What is a deepfake?
A deepfake is AI made or AI altered media that imitates a real person. The US Government Accountability Office puts it this way: "A deepfake is a video, photo, or audio recording that seems real but has been manipulated with AI." The word joins "deep learning" and "fake", and MIT Sloan dates it to late 2017, when a Reddit user shared face swapped videos made with open source software.

People also search for deepfakes in the plural, and for a deepfake video, image or voice. The idea is the same in every case. A model learns how one person looks or sounds, then produces new footage or audio of them that was never recorded.
The bar to make one is low. GAO wrote in 2020 that "Anyone with basic computer skills and a home computer can create a deepfake." So this page is about checking, not making.
How are deepfakes made?
In concept, a deepfake is a neural network trained on many pictures or recordings of one person until it can produce new ones. GAO names two techniques. Autoencoders learn to compress a face and rebuild it. Generative adversarial networks, or GANs, pit two networks against each other: one forges, the other tries to spot the forgery, and each round the fakes get harder to catch.
MIT Sloan describes the classic face swap in one line. Train one autoencoder on the target's face and another on many faces, then pair the second one's encoder with the target's decoder, so the target's face appears on someone else's body. Voice clones work on the same idea with audio.
This page stops at the concept on purpose. It doesn't give steps for making a fake of a real person, and neither should any tool you trust. Making sexual or deceptive fakes of real people is what the laws below target.
What are real examples of deepfakes?
Four dated cases show the range, from a bank transfer to a phone call to a flood of videos. Each row is what the named source reports, and where the source itself flags doubt, the table says so.
| When | Case | What was reported | Source |
|---|---|---|---|
| March 2019 | UK energy firm voice call | A chief executive was reportedly tricked by a fake of his boss's voice into sending EUR 220,000; later reports questioned the evidence | MIT Sloan |
| 2019 | Deeptrace count | More than 14,000 deepfake videos online, double 2018, and 96% were pornography | MIT Sloan, citing Deeptrace |
| January 2024 | Biden robocall | A deepfake voice of the president went to New Hampshire primary voters; the FCC proposed a $6 million fine in May 2024 | FCC |
| Reported to police January 2024 | Arup video call | A finance worker joined a call with fake colleagues and sent HK$200 million (about US$25.6 million) in 15 transfers | CNN |
The Arup case is the one to remember. CNN reports that the worker first suspected a phishing email because it asked for a secret transaction, and dropped his doubts after the video call because the attendees looked and sounded like colleagues. His instinct was right. The call overrode it.
The Biden call has a legal sequel. The FCC fined consultant Steve Kramer $6 million in September 2024, and a New Hampshire jury acquitted him of all 22 criminal counts on 13 June 2025, according to NHPR. NHPR doesn't say whether the fine was paid. A deepfake can be real fraud and still hard to convict.
Are all deepfakes harmful?
No. GAO lists three potentially benign uses: entertainment, where a movie uses voices and likenesses for a creative effect when the performers are not available; e-commerce, where shoppers could virtually try on clothing with their own likeness; and communication, where speech synthesis can make a person appear to speak another language. Consent is the dividing line. The harm comes from imitating a real person without their say.
How can you spot a deepfake?
Look for small physical flaws, then check the source. The FBI's December 2024 alert lists distorted hands or feet, odd teeth or eyes, blurry or irregular faces, strange glasses or jewelry, incorrect shadows and watermarks. In video it adds lag, lips that don't match the audio and unnatural motion. For voice, it says to listen to tone and word choice and compare them with how the person normally speaks.
MIT Sloan adds blinking that looks off, eyebrows that don't fit the face, misplaced hair, skin that looks airbrushed, and reflections in glasses that ignore the lighting. Treat each as a hint. Models fix visible flaws fast, and a clean clip proves nothing.
GAO's warning from 2020 still frames the problem: "Current tools cannot perform a complete and automated analysis that reliably detects deepfakes." That is a six year old finding, so read it as the starting point of an arms race, not today's score. For how automatic detectors work and where they fail, see how AI detectors work.
The checks that hold up best don't depend on pixels at all. Agree on a private code word with family. Hang up and call the bank or the colleague back on a number you look up yourself. MIT Sloan's advice to businesses is the same in spirit: use a secret question at the start of a sensitive call.
What mistakes make people fall for a deepfake?
Four habits cause most of the damage, and each has a plain fix.
- Trusting a familiar face or voice. The fix: treat any unusual request for money or secrets as unverified until a second channel confirms it, as Arup's worker learned.
- Hunting for glitches in a polished clip. The fix: check where the clip came from and who else reported it.
- Believing a detector score as final. The fix: use it as one input, since GAO said in 2020 that even perfect detection might not stop a fake from working as disinformation.
