Skip to content

What Is a Deepfake? How Fakes Are Made and Spotted

A person seen from behind watching a laptop video call that shows a glowing silhouette in a dim home office


A person seen from behind watching a laptop video call that shows a glowing silhouette in a dim home office

A deepfake is an image, video, or audio clip made or changed with artificial intelligence so a real person appears to say or do something they never said or did.

The word blends deep learning and fake, and it entered wide use in 2017. Early examples were rough face swaps. Now phone apps and cloud tools build convincing fakes without a studio.

How a deepfake is made

Most deepfakes start with training data: photos, video, and audio of one person. The classic method is a generative adversarial network, or GAN. One network makes fakes, a second judges them, and Norton explains how the two refine each other over repeated rounds.

Newer tools use diffusion models, which refine noise step by step into an image or video from a text prompt or a reference photo. The Midjourney guide covers an image generator of this kind, and the Sora guide covers a video generator in the same class. Voice cloning learns pitch, rhythm, and phrasing from samples, then sometimes reads new words in that voice from seconds of audio. The ElevenLabs guide covers voice cloning, and the AI voice changer guide covers tools that change a voice as you speak.


A monitor showing a grid of face sketches that progress from rough outlines to a finished portrait, with a keyboard and notebook on the desk

Where deepfakes do real work

Studios have de-aged actors, finished a performance after an actor died, and rebuilt a voice for narration. Examples include a de-aged Harrison Ford in Indiana Jones and the Dial of Destiny, a recreated voice for Anthony Bourdain in Roadrunner, and Peter Cushing in Rogue One.

Teachers use the same methods to stage a speech by a historical figure, and dubbing teams revoice films in other languages with matched lip movements. Voice cloning can restore a voice for someone who loses speech to illness or surgery. Consent and disclosure separate these uses from the harmful ones.


A film set with a professional camera on a tripod, studio lights, a chair, and a monitor showing a face

Where deepfakes cause harm

The costliest known case started with an email. In January 2024, a finance worker at engineering firm Arup in Hong Kong joined a video call with people who looked and sounded like senior staff, including the chief financial officer. Every other person on the call was fake.

The worker approved 15 transfers totaling HK$200 million, about $25 million, to accounts the criminals controlled. Police disclosed the case, and Arup confirmed that fake voices and images were used. Its systems were never breached. The attack went after trust, not software.

Elections bring a second harm. A fake clip of a candidate can travel further in an hour than a correction travels in a week. X and Facebook have banned malicious deepfakes, but enforcement arrives after the first wave of shares. Fake intimate images, made without consent, are used to harass and blackmail, and copies spread faster than takedown requests.


A smartphone showing an incoming call screen on a wooden desk next to a wallet and a notebook

What is deepfake detection, and why it lags behind

Detection tools look for what the eye misses: odd pixel patterns, inconsistent lighting, and lip movements that drift from the audio. The better tools combine several signals instead of trusting one. They help. They do not settle the question.

Here is the complication. The same contest that builds a deepfake also trains its detector. A GAN improves by fooling a judge, so stronger detectors teach the next generator to hide better.

A clean result from one tool means that tool found nothing. It does not prove the clip is real.

Provenance attaches a signed record of where a file came from and what was done to it. The C2PA FAQ describes Content Credentials as a signed, tamper-evident record of a file’s history. A screenshot or a re-save strips it. A credential says nothing about whether the scene is true, only that the record is unaltered since signing.


A hand holding a magnifying glass over a tablet showing a paused video frame of a man

What the law says in the US, the EU, and the UK

In the United States, the TAKE IT DOWN Act became federal law on 19 May 2025. It criminalizes knowingly publishing certain intimate images without consent, including realistic AI fakes the law calls digital forgeries. Penalties reach two years in prison for adults, and three years where the person depicted is a minor.

The Congressional Research Service summary sets out the offenses and platform duties. Covered platforms must remove a reported image within 48 hours of a valid request, enforced by the Federal Trade Commission. The Act covers intimate images only. It does not ban deepfakes in general.

The European Union took the labeling route. Article 50 of the EU AI Act requires machine-readable marks on AI-generated output, and publishers of deepfakes to clearly disclose that the content was artificially generated or manipulated. The rules have applied since 2 August. The European Commission FAQ on Article 50 explains the provider and deployer duties.

Sharing or threatening to share an intimate image without consent, including a fake one, is a criminal offense in England and Wales under the Online Safety Act 2023. No proof of intent to cause distress is required. The Data (Use and Access) Act 2025 added an offense for creating such an image without consent. No single UK law covers every deepfake; political fakes and fraud fall under privacy, data protection, defamation, and copyright law.

This section is general information, not legal advice.


A stack of paper documents, a gavel, and reading glasses on a desk next to a laptop showing a paused video

How to spot a deepfake before you share it

Start with the face and the voice. Look for blinking that looks wrong and expressions that do not match the words. Watch the edges where face meets neck for blur or slipping.

Listen for flat delivery and strange pronunciations. Lip movements that drift from the sound are another sign. Slow the clip down and watch the mouth. Cheap fakes fall apart under that look.

Then check the source. Ask who posted it first, and whether a trusted outlet or the person depicted has confirmed it. Search the claim in words. The ChatGPT Search guide covers checking a claim against live web sources.

For calls and messages, use a second channel: if someone asks for money or a code, hang up and call back on a trusted number. When in doubt, verify first and share nothing until you have.


A man at a cafe table holding a phone showing a video while a laptop shows a news page

Frequently asked questions about deepfakes

What is a deepfake in simple terms?

A deepfake is a photo, video, or audio clip changed with AI to show a real person saying or doing something they never did.

Are deepfakes illegal?

Not in general. The US TAKE IT DOWN Act covers AI fakes of intimate images, and fraud and harassment break the law in many places.

Can you detect a deepfake reliably?

No single check is reliable alone. Detectors, visual checks, and provenance records each help, and each can fail. Combine checks and verify the claim independently.

What is the difference between a deepfake and an AI-generated video?

A deepfake depicts a real person, place, or event falsely, so a viewer takes it as authentic. An AI-generated video can be entirely fictional.

What should I do if someone makes a deepfake of me?

Save links and screenshots, with dates, then report the content and ask for removal. For threats, blackmail, or fraud, report it to the police as well.

Leave a Comment