A deepfake is real audio or video that AI has altered to show a real person doing or saying something they did not. An AI-generated video is made from scratch by a model, usually from a text prompt, with no original footage underneath. In everyday use the two terms overlap, and the law and platform rules increasingly cover both, but the distinction matters because the way you check each one is different.
One scope note first: this site checks text, not video. Nothing here is a video detector, and we do not offer one. This guide is background reading on terms and verification.
Key takeaways
- Deepfake means manipulation of existing media (face swap, expression change, lip-sync, voice cloning). AI-generated video means synthesis from a prompt. All deepfakes are AI-made; not all AI video is a deepfake.
- Real footage exists behind a manipulated deepfake, so you can often compare it with an original. A fully synthetic clip has no original, so provenance and context carry more weight.
- Regulators and platforms mostly care about realistic content that could be taken as real, whichever way it was made.
- No visual tell or detection tool is conclusive, and the older tells go out of date as generators improve.
What is a deepfake?
The US Government Accountability Office defines a deepfake as “a video, photo, or audio recording that seems real but has been manipulated with artificial intelligence technologies.” Its explainer lists the main techniques: face swapping (one person’s face placed on another person’s body), facial or expression manipulation (one face imitating another’s expressions), lip-sync (mostly the mouth region is altered to match a new audio track) and voice cloning (a synthetic copy of a person’s voice that can say anything). That report dates from 2020, so it describes the underlying idea rather than today’s tools. See the GAO science and tech spotlight on deepfakes.
The key feature is a real starting point: an existing video, photo or recording of a real person, changed by software. The result usually aims to look like a particular person did or said something.
What is an AI-generated video?
An AI-generated video is produced by a generative model, typically a text-to-video or image-to-video system, from a prompt or a still image. No camera recorded the scene. OpenAI’s Sora, released publicly in December 2024, was one example of a text-to-video product; OpenAI’s help center now says Sora is discontinued. Other text-to-video tools remain in use, and the categories described here apply to them equally.
A synthetic clip can show an invented person, an imaginary place, or a realistic-looking person who resembles someone real. It becomes a deepfake in the ordinary sense of the word when it is built to pass as a specific real person or real event. A fantasy landscape or a cartoon is AI-generated but nobody would call it a deepfake.
How do deepfakes and AI-generated videos compare?
| Deepfake | Fully AI-generated video | |
|---|---|---|
| Starting point | Existing footage, photo or recording of a real person | A text prompt, a still image or both |
| Typical methods | Face swap, expression manipulation, lip-sync, voice cloning | Text-to-video or image-to-video generation |
| Is there an original? | Usually yes, which can be compared | No |
| Typical intent | To make a real person appear to say or do something | Anything from art and ads to deception |
| Can be both? | Yes, a generated clip of a real person’s likeness is often called a deepfake | Yes, if it imitates a real person or event |
| Strongest check | Find and compare with the source footage; look for corroboration | Check provenance data and where the clip first appeared |
Real cases mix these methods. A fake video call may combine a manipulated face, a cloned voice and a script, and the reporting rarely says which tool did what.
What do documented cases look like?
A fake executive call at Arup (2024). The engineering firm Arup confirmed to Fortune that “fake voices and images were used” in a fraud against its Hong Kong office. An employee joined a video call with what appeared to be the company’s CFO and colleagues, and transferred about HK$200 million (US$25.6 million) across 15 transactions. The employee became suspicious and checked with the UK head office, which exposed the fraud. Public reporting does not say which techniques produced the faces and voices. See Fortune’s report.
A fake surrender message from Ukraine’s president (2022). In March 2022 an altered video showed Volodymyr Zelensky apparently telling soldiers to lay down their arms. Meta removed it under its manipulated media policy, and Zelensky quickly posted a video to deny it. The clip also appeared after a Ukrainian news site’s ticker was hacked. See TechCrunch’s coverage and Euronews.
A cloned voice in a robocall (2024). A call imitating President Biden’s voice told New Hampshire voters to skip a primary. The US Federal Communications Commission ruled on 8 February 2024 that AI-generated voices in robocalls count as “artificial” under the Telephone Consumer Protection Act. This is an audio-only deepfake, which shows why video is not the only medium to consider. See The Record’s report.
We are not aware of a comparably well-documented case that we could verify for a fully text-to-video clip being used in a fraud, so we do not offer one.
How is each one usually checked?
Detection and verification are different jobs. Detection asks whether pixels or sound carry signs of manipulation. Verification asks whether the claim the media makes is true, and it often works without any software.
For a suspected deepfake, the original is the most useful clue. Search for the same event from other outlets, look for the unaltered footage, and ask the supposed speaker or their organization directly. The Arup case was exposed by asking the head office, not by analyzing pixels. Automated detectors look for artifacts the manipulation leaves behind; the GAO describes this as a “cat and mouse” contest in which better detectors lead to better fakes.
