• Communications: misinformation and how to check it

What is a deepfake?

Written with AI (How we create content)Target level: Grade 12Last updated: Report an error
Quick answer

A deepfake is a video, audio clip or image made or altered with artificial intelligence so that a real person seems to say or do something they never said or did. The name, from deep learning and fake, comes from a Reddit user in 2017. Deepfakes can deceive, defraud and humiliate, and they also let liars dismiss real recordings as fake.

The topic explained

A face, a voice, a fake

A deepfake is synthetic media: a video, voice recording or picture produced or changed by artificial intelligence so that it shows a real person saying or doing what they never said or did. The word joins deep learning, the AI technique behind it, with fake. It comes from a Reddit user called “deepfakes,” who in late 2017 began sharing face-swapped videos, many of them pornographic.

That origin is not a footnote. In October 2019 a report by the Dutch company Deeptrace estimated that 96% of the deepfakes it found online were pornographic. Such videos place a real person's face into sexual images they never agreed to. The most common harm of deepfakes, in other words, is not a fake speech that changes an election but the humiliation of people who never consented. Today related tools also clone voices from short recordings and produce faces of people who do not exist.

How they are made, in brief

Early face swaps used autoencoders: a network learns to compress thousands of pictures of a face into its essential features and to rebuild the face from them. Train one rebuilder on person A and another on person B, feed B's expressions into A's rebuilder, and A's face appears with B's movements.

A second technique, the generative adversarial network (GAN), was introduced by Ian Goodfellow and colleagues at the University of Montreal in 2014. Two networks compete: a generator produces fake images, and a discriminator tries to tell them from real ones. Each round, the discriminator's verdict teaches the generator to fool it better, until the fakes are hard to tell apart. The design explains why detection is a race: any flaw a detector learns to spot is a flaw the next generator learns to avoid.

Diagram of an adversarial network: the generator sends a fake image to the discriminator, and the discriminator sends back its verdict, real or fake, so that both improve.GeneratorDiscriminatorfake imagereal or fake?
A sketch of a generative adversarial network, Goodfellow and colleagues, 2014.

Real cases

In 2019 the head of a British energy company received a phone call from what sounded exactly like the chief executive of its parent company, asking for an urgent payment. He transferred 220,000 euros to a Hungarian bank account; the voice was an audio deepfake. In 2021 a Belgian visual-effects artist, Chris Umé, posted TikTok clips of a Tom Cruise lookalike whose face had been replaced; they drew tens of millions of views. He presented them openly as a demonstration, which shows that the technology itself is not the problem: deception and consent are.

Timeline of deepfakes: 2014 adversarial networks; 2017 the term, from a Reddit user; 2019 a voice-fraud of 220,000 euros; 2019 a report finds 96% of deepfakes online pornographic; 2021 fake Tom Cruise clips on TikTok.2014Adversarial networks2017The term, on Reddit2019Voice fraud: €220,0002019Report: 96% pornographic2021Tom Cruise clips on TikTok
From a lab technique to a tool anyone can use, in less than a decade.

Checking, and the liar's dividend

Guides list visual clues: unnatural blinking, lighting that does not match, blurred edges around the face, lips out of step with the words. They help, but they are a moving target, and even the best automated detectors make many mistakes: the winning model of a large detection challenge was right only 65% of the time on 4,000 videos. Safer questions are about the source: who posted it first, does any reliable outlet report it, and can I check directly, for example by calling the person back on a number I already know.

The legal scholars Robert Chesney and Danielle Citron warned in 2018 of a second danger, the liar's dividend: once everyone knows fakes exist, a person caught on a genuine recording can claim it is a deepfake. The common mistake is to conclude that nothing can be believed. The better conclusion is that trust must move from the picture to its source.

