TL;DR. A deepfake is an image, video or audio clip that AI has edited or generated, including clips that make a real person seem to say or do something they never did. People were right about two in three videos in one study, and the best commercial video detector in a 2024 benchmark scored 78%. Check the source before the pixels.
Fortune reported on May 17, 2024 that an employee at the Hong Kong office of the firm Arup had joined a video call with deepfakes of the company’s chief financial officer and other staff. After first doubting the call, the employee sent $25.6 million (HK$200 million) in 15 transactions. Arup said: “We can confirm that fake voices and images were used.”
A deepfake is an image, video or audio clip that artificial intelligence has edited or generated. The ones that cause harm, including every case on this page, look or sound like a real person. In plain terms: it is a recording that seems to show someone but was made, or changed, by software.
Last reviewed: October 28, 2026. Reviewed quarterly.
What a deepfake actually is
A deepfake is media that has been “edited or generated using artificial intelligence,” in Wikipedia’s wording, and the name blends “deep learning” with “fake.” The word “originated in late 2017 from a Reddit user named” deepfakes, who shared face-swapped videos in a community called r/deepfakes. Since then it has come to cover far more than face swaps.
Two things set a deepfake apart from an ordinary edited photo:
- AI did the work. The face, voice or movement was produced or altered by a trained model, not drawn by hand or cut together from real clips.
- It can stand in for a real person. U.S. law, when it addresses deepfakes of intimate images, speaks of an “identifiable individual.” That is where the harm comes from: someone’s face or voice used without them.
A deepfake is not the same as the low-quality, mass-produced AI content people call AI slop, which is a flood rather than a targeted fake. It is also broader than AI voice cloning, which is one kind of deepfake, the audio kind.
How a deepfake is made
Wikipedia describes several ways to make a deepfake. One of them is the autoencoder, and here is the plain version of how it makes a face swap, followed by the technical detail.
In technical terms, “Deepfakes rely on a type of neural network called an autoencoder,” as Wikipedia puts it: an encoder compresses a face into a latent representation, and a decoder trained on the target rebuilds it, which makes face swapping possible. A second method uses a generative adversarial network, or GAN, in which a generator and a discriminator train against each other, which pushes the fakes toward realism. A third, lip-sync, edits a video so a person appears to mouth words from a different audio track.
Voices need the least material. According to Wikipedia, voices can be cloned from only seconds of speech, and the FBI’s warning of December 2024 covers cloned audio as well as fake video.
The four forms you will meet
Four forms show up in published sources. The FBI’s December 3, 2024 public service announcement (I-120324-PSA) names the audio and video ones as they are used in fraud.
| Form | What it is | Example from a published source |
|---|---|---|
| Face swap | One person’s face put on another person’s video | The 2017 Reddit community that gave the word its name |
| Lip-sync | A real video edited so the mouth matches new audio | Described by Wikipedia as one technique |
| Voice clone | AI audio that mimics a real person | FBI: clips that mimic a loved one’s voice to demand urgent money |
| Live fake video | Fake video in a real-time call | FBI: fake video chats posing as company executives or law enforcement |
The FBI adds two more uses of AI video: videos sent privately to “prove” that an online contact is a real person, and fabricated promotional videos for investment fraud.
Where deepfakes do harm
The cases below come from published reporting or from the text of a law.
| Case | What happened | Source |
|---|---|---|
| Arup, reported May 2024 | A Hong Kong employee sent $25.6 million in 15 transfers after a video call with fake colleagues | Fortune, May 17, 2024 |
| A Senate race, October 20, 2026 | Two videos of a man accusing a candidate of abuse; PolitiFact rated the claim Pants on Fire | PolitiFact, October 22, 2026 |
| Impersonated voices | AI audio used to impersonate people and get into bank accounts | FBI, December 3, 2024 |
| Intimate “digital forgeries” | Made a federal crime to publish, with a 48-hour takedown duty for platforms | TAKE IT DOWN Act, May 19, 2025 |
The Senate-race case is worth a second look because of how it was built. The Foundation for Defense of Democracies (FDD) reports that operators began in September by creating an X account that supposedly belonged to a real man, a graduate of a middle school. The real man told the fact-checking site Lead Stories that the person in the videos was not him. The account borrowed an ordinary person’s name and profile picture.
FDD adds a detail that matters for anyone who uses a chatbot to check things: it says AI search tools “quickly picked up and repeated the claim,” with Google’s Gemini and Microsoft Copilot presenting the accusation as real until fact-checking reports appeared.