- Sharing before checking. The fix: wait, search for the claim, and read how to fact check AI answers for the same routine applied to text.
Are deepfakes illegal in the US?
Some are, and the law got much tighter in 2025 and 2026. The TAKE IT DOWN Act was signed on 19 May 2025. Since 19 May 2026 it requires covered platforms to remove intimate images shared without consent, including AI made "digital forgeries", and known identical copies, within 48 hours of a valid request. The FTC enforces it and lists civil penalties of $53,088 per violation. The Act covers intimate images, not every deepfake.
State law fills in the rest. Ballotpedia counted 45 states with laws on sexually explicit deepfakes in its release dated 30 July 2025, up from 32 in January 2025, and 28 states with laws on political deepfakes. Check your own state's statute for the current rule.
A deepfake used to defraud someone is fraud under existing law whatever the tool. Beyond that, whether a fake is illegal depends on what it shows, where, and why it was shared. This page is general information, not legal advice. A lawyer can tell you what applies to your case.
How do you report a deepfake?
Report it where it was posted, then to the agency that fits the harm. For intimate images shared without consent, use the platform's removal form first. If the platform doesn't act within 48 hours or has no process, the FTC takes reports at takeitdown.ftc.gov.
For fraud, report to the FBI's Internet Crime Complaint Center at ic3.gov. Its alert asks you to include the scammer's identifying details, transaction records and a description of how the contact occurred. If you were targeted by a cloned voice, say so in the report and tell your bank.
If you are in danger or a child is involved, contact local police first. Keep screenshots, links and dates before anything is taken down.
What should you ask an AI about deepfakes?
An AI chatbot is good for explaining and planning, and poor for judging whether a specific clip is fake. These five prompts use it for what it does well. Paste them into Claude, ChatGPT, Gemini or Whizi.
- Spot the story: "Here is a claim and where I saw it: [paste]. List five checks I can do without special software, and the search terms for each."
- Family code word: "Help me set up a code word plan for my family, including what to do if someone calls in a panic asking for money. Keep it under 200 words."
- Workplace rule: "Draft a one page policy for [type of company] saying any payment request made on a video call must be confirmed on a second channel. Include who approves."
- Teach a class: "Write a 10 minute lesson for [age group] on what deepfakes are and how to check a clip. Use no instructions for making one."
- Check my law question: "I am in [state]. What questions should I ask a lawyer about a deepfake of me, and what evidence should I save first? Don't give legal advice."
Web search matters for the first prompt, because laws and tools change. Ask the model to give a link for every legal claim and open each one.
Where does Whizi fit?
Whizi is one app for the explaining, planning and drafting above, instead of a pile of single purpose tools. Web search is on by default in each chat, so a question about a recent law starts with fresh sources. The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card.
Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. Compare mode puts one question to two models side by side, which helps when you want a second opinion on a policy draft or a legal question list. You can also attach a screenshot and ask the model what looks odd.
Whizi doesn't verify whether a clip is real. No chatbot can promise that, so use it to plan checks, not to settle them.
What can deepfake checks not do?
They can't give certainty. A flaw can come from poor compression, and a clean clip can still be fake. GAO's paper is also from February 2020, so it describes detection methods that have since changed.
No visual check replaces confirming through a second channel, and no list of tells will stay current. Use the signs above as prompts to verify, and treat anything that asks for money or secrets as unverified until someone you trust confirms it.
- Pause on any clip or call that asks for money, a password or a secret
- Check the source of the clip, not just the face in it
- Confirm by a second channel before you move money
- Agree a code word with family and a callback rule at work
- Save evidence, then report to the platform, the FBI or the FTC
Frequently asked questions
What is a deepfake video?
A deepfake video is footage in which AI has swapped, changed or invented a real person's face or speech so that it seems real. GAO calls it a manipulated "video, photo, or audio recording". The Arup call, reported to police in January 2024, is a documented case of a fake video call used for fraud.
What is a deepfake scam?
A deepfake scam uses a fake face or voice to get you to send money or give up a secret. In the Arup case a finance worker sent HK$200 million, about US$25.6 million, in 15 transfers after a call with fake colleagues, CNN reported. Confirm any payment request on a second channel.
How can you tell if a video is a deepfake?
Check for odd hands, teeth, eyes, shadows, lag and lips that don't match the audio, then confirm the source. The FBI lists these signs in its December 2024 alert. None of them proves a clip is fake, so check where the clip came from before you trust or share it.
Is it illegal to make a deepfake?
It depends on the content and the state. The TAKE IT DOWN Act covers intimate images shared without consent, and Ballotpedia counted 45 states with laws on sexually explicit deepfakes as of its 30 July 2025 release. Fraud with a deepfake is fraud under existing law. Ask a lawyer about your case.