How do you spot a deepfake?
Often you cannot spot a deepfake by looking. In a 2026 University of Florida study, people judged fake face photos no better than chance, though they sorted real and fake videos correctly about two-thirds of the time. The safer test is the request: check any urgent plea for money another way.
Picture a grandmother whose phone rings late at night. The voice sounds like her grandson. He says he is in trouble and needs money right away. This is an illustration, not a real case. She knows that voice. She has known it his whole life. She reaches for her purse.
That is the problem. Tools built on artificial intelligence (AI) can now copy a face or a voice. The Federal Bureau of Investigation (FBI) warns that criminals make short audio clips in a loved one’s voice to pose as a relative in a crisis. So a familiar voice is not proof on its own.
Contents
Why is it hard to spot AI fakes?#
Because detection and fakes keep trading places, and the FBI says AI-made content is getting harder to detect. A deepfake is a picture, video or voice made or changed by AI to show someone doing or saying something they never did. Newer tools can remove the flaws people learned to spot.
The FBI says so plainly. In its annual internet crime report, covering last year, it calls AI-made content “increasingly difficult to detect and easier to make.”
The FBI’s Internet Crime Complaint Center received 22,364 complaints that reported AI-related information last year. Losses in those complaints topped $893 million, close to a billion dollars. Those are only the reported cases.
The machines that check for fakes face the same race. The National Institute of Standards and Technology, the federal agency that sets technical standards, calls automated detection “a constant cat-and-mouse game.” As soon as a new detection method appears, the tools that make fakes improve. It is like a counterfeit bill that learns to pass whatever test the store clerk uses.
What are some warning signs that might be a deepfake?#
The FBI’s public guidance lists the flaws that still show up in fakes: odd hands, teeth and eyes, shadows that fall wrong, and sound that lags the picture. They are worth knowing, as long as you do not count on finding them. Newer tools can remove some of them.
In images and video, the FBI says to look for:
- Hands, feet, teeth or eyes that look wrong.
- Blurry faces, or odd glasses and jewelry.
- Shadows that fall the wrong way.
- Lag, or movement that looks stiff.
On a call, listen to the tone and the choice of words. A cloned voice can sound right and still say things the real person would never say.
The FBI’s report also describes job applicants using fake voices, possibly voice deepfakes, in online interviews. The lip movement of the person on camera does not fully match the audio. Coughs and sneezes may not line up with what the camera shows. The FBI says the goal generally appears to be getting into private computer networks. It sees no significant dollar loss tied to that pattern.
The same habit of checking helps with claims about AI rules too.
These signs raise or lower the odds. They never settle the question.
Can you tell if a video is real or fake?#
Video is the one place people still have an edge, and it is small. In a 2026 University of Florida study, people sorted real and fake videos correctly about two-thirds of the time. On still photos of faces, they did no better than chance.
The researchers tested thousands of participants and published the results in January 2026, in the journal Cognitive Research: Principles and Implications. Computer programs were up to 97 percent accurate at spotting the fake face photos.
Videos turned out differently. On video, people did better than the computer programs. The lead author said the team was “a little shocked” by that result, as the University of Florida reported.
Two-thirds is better than a coin toss. It still means being wrong about one time in three. That is not safe. Nobody should bet savings on odds like that.
The researchers suggest that for still images, automated detection may now do better than human judgment alone. For video, a person’s eye still adds something. Neither is enough to send money on.
How to protect yourself from AI scams?#
Check the request. Ignore the face. A fake can copy how someone looks and sounds. It cannot copy a fact only your family knows, or a phone number you already have. That habit is far harder for a fake to beat. Hang up first.
The FBI’s report shows where the money is lost. The largest losses with an AI link came from investment scams, not from the family emergency call this article opened with. Among the smaller scams, business email scams cost the most.
Investment scams with an AI link cost victims more than $632 million last year. Business email scams with AI cost more than $30 million. Romance scams with a likely AI link cost more than $19 million. The FBI counts family emergency scams, where a cloned voice mimics a loved one in distress, inside that romance group. They cost more than $5 million. Job scams involving AI cost almost $13 million.
Even the largest of those numbers is a floor. Total investment scam losses topped $8 billion, more than twelve times the figure with an AI link. The FBI says many victims do not realize how far AI may be involved.
The FBI’s advice comes down to a few habits:
- Agree on a secret word or phrase with your family to confirm who is calling.
- If a call asks for money, hang up. Look up the number yourself and call back.
- Keep less of your face and voice online, and make social media accounts private.
A secret word costs nothing and cannot be copied from a video. It works like the knock a child learns before opening the front door.
Older adults report internet crimes of all kinds in large numbers. The FBI’s crime center received 201,266 complaints from people 60 and older last year, 37 percent more than the year before.
This site has reported on fraud losses after 60, and the same habits apply there.
Are deepfakes illegal, and what can you do?#
Some uses are, and at least one federal rule now forces platforms to act. It is the Take It Down Act, and it covers intimate images posted without consent, including fakes. For other fakes, such as a cloned voice used to ask for money, the path is a fraud report.
The Federal Trade Commission began enforcing it on May 19, 2026. Platforms must remove an intimate image, and known identical copies, within 48 hours of a valid request from the victim.
The agency’s consumer guidance says the law applies to real images, digitally altered images and deepfakes made with AI. A victim can report a platform that fails to act at TakeItDown.ftc.gov.
That rule is narrow. It covers intimate images only. Fraud is different. A fake voice used to take money is a fraud case instead. The FBI asks victims of financial fraud to file a report with its Internet Crime Complaint Center at ic3.gov. This article covers the federal rule only, not state laws.
This site has also written about AI outrunning the public.
Back to the grandmother and her late phone call. The voice may be perfect. The safer test is not the picture. It is the request. She can hang up, call her grandson on the number she already has, or ask for the family’s secret word.
None of that needs expertise. It needs a pause. One minute before she sends anything is enough. Her family’s own check beats her ears.
Source: University of Florida News, deepfake detection study.