Can you spot a deepfake?
Not reliably by eye, and not reliably by software on video, which is the main fact to carry away. In a University of Florida study, people did no better than chance on deepfake face images, while programs were “up to 97% accurate” on them. On video the picture flipped: people got real and fake videos right “about two-thirds of the time,” and the algorithms’ performance “declined sharply” to chance levels. The researchers note the study used specific faces and videos under controlled conditions.
The Deepfake-Eval-2024 benchmark, built from deepfakes that circulated online in 2024, tells a similar story for tools. Its video set has 1,072 real and 964 fake clips, and the best commercial video detector scored 0.78 accuracy.
Even Google’s own tool is not a verdict. On October 21, 2026, Full Fact reported that Google’s public SynthID Detector had flagged five apparently genuine images as made or edited with Google AI. Google rolled out an update on October 15, and in Full Fact’s report the five come back as “Unsure.” Full Fact’s conclusion is that the tool is useful but should be combined with other checks. For how to do that step by step, see our guide to checking whether a video is real.
The advice that holds up across sources is to start with where the clip came from. Full Fact says to find the highest-quality version, search for the original footage, and check whether trustworthy outlets reported the same thing, and it warns that there are “few guaranteed ways” to tell. Two sibling terms help here: SynthID is an invisible watermark, and Google’s October 7 announcement says its detector covers content from Google and partners such as OpenAI, NVIDIA and Kakao. C2PA (Content Credentials) is a record of where a file came from. Neither is a verdict: Google’s help page says a missing watermark means the content “could have been created by other AI systems,” and that Content Credentials don’t tell you whether a file is definitely AI-generated.
What the law says (United States, federal)
One U.S. federal law that addresses deepfakes directly is about intimate images. The TAKE IT DOWN Act, signed on May 19, 2025 as Public Law 119-12, defines a “digital forgery” as an intimate visual depiction of an identifiable individual created with software, machine learning or AI that a reasonable person would find indistinguishable from an authentic one.
It does two things. First, it makes it unlawful to knowingly publish such an image using an interactive computer service, with penalties of up to 2 years in prison where the victim is an adult and up to 3 years where the victim is a minor. Second, it requires covered platforms to remove a valid request “as soon as possible, but not later than 48 hours after receiving such request,” and to make reasonable efforts to remove known identical copies. The Act gave platforms one year from enactment to set up a removal process, and the FTC enforces its notice-and-removal requirements.
That is all this page claims about the law. Other deepfakes, such as a fake executive on a video call or a fake political clip, may fall under other laws or none, depending on the use and the place. This page is not legal advice.
What this means for accountants and bookkeepers
A deepfake in finance can be a request to move money, and the Arup case is an example: a video call with deepfakes of the CFO and other staff, and 15 transfers. The FBI lists fake video chats posing as company executives among the ways criminals use AI, and its general advice is to verify by hanging up and contacting the organization through a number you already have.
For an accounting team, that becomes a rule: no change to payment details and no unusual transfer is approved on a call or video alone, however familiar the face. Confirm by calling a number from your own records. The Arup employee only learned the truth after checking with the company’s U.K. head office, which is the same idea after the fact.
The limit: a rule only works if someone is allowed to slow a senior person down. If staff are afraid to question the CFO, the best procedure on paper won’t help.
Next step: Spot the Scam Before You Pay: AI Payment-Fraud Checks (first 2 lessons free) teaches screening payment emails with a free AI tool, then verifying by phone.
What this means for small business owners
For a small business, the deepfake risk is that someone who sounds like you, or like your supplier, asks an employee or a customer to act. The FBI says criminals create short clips that mimic a loved one’s voice, and it advises people to limit their online photos and voice recordings, set social media accounts to private and accept followers only from people they know. A business owner who posts long videos of themselves is handing over the raw material, which is a trade-off, not a reason to disappear.
Practical steps that follow from the FBI’s advice: agree a spoken code word with whoever approves payments, so a voice alone never authorizes anything; and make a call-back to a known number the standard answer to any urgent request that arrives by voice or video.
The limit: none of this makes you safe from a fake. It makes a fake costly to act on.
Next step: AI for Small Business (first 2 lessons free) walks through using AI in a small business, which is where you decide which of these checks fit your own workflows.
