheeks and forehead match the age of the neck, hands and hair? Swapped faces often look airbrushed next to an untouched body.
  • Edges that shimmer. Watch the hairline, jaw and ears when the head turns. Blurring, a faint outline or a face that slides slightly against the head are common artifacts.
  • Facial hair and moles. A beard that looks painted on, or a mole that appears and disappears between shots, deserves a second look.

Eyes, glasses and mouth

  • Reflections and glare. Light in both eyes, and on glasses, should match the light sources in the room. Mismatched or missing reflections are a strong hint.
  • Blinking. Early deepfakes blinked rarely. Modern ones blink normally, so blinking is now a weak signal at best.
  • Lip sync. Sounds such as “b”, “m” and “p” require closed lips. If the mouth stays open on those sounds, or the teeth look like a single white strip, the audio and video may not belong together.

Light, shadows and motion

  • Shadows that disagree. The shadow of the nose and chin should fall the same way as shadows on the wall behind.
  • Frame flicker. Step through the video slowly. A face that briefly distorts when a hand passes in front of it, or when the person turns sideways, is a classic failure point.
  • Stiff body language. A lively face on a frozen neck and shoulders can mean only the face was generated.

Watch the clip at least once on a larger screen with the sound up. Many of these clues are invisible at phone size and full speed.

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

Fully generated photos rarely have a face-swap seam, so you look for different errors: things that don’t make physical or logical sense.

  • Hands and small objects. Extra or fused fingers are less common than they were, but jewelry, watch straps, glasses frames and buttons still often merge or break.
  • Text in the scene. Signs, banners, badges and license plates may contain letter-like shapes that read as nonsense up close.
  • Backgrounds. Crowds with repeated or melting faces, windows that don’t line up and railings that dissolve into walls are common tells.
  • A too-perfect look. Glossy skin, ideal lighting from no visible source and a magazine-cover finish on what claims to be a snapshot from a protest.
  • Missing history. A real news photo usually has a photographer, an agency and other images from the same scene. A single, perfect image with no siblings is suspicious.

Technical checks can add evidence. Reading a file’s metadata (EXIF) with a free viewer may show the camera model or editing software, though most platforms strip metadata on upload, so an empty result means little.

How do you detect a cloned voice?

Voice clones are now the most common deepfake in fraud because they are cheap to make and work over the phone, where you can’t see anything at all. Listen for:

  • Urgent words in a flat tone, or an emotion that stays the same when the topic changes.
  • No breaths at all, breaths in odd places, or a rhythm that is too even.
  • Studio-clean audio from someone who claims to be at a crash scene or a police station.
  • A voice that dodges personal questions and keeps drifting back to its script.

Honestly, though, the defense that works best is not listening harder. Hang up and call the person back on a number you already have. Families and small offices can also agree on a code word that a caller must say before anyone sends money or passwords.

Can a live video call be a deepfake?

Yes. Real-time face and voice swapping now runs on consumer hardware, and it has already been used for large frauds. In early 2024, police in Hong Kong said an employee at the engineering firm Arup transferred about $25 million after a video meeting in which the company’s finance chief and other colleagues were all deepfakes.

If something feels off on a call, ask for things a live swap handles poorly:

  • Turn your head fully to the side and back.
  • Pass a hand slowly in front of your face.
  • Stand up, or hold an object near your face.
  • Answer a question only the real person would know.

Then confirm any request for money, credentials or confidential files through a separate channel. A real colleague will not mind the call-back.

What are Content Credentials and SynthID, and can they prove a clip is real?

Provenance tools record where a file came from instead of guessing from its pixels. They are useful, but none of them is universal yet.

Content Credentials (C2PA)

Content Credentials is an open standard developed by the Coalition for Content Provenance and Authenticity (C2PA), backed by companies including Adobe, Microsoft, Google, OpenAI, Sony and the BBC. A camera or editing app can attach a signed record to a file showing how it was made and edited. Where it is supported, a small “CR” pin appears on the image. Clicking it, or uploading the file to the free verify tool, shows the history.

SynthID watermarks in Google’s AI content

Google DeepMind’s SynthID embeds an invisible watermark in images, video, audio and text made with Google’s AI tools, designed to survive cropping, filters and compression. Google offers checks through its own products and a SynthID Detector portal.

What provenance can’t tell you

  • A missing label proves nothing: most cameras and apps don’t add credentials, and platforms often strip them.
  • A watermark only shows up for content made with tools that add it. Open-source generators usually don’t.
  • A credential confirms a file’s history, not that what it shows is true or fairly framed.

Do deepfake detection tools actually work?

Detection tools give useful signals but not verdicts. They are trained on known generators and lose accuracy on new ones, on heavily compressed social video and on short clips. Use them as one input alongside the source checks above.