For a suspected fully generated clip, there is nothing to compare against, so the checks shift to provenance. Some generators attach visible watermarks or Content Credentials (C2PA) metadata, which can show where a file came from. C2PA says credentials can be separated from a file, and that tampering invalidates the signature. So a present credential is useful evidence and a missing one proves nothing. Context still matters: who posted it first, whether a camera-original exists, and whether any reputable outlet corroborates it.
For both, a detector’s output is a judgment, not a verdict. A 2025 research benchmark, Deepfake-Eval-2024, collected deepfakes that actually circulated online in 2024 and found that open-source detectors lost roughly 45% to 50% of their AUC (a standard score of how well a classifier separates real from fake) compared with older academic benchmarks. Commercial and fine-tuned models did better but did not reach the accuracy of human forensic analysts. The practical point is that tools tested on old datasets can fail on new fakes. Our guide to telling whether a video is AI-generated walks through a practical checking order, and how AI video detection works explains what such tools look for.
Be careful with visual tells like odd blinking, warped hands or garbled background text. They were true of some generators at some times, and they are not rules. A clip with none of them can still be fake, and a real one can show some through compression.
What do the law and platform policies say?
The rules are changing, so check the current text for your jurisdiction. These are the points we could verify.
- EU AI Act. Article 3(60) defines a deep fake as “AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.” That wording covers both manipulated and fully synthetic content. Article 50(4) requires deployers of such systems to disclose that the content was artificially generated or manipulated, with a lighter duty for evidently artistic or satirical work, and Article 50(2) requires providers to mark outputs in a machine-readable way. See Article 50 of the AI Act. The transparency duties are scheduled to apply from 2 August 2026, and a 2026 amendment package reportedly postpones the provider marking duty to 2 December 2026, though a law firm’s summary noted uncertainty over which systems that covers. Confirm the current position with an official source.
- United States. The TAKE IT DOWN Act, signed on 19 May 2025, criminalizes knowingly publishing intimate images of an identifiable person without consent, including AI-created “digital forgeries,” and requires covered platforms to remove such material within 48 hours of notice. Platforms had until 19 May 2026 to set up the process. See the Congressional Research Service summary. We did not verify general federal or state laws on political or commercial deepfakes, so we make no claim about them.
- YouTube. Creators must disclose realistic altered or synthetic content, for instance content that “makes a real person appear to say or do something they didn’t do,” “alters footage of a real event or place” or “generates a realistic scene that didn’t actually occur.” Unrealistic content and minor edits are exempt, and repeated non-disclosure can lead to labels, removal or loss of Partner Program access. See YouTube’s policy page.
Notice that the EU definition and YouTube’s rule both focus on realism and deception, not on the technique. In policy terms, the deepfake-versus-generated split matters less than whether a viewer could mistake the content for real.
A short checklist for telling which one you are looking at
- Is there a real person or event it claims to show? If so, look for the original footage and independent coverage.
- Does the clip start from a real recording? A familiar speech or interview with changed mouth movements or audio suggests manipulation of existing media.
- Is the whole scene improbable or unsourced? A clip with no origin and no camera-original deserves provenance checks first.
- Does it carry credentials or a watermark? Treat their presence as evidence and their absence as no information.
- Can someone confirm it out of band? Call the person back on a known number, as in the Arup case.
Where this site fits
This site does not check images, audio or video. Our AI-generated text detector labels passages of written text, and the methodology page explains what it can and cannot tell you. If a suspicious video comes with a transcript, script or caption you want to review, you can check that text, but the result says nothing about whether the video itself is real.
Frequently asked questions
Is every deepfake AI-generated?
Yes, in the sense that AI is involved, but “AI-generated” is not always the best description. A deepfake usually modifies real footage with AI, so only part of the final clip is machine-made. Fully AI-generated video is made entirely by a model. Many people use the words interchangeably, which is why the context of a source matters.
Can a text-to-video clip be a deepfake?
It can, if it is made to look like a specific real person or real event and could fool viewers. The EU AI Act’s definition covers “AI-generated or manipulated” content that resembles real persons or events, so it applies to both. A clip of an imaginary character or place is AI-generated but not usually called a deepfake.
Is a deepfake illegal?
It depends on what it shows and where. Some uses are specifically illegal, such as nonconsensual intimate deepfakes under the US TAKE IT DOWN Act, and AI-voiced robocalls under the FCC’s ruling. Many deepfakes are not covered by a dedicated law, though fraud, defamation and other existing laws can apply. This is general information, not legal advice.
Can software reliably detect deepfakes?
Not reliably. Benchmarks on fakes found in the wild show sharp drops from lab results, and the best tools still trail forensic experts. Detectors are one input, and you should combine them with provenance and source checks.