What this means for customer support teams
Customer support is where voice deepfakes meet accounts. The FBI says criminals use AI-generated audio of individuals to impersonate them and gain access to bank accounts, and a phone line is where that attempt arrives. Treat a caller who sounds exactly like the account holder as one signal, not proof.
The fix is in the process: treat the voice as one signal and never the only one. Ask for something the caller could not know from a public video, use a call-back to the number on file for sensitive changes, and make it easy for an agent to say “I’m going to verify this another way” without penalty. A script that tells agents to stop and verify is worth more than any detector.
The limit: stricter verification adds friction for real customers. Choose which actions get the extra step, such as payment changes and credential resets, instead of adding it everywhere.
Next step: AI-Powered Customer Support Systems (first 2 lessons free) covers building AI into a support operation, a good place to design the verification step alongside it.
What this means for teachers
For a teacher, the deepfake question can arrive as a student problem: a fake image or video of a classmate. The law is clearest for one case. Under the TAKE IT DOWN Act, publishing an intimate digital forgery of an identifiable person is unlawful, with a longer maximum penalty where the victim is a minor, and covered platforms must remove a valid request within 48 hours.
What a teacher can do is limited and should stay inside school policy. Your school’s safeguarding lead and the platform’s own removal form are the first calls, and the person affected should not be asked to look at the material again. Teachers also have a smaller, constant job: teaching students to ask where a clip came from before they share it.
The limit: this page cannot tell you what your school, state or country requires. Ask your administration.
Next step: Teaching with AI (first 2 lessons free) is the starting point for using AI in the classroom, and Trust, but Verify: Fact-Checking AI (first 2 lessons free) builds the verify habit you can pass on to students.
Common misconceptions
“A deepfake always looks fake.”
Some do, and the FBI lists the flaws to look for: distorted hands or feet, unrealistic teeth or eyes, irregular faces, inaccurate shadows and unnatural movement. But in the University of Florida study people got videos right about two-thirds of the time, which means about one in three was misjudged. Full Fact says a deepfake may lack the usual signs and a genuine recording may show them.
“A detector will tell me.”
Full Fact says not to rely on AI detectors alone. In the Deepfake-Eval-2024 benchmark none of the commercial models reached 90% accuracy, and Full Fact found likely false positives in Google’s public tool. Use a detector as one input, after the source check.
“If someone says it’s a deepfake, it is.”
Calling a real video a deepfake is also a mistake, and it has happened. On October 19, 2026, the Oregon Journalism Project published a story that called a video of an upside-down American flag a “deepfake,” then unpublished it on October 20 after new images showed the flag had in fact been upside down. The lesson is the same in both directions: trace the clip before you label it.
“Deepfakes only target celebrities and politicians.”
The Arup case involved an employee at the firm’s Hong Kong office, and the Senate-race videos borrowed a real private person’s name and profile picture. Wikipedia says a voice can be cloned from only seconds of speech, which is why the FBI advises limiting the photos and recordings you post.
See also
Courses, related terms and posts about deepfakes and verification.
Courses on verifying what you see (first 2 lessons free)
- Trust, but Verify: Fact-Checking AI: a short verify habit for AI answers and claims
- Spot the Scam Before You Pay: AI Payment-Fraud Checks: screen payment emails, then verify by phone
- AI for Fraud Detection & Forensic Accounting: AI-assisted fraud procedures
- Falsely Flagged by an AI Detector: what to do when a detector gets it wrong
- AI for Journalists & Media Professionals: research and fact-checking with AI
- AI Fundamentals: how AI works, in plain language
Related terms
- AI voice cloning: the audio kind of deepfake
- SynthID: Google’s invisible watermark
- C2PA: Content Credentials
- AI detector: why detection tools miss
- AI slop: low-quality AI content in your feed
- AI hallucination: when a chatbot states something false with confidence
Skills
- Phishing Email Detector: analyze a suspicious email
- Critical Thinking Navigator: evaluate information with proven frameworks
Related posts
- Is This Video Real? A 4-Step Check Before You Forward It: the step-by-step method
- The AI Voice-Cloning Scam: How It Works and How to Stop It: protecting your family
- How to Tell If an Image Is AI-Generated in 2026: the photo version
- Do AI Detectors Actually Work in 2026? The Honest Answer: how reliable the tools are
The bottom line
A deepfake is AI-made or AI-edited media that borrows a real person’s face or voice, and the studies and tests cited here suggest that neither your eyes nor current detectors settle it. What works is older than the technology: find out where the clip came from, and don’t act on a face or a voice alone when money or a reputation is at stake. FindSkill.ai’s verify-first courses are built around that habit.