ToolWhat it doesBest forLimit
Google Lens / TinEyeFinds where an image or frame appeared beforeRecycled or altered footageSays nothing about brand-new fakes
InVID-WeVerify pluginSplits video into keyframes, reverse searches them, reads metadataVerifying viral videosBuilt for journalists; takes practice
Content Credentials verifyReads the signed history of a fileFiles from supporting cameras and appsMost files carry no credentials
SynthID DetectorLooks for Google’s AI watermarkContent made with Google toolsBlind to other generators
DeepFake-o-Meter (University at Buffalo)Runs several research detectors on an uploadA second opinion on a face videoResults can disagree; not a verdict

Be wary of any app that promises near-perfect accuracy. Researchers regularly report large drops in performance when detectors move from lab datasets to real social media clips.

Which context red flags matter more than pixels?

The strongest clues often sit outside the frame. Ask what the clip wants from you:

  • Urgency. “Share before they delete this” or “send the money today” is pressure, not information.
  • A request. Money, gift cards, crypto, a password or a login code turn a strange clip into a likely scam.
  • Perfect timing. A shocking clip of a candidate that appears days before an election, with no original source, deserves extra suspicion. We covered the rules for AI-generated political ads in the 2026 midterms separately.
  • Out of character. A public figure saying something wildly unlike their record, with no other outlet reporting it.
  • Low-quality reposts. Fakes often circulate as screen recordings of screen recordings, which hides artifacts and strips provenance.

Are deepfakes illegal in the United States?

Making a deepfake is not illegal in itself, but using one for fraud, harassment or certain political and robocall purposes can be. The main federal rules so far:

  • AI voices in robocalls. In February 2024, the Federal Communications Commission ruled that AI-generated voices count as “artificial” under the Telephone Consumer Protection Act, making voice-cloned robocalls without consent illegal (FCC announcement). The ruling followed a January 2024 robocall that imitated President Biden’s voice before the New Hampshire primary; the FCC later fined the political consultant behind it $6 million.
  • Non-consensual intimate images. The TAKE IT DOWN Act, signed in May 2025, makes publishing non-consensual intimate images, including AI-generated ones, a federal crime and requires covered platforms to remove them within 48 hours of a valid request.
  • Fraud. Existing wire fraud and identity theft laws apply when a deepfake is used to steal money.

Many states add their own rules, especially on deepfakes in election ads and intimate images, and they differ widely. Outside the US, the EU AI Act requires that AI-generated or manipulated content be disclosed, with those transparency rules applying from August 2026.

What should you do if you think you’ve found a deepfake?

  1. Don’t share it, even to debunk it. If you must discuss it, describe it or use a screenshot marked as fake.
  2. Save evidence: the link, the account name, the date and a screen recording.
  3. Report it to the platform using its synthetic or manipulated media option.
  4. If money or accounts are involved, contact your bank immediately and report to the FBI’s Internet Crime Complaint Center (IC3) and the FTC at ReportFraud.ftc.gov.
  5. If it shows you or someone you know in intimate imagery, use the platform’s removal process and keep a copy of your request.
  6. Tell people close to you. Voice-clone scams often target relatives; a quick warning and a family code word help.

How do you talk to family and friends about deepfakes?

Lead with habits, not fear. Telling an older parent that “anything could be fake” tends to produce either panic or cynicism about all news. A shorter message works better:

  • Agree on a code word for emergencies and a rule: no money moves without a call-back.
  • Practice the three-question check together: who posted it, who else reports it, what does it want from me?
  • Share a debunked example from a fact-checking site so the idea is concrete.
  • Make it normal to ask, “Is this real?” in the family chat without anyone feeling foolish.

For newsrooms and newsletter writers, the same habits apply with higher stakes: write down your verification steps, as we do in our editorial standards, and say in print how you checked a viral clip.

Frequently asked questions

What is the easiest way to spot a deepfake?

Check the source before the pixels. Find who posted the clip first, see whether established outlets report the same thing and run a reverse image search on a frame. Most fakes fail these checks before you need to study faces or lighting.

Can you still spot a deepfake by blinking?

Not reliably. Early deepfakes blinked rarely, but modern generators produce normal blinking. Treat blinking as a weak hint at most and look instead at mismatched reflections, lip sync on b, m and p sounds, shimmering edges and the source of the clip.

Are free deepfake detectors accurate?

They give useful signals but not verdicts. Detectors are trained on known generators and lose accuracy on new tools, compressed social video and short clips. Use them alongside source checks, reverse image search and provenance labels such as Content Credentials.

How can I tell if a phone call uses a cloned voice?

Listen for flat emotion, odd breathing and audio that is too clean for the claimed setting, but the safest test is to hang up and call back on a number you already know. Families can also agree on a code word for emergencies.

Is it illegal to make a deepfake in the US?

Making one is not illegal in itself, but using it for fraud, non-consensual intimate images or robocalls can be. The FCC ruled in 2024 that AI-generated voices in robocalls fall under the TCPA, and the TAKE IT DOWN Act of 2025 covers intimate deepfakes. Many states add election and privacy rules.

From the blog of The Smattering, where we write plainly about satire, media literacy and how we work. More posts.

Get the next one first

Free satire in your inbox once or twice a week.