Synthetic AI media — image, video, or audio — that convincingly shows a real person doing or saying something they never did.
Last verified · 2026-10-15 · by Moe Ameen
A deepfake is synthetic media generated or altered by AI to depict a specific real person doing or saying something they never did. The word blends "deep learning" and "fake," and the category spans the full family of synthetic likeness: face swaps, fully AI-generated video of a real face, AI-altered photos, and cloned speech. The common thread is that a real identity — a face, a voice, or both — is reproduced closely enough to be mistaken for the genuine article. Like most of the synthetic-media vocabulary, the term is widely used with a criminal connotation because it went mainstream through abuse, but the artifact itself is neutral. What separates a crime from a production tool is consent, ownership, and disclosure, not the technology.
Under the hood, the capability traces to the generative adversarial network (GAN), introduced by Ian Goodfellow and colleagues in 2014, in which a generator network produces fakes and a discriminator network tries to catch them, each improving by competing with the other until the output is hard to distinguish from real. Diffusion models later pushed quality and ease much further. The practical consequence for 2026 is how little source material a convincing fake now needs: a handful of public photos or seconds of audio is enough, which means any creator with a public face or voice is already a viable target. The [deepfake voice](/glossary/deepfake-voice) is simply the audio-only branch of this same family.
For creators the stakes are concrete, because a face and a voice are the brand. The abuse end runs from non-consensual intimate imagery to fabricated endorsements in [AI UGC ads](/glossary/ai-slop) that stack an unauthorized likeness onto a deceptive claim. The legitimate end is large and growing: consented, disclosed synthetic video of your own likeness used to dub content, run a recurring on-brand presenter, or scale a persona. The defensive layer built specifically for the abuse case is [likeness detection](/glossary/likeness-detection) — platform tooling that scans uploads for an enrolled person's face or voice — which sits alongside watermarking and provenance credentials in the wider detection stack.
The underlying math is older than the word. GANs arrived in 2014; by 2017 researchers were generating sharp synthetic faces. The term "deepfake" itself was coined in late 2017 by a Reddit user operating under that handle, who ran a subreddit for face-swapped pornography made with open-source tools; Reddit banned the community in February 2018 for involuntary pornography. That origin shaped the early reality — a widely cited 2019 Deeptrace report found roughly 96% of the deepfake videos it catalogued were pornographic, nearly all depicting women non-consensually — a figure worth treating as historical framing for how skewed early use was, not as a current measurement, since it predates modern diffusion models.
The 2020s turned a research curiosity into a mass phenomenon. Consumer diffusion and video models made convincing fakes cheap, and the flashpoints piled up: sexually explicit AI images of Taylor Swift spread across X in late January 2024, prompting renewed calls for legislation; OpenAI's Sora app launched a "Cameos" feature on September 30, 2025 that let a verified user drop their own face and voice into generated clips — but Sora 2 also proved able to generate OTHER real people, including Bryan Cranston and deceased figures, without consent. Cranston brought his case to SAG-AFTRA, and Robin Williams's daughter Zelda Williams separately and publicly asked people to stop sending her AI videos of her late father. By October 20, 2025, OpenAI, SAG-AFTRA, Cranston, and major talent agencies issued a joint statement in which OpenAI committed to stronger guardrails against unauthorized likeness and voice replication. In October 2026 a verified Instagram "fan" account posted an AI-generated selfie of actress Mia Goth that many outlets and viewers took as genuine before Meta removed it for impersonation — a case that captured the platforms' default posture of labeling AI-modified content rather than banning depictions of real people outright.
Regulation moved to catch up. In the US, the TAKE IT DOWN Act was signed on May 19, 2025, criminalizing the knowing publication of non-consensual intimate imagery including AI deepfakes, with a notice-and-takedown duty (removal within 48 hours of a valid report) that covered platforms had to meet by May 19, 2026, enforced by the FTC. In the EU, the AI Act's Article 50 transparency rules take effect August 2, 2026, requiring providers of generative systems to mark synthetic output in a machine-readable form and deployers to disclose deepfakes, under a definition broad enough to cover any AI content resembling real people or events.