Frequently asked questions
What is a deepfake? A deepfake is an image, video or audio clip that has been edited or generated using artificial intelligence. The name blends “deep learning” and “fake” and goes back to a Reddit user called deepfakes who shared face-swapped videos in late 2017. Today the word covers face swaps, lip-synced videos, cloned voices and fake live video calls.
How do deepfakes work? Many deepfakes use a neural network called an autoencoder: an encoder compresses a face into a compact description, and a decoder trained on a target person rebuilds a face from it, which makes a face swap possible. Other methods use two networks that train against each other until the fakes look real, or edit a video so a person seems to mouth a different audio track. Voice clones can be made from only seconds of speech, according to Wikipedia.
Can AI detectors spot deepfakes? Not reliably. In the Deepfake-Eval-2024 benchmark, which used deepfakes that circulated online in 2024, the best commercial video detector scored 0.78 accuracy and none of the commercial models reached 90%. Full Fact advises against relying on detectors alone and reported that Google’s public SynthID Detector flagged five apparently genuine images, which Google then updated the tool to treat as unsure.
Are deepfakes illegal? It depends on what the deepfake shows and what it is used for. The U.S. TAKE IT DOWN Act, signed on May 19, 2025, makes it unlawful to knowingly publish an intimate depiction or a “digital forgery” of an identifiable person, and it requires covered platforms to remove a valid request within 48 hours. Other deepfakes fall under other laws or none, depending on the place and the use, and this page is not legal advice.
How can I protect myself from a deepfake scam? The FBI’s advice is to agree a secret word or phrase with your family, limit the photos and voice recordings you post, and verify a caller by hanging up and contacting the organization through a number you already have. Don’t send money, gift cards or cryptocurrency to someone you have only met online or by phone. For a business, the same idea becomes a call-back to a known number before any payment changes.
What should I do if someone makes a deepfake of me? If it is an intimate image, the TAKE IT DOWN Act lets you ask a covered platform to remove it, and the platform must act within 48 hours of a valid request. For other kinds of deepfake, start with the platform’s own reporting tools and keep the original posts and links. For legal options, talk to a lawyer where you live.
Sources
- Wikipedia, “Deepfake,” accessed 2026-10-28. https://en.wikipedia.org/wiki/Deepfake
- FBI Internet Crime Complaint Center, “Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (I-120324-PSA),” accessed 2026-10-28. https://www.ic3.gov/psa/2024/psa241203
- Fortune, “Arup revealed as victim of $25 million deepfake scam involving Hong Kong employee,” accessed 2026-10-28. https://fortune.com/europe/2024/05/17/arup-deepfake-fraud-scam-victim-hong-kong-25-million-cfo
- U.S. Government Publishing Office, “TAKE IT DOWN Act, Public Law 119-12,” accessed 2026-10-28. https://www.govinfo.gov/content/pkg/PLAW-119publ12/html/PLAW-119publ12.htm
- arXiv, “Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024,” accessed 2026-10-28. https://arxiv.org/abs/2503.02857
- University of Florida, “Machines spot deepfake pictures better than humans, but people outperform AI in detecting deepfake videos,” accessed 2026-10-28. https://news.ufl.edu/2026/02/deepfake-detection/
- Full Fact, “How to spot deepfake videos and AI audio,” accessed 2026-10-28. https://fullfact.org/toolkit/how-to-spot-deepfakes/
- Full Fact, “Google updates AI detector after Full Fact investigation finds it flagging apparently real images,” accessed 2026-10-28. https://fullfact.org/toolkit/google-synthid-detector-flagging-apparently-real-images/
- Google, “SynthID Detector for everyone,” accessed 2026-10-28. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/
- Foundation for Defense of Democracies, “Russian Operatives Plant Deepfake Targeting U.S. Midterm Elections,” accessed 2026-10-28. https://www.fdd.org/analysis/russian-operatives-plant-deepfake-targeting-u-s-midterm-elections/
- PolitiFact, “Russian disinformation group targets James Talarico with baseless claim in AI-generated video,” accessed 2026-10-28. https://politifact.com/factchecks/2026/oct/22/tweets/russian-disinformation-group-targets-james-talarico-with-baseless-claim-in-ai-generated-video/
- Oregon Journalism Project, “Editor’s note, October 20,” accessed 2026-10-28. https://www.oregonjournalismproject.org/editors-note-october-20