Table of four questions for checking a suspect video: where did it come from, who posted it first and when; who else reports it, reliable news outlets; do face and voice match, lips, lighting and edges; can I verify directly, call back or ask in person.QuestionWhat to checkWhere did it come from?Who posted it first, whenWho else reports it?Reliable news outletsDo face and voice match?Lips, lighting, edgesCan I verify directly?Call back, ask in person
Visual clues help, but the source and direct checks matter more.
Free printable worksheet

Practice: deepfakes

For 12th grade communications students. Every office, person, call and video here is made up. Answer from what is printed on the page, then check yourself against the answers.

Free, no sign-up. The answer key prints on the second page. Regenerating a new worksheet on the same topic needs a free account.

Free printable worksheetlessoons.com
Practice: deepfakesName:
  1. Where does the word deepfake come from, and why does the page say its origin is not a footnote?

  2. Fill in the missing message in the diagram, and explain why the two networks make each other better.

    Diagram of an adversarial network: the generator sends a fake image to the discriminator, and the discriminator sends back its verdict, real or fake, so that both improve.GeneratorDiscriminatorfake image?
    A sketch of a generative adversarial network, Goodfellow and colleagues, 2014.
  3. Fill in the missing event on the timeline.

    Timeline of deepfakes: 2014 adversarial networks; 2017 the term, from a Reddit user; 2019 a voice-fraud of 220,000 euros; 2019 a report finds 96% of deepfakes online pornographic; 2021 fake Tom Cruise clips on TikTok.2014Adversarial networks2017?2019Voice fraud: €220,0002019Report: 96% pornographic2021Tom Cruise clips on TikTok
    From a lab technique to a tool anyone can use, in less than a decade.
  4. A detector is right 65% of the time. On 4,000 videos, how many does it get right and how many wrong? What does this mean for relying on detectors?

  5. In a made-up case, a bookkeeper gets a voice call from her manager asking her to transfer money urgently to a new account. What should she do, and why?

  6. Fill in the missing question in the table.

    Table of four questions for checking a suspect video: where did it come from, who posted it first and when; who else reports it, reliable news outlets; do face and voice match, lips, lighting and edges; can I verify directly, call back or ask in person.QuestionWhat to checkWhere did it come from?Who posted it first, when?Reliable news outletsDo face and voice match?Lips, lighting, edgesCan I verify directly?Call back, ask in person
    Visual clues help, but the source and direct checks matter more.
  7. What is the liar's dividend, and who named it?

  8. A student says: “Since anything can be faked, I won't believe any video again.” Write a short answer.

More topics and worksheets at lessoons.com© 2026 Lessons

Common questions

How can you tell if a video is a deepfake?

Look for clues such as odd blinking, mismatched lighting, blurred edges and lips out of step with the words, but do not rely on them alone. Ask who posted it first, whether reliable outlets report it and whether you can check directly with the person.

Are deepfakes illegal?

It depends on the country and the use. Making a deepfake as clearly labeled art or satire is usually allowed; using one for fraud, impersonation or sexual images without consent is a crime or a civil wrong in many places, and new laws keep being added.

What is the difference between a deepfake and a cheapfake?

A deepfake is made with AI. A cheapfake is misleading media made with simple tools: a real video slowed down, cut or given a false caption. Both mislead; the cheapfake needs no technology.

Who invented deepfakes?

No single person. The word comes from a Reddit user in 2017, and the methods build on deep learning research, including the generative adversarial networks introduced by Ian Goodfellow and colleagues in 2014.

What is the liar's dividend?

A term from Robert Chesney and Danielle Citron (2018): because people know convincing fakes exist, someone caught on a real recording can falsely claim it is a deepfake and escape responsibility.

Still have a question? Ask Lessi

Any question about this subject. Lessi answers right here, step by step.

Try asking

Asking needs a free account; your question is kept while you sign up. Answers are written by AI and can contain mistakes. Lessi doesn’t give medical or legal advice.

Learn it with a teacher

Learn this topic live with Lessi

Lessi, the AI teacher, explains it out loud on a board, with short quizzes, at your pace. The first 60 minutes are free.