| Platform | Behavior |
|---|---|
| Generative-video apps (Sora, etc.) | The newest and most contested surface — a verified user can drop their own face and voice into a generated clip via Sora's "Cameos" feature, consent captured at the generation step itself rather than enforced after upload. But the same model could also be prompted to generate OTHER real people without their consent, which is what drove OpenAI's October 2025 guardrail tightening after pressure from SAG-AFTRA and Bryan Cranston. |
| Meta (Instagram / Facebook) | Does not ban AI-generated depictions of real people as a rule; it labels AI-modified content and removes accounts that impersonate a public figure or run fraudulently. The Mia Goth case showed the gap: an AI image on a verified "fan" account spread widely before removal. Realistic synthetic media must still be disclosed under Meta's AI-content rules. |
| YouTube | Requires creators to label realistic AI-generated or synthetic content via an "altered or synthetic content" disclosure, and separately runs likeness detection that lets enrolled creators find and request removal of AI content using their face. Undisclosed synthetic media that could mislead viewers risks removal or demonetization. |
| TikTok | Labels AI-generated content, reads C2PA provenance signals to auto-tag some uploads, and for ads requires advertisers using a voice clone or digital likeness to submit consent documentation. An undisclosed clone of a real person falls under its impersonation and manipulated-media policies. |
| EU AI Act (Article 50) | Not a platform but the regulatory driver. From August 2, 2026, anyone deploying a deepfake must disclose it, and providers of the generative systems must mark output in a machine-readable, detectable format — pushing the whole industry toward provenance signals that detection tools can read. |
The reflex is to file every deepfake under threat, but for a working creator the more useful question is provenance: not "is this synthetic?" but "whose likeness is it, who consented, and is it disclosed?" That reframing is the difference between a defensive crouch and a workflow, because the exact capability behind the celebrity-likeness scandals is also what lets you dub yourself into another language, run a consistent on-brand presenter across a hundred posts, or scale a persona — provided the face and voice are yours or licensed and you label the output. Kompozy sits deliberately in that lane: its persona video is built on HeyGen's consented avatar-and-voice model, tied to an [AI Influencer persona](/glossary/persona-shorts) you set up and control, and published to your own channels, so the synthetic likeness you ship has a clean provenance chain rather than a stolen one. The disclosure decision then lives in one governed place — a per-platform-account toggle in Settings → Publishing — so the honest AI-generated label ships with every post from that account instead of needing to be remembered post by post. Two habits make you resilient on both sides of the line: assume your public face and voice are already fakeable and know the [likeness-detection](/glossary/likeness-detection) tools that let you find misuse, and keep your own synthetic-likeness use consented and disclosed so you build trust with platforms and audiences instead of tripping their manipulated-media filters. Owning your synthetic identity is now something you do on purpose — or something that gets done to you.
A deepfake is synthetic media — an image, video, or audio clip — generated or altered by AI to depict a specific real person doing or saying something they never did. The name blends "deep learning" and "fake." The category covers face swaps, fully AI-generated video of a real face, AI-edited photos, and cloned voices. The term is neutral: the same technology powers both non-consensual abuse and consented, disclosed creator tools.
The underlying method, the generative adversarial network, was introduced by Ian Goodfellow and colleagues in 2014. The word itself was coined in late 2017 by a Reddit user with the handle "deepfakes," who ran a subreddit for face-swapped pornography; Reddit banned the community in February 2018. The term now covers a far broader range of AI-generated likeness than those original face-swap videos.
The technology is legal and has legitimate, consented uses; specific abuses break the law. In the US the TAKE IT DOWN Act, signed May 19, 2025, criminalizes knowingly publishing non-consensual intimate imagery including AI deepfakes and requires covered platforms to remove it within 48 hours of a valid report. The EU AI Act requires anyone deploying a deepfake to disclose it from August 2, 2026. Using a consented or licensed likeness with disclosure to make your own content is legitimate.
Usually not reliably. In blind tests humans perform only modestly better than chance at flagging synthetic media, and quality keeps rising. Detection is more dependable when it reads a planted signal — an invisible watermark or a C2PA provenance credential — than when it relies on spotting visual artifacts, which modern models increasingly avoid.
Yes, with the right provenance. Synthetic video or audio of your own likeness, or a licensed one, used with disclosure is a production tool, not an impersonation — this is how consented avatar and persona video works. You still need to apply each platform's AI-generated label where its policy requires it; consent covers whose likeness you used, disclosure covers telling viewers it is AI.
Enroll in platform likeness-detection tools where available (YouTube ships the clearest version) so you can find and request removal of AI content using your face, keep consent records for any synthetic likeness you create, and assume a handful of your public photos or seconds of audio is enough source material for a fake. The durable posture is to own and disclose your own synthetic identity rather than leave it to be cloned without you.