In 2017 a Reddit user named 'deepfakes' posted a tool that swapped celebrities' faces into pornographic videos. Seven years later the technology that powered that stunt, Generative Adversarial Networks (GANs), is a standard fixture in Hollywood post-production, game development, and music licensing. The same architecture that reduces the cost of a visual-effects shot from thousands of dollars to a GPU cycle also produces fabricated media that no major platform can reliably spot.
The cost of that ambiguity is unevenly distributed. Studios gain faster pipelines, cheaper personalization, and new creative options. Actors, musicians, and voice performers face the prospect of their likenesses being used without consent, for a single session fee, or after death. Policymakers in Beijing and Brussels have passed rules requiring labels on AI-generated material. Washington has not, and no federal deepfake law exists as of March 2024.
Generative AI creates genuine economic value in media. It also destroys trust. The technical arms race between generation and detection is unlikely to produce a winner.
How GANs and Diffusion Models Changed the Economics of Production
GANs train a generator network against a discriminator network. The generator produces images. The discriminator judges whether they are real. Over thousands of rounds the generator becomes good enough that the discriminator cannot tell the difference. That competition produces the high-resolution fabricated faces, voices, and motion seen in contemporary deepfakes.
Film and television
Visual-effects houses use GANs and diffusion models for de-aging, background replacement, and digital makeup. A studio can now generate a realistic crowd scene entirely from artificial actors, avoiding the cost of hiring hundreds of extras and the later cost of clearing their likenesses for streaming. Disney's Industrial Light & Magic and multiple independent houses have deployed these tools in production. The cost per shot for certain classes of VFX work has fallen by an estimated 40 to 60 percent since 2020, though that figure is not audited.
Music and gaming
Music labels license AI tools that separate vocals from instrumentals, generate backing tracks in a specific artist's style, or produce fabricated voice performances for interactive characters. Game studios use procedural generation driven by large language models to create dialogue trees and ambient audio without recording hundreds of hours of session time. The result is faster iteration, lower variable cost per unit of output, and a broader palette of sonic textures.
The Deepfake Incidents That Reshaped the Conversation
The term emerged on Reddit in 2017. By 2019, a GAN-generated video of a politician speaking words they never said had circulated in Gabon and contributed to an attempted coup, according to journalists who covered the event. In 2022, a fake video of Ukrainian president Volodymyr Zelenskyy appeared on a hacked news site; the Ukrainian government took the site down within hours.
Financial harm arrives faster. In 2023, a finance worker in Hong Kong transferred USD 25 million after attending a video call in which every participant, including the company's CFO, was a deepfake. The worker recognized the CFO's voice and mannerisms. The fraud was not detected until the worker checked with the real CFO.
Reputational harm is the most common category. Fabricated non-consensual pornography, nearly all of it targeting women, accounts for the majority of deepfake material online according to several academic surveys. The social platforms that host it rely on detection tools that have false-positive rates high enough that automated removal often flags legitimate journalism or satire.
Detection Technology Runs Behind Generation
Google, Meta, and Microsoft have each released detection tools. Google's Photo DNA and SynthID, Meta's internal classifiers, and Microsoft's Video Authenticator all work on the same principle: they look for artifacts in the fabricated image that the generator did not eliminate. The problem is that each new generation of GANs and diffusion models eliminates those artifacts.
A detection model trained on deepfakes from 2020 will miss most deepfakes from 2023 because the generator learned to suppress the specific statistical signals the detector used. The arms race favors the attacker. An academic paper published in 2023 tested seven leading detection models against a state-of-the-art generator and found that accuracy fell below 50 percent for the newest videos. The same paper noted that adding mild compression, the kind social media platforms apply automatically, dropped detection to near chance.
No major platform currently claims to catch all fabricated material. Meta labels AI-generated images that its users upload, but only when the image contains a C2PA provenance signal, a cryptographic watermark that most generators do not embed. Google's SynthID can watermark images at creation time, but the watermark is fragile enough that a screenshot or a crop removes it.
Regulation: China, Europe, and the Gap in the United States
China introduced a regulation in 2023 requiring that all deepfake material carry a clear label. The regulation, issued by the Cyberspace Administration of China, applies to any fabricated media that could mislead the public. Platforms that distribute unlabeled deepfakes face fines. Enforcement has been uneven, but the rule establishes a legal obligation that does not exist in most other jurisdictions.
The European Union's AI Act, which passed in 2023 and enters into force in phases, includes a transparency obligation for AI-generated material. Any system that produces fabricated text, image, audio, or video must disclose that the output is AI-generated. The obligation applies to deployers, not developers, and carries fines of up to 3 percent of global annual turnover. The AI Act does not ban deepfakes. It requires labeling.
The United States has no federal deepfake law as of March 2024. Several bills have been introduced in Congress, including the DEEPFAKES Accountability Act and the Protecting Consumers from Deceptive AI Act, but none has passed both chambers. State-level laws exist in California, Texas, and a handful of other states, mostly focused on election deepfakes and non-consensual pornography. The patchwork means a deepfake that is illegal in California can be perfectly legal in Florida.
Social Platforms Police What They Can See
YouTube, owned by Google, requires creators to label AI-generated material that could be mistaken for real footage. Failure to label can result in a takedown or demonetization. The policy, announced in November 2023, covers fabricated realistic material but exempts obviously fantastical material such as animation. Enforcement relies on the creator's honesty and on automated scans that flag unlabeled uploads.
Meta's policy for Facebook and Instagram treats AI-generated material on a sliding scale. Fabricated media that is clearly parody or satire is allowed. Media that realistically depicts a person saying something they did not say is removed if it violates the platform's bullying or harassment policies. Meta also adds an 'AI info' label to material that its automated systems identify as fabricated, but the company acknowledges that the label is applied inconsistently.
TikTok introduced a similar labeling requirement in 2023. The platform's systems can detect uploads generated by a subset of popular AI tools. The detection rate for OpenAI's DALL-E 3, for example, is higher than for open-source models because DALL-E 3 embeds a watermark that TikTok's systems recognize. Open-source models rarely embed any watermark.
Actors, Musicians, and the Question of a Dead Performer's Likeness
The Screen Actors Guild-American Federation of Television and Radio Artists (SAG-AFTRA) negotiated during its 2023 strike to establish guardrails on AI use of performers' likenesses. The contract that ended the strike in November 2023 requires studios to obtain consent from performers before using AI to generate a digital replica, and to negotiate separate compensation for that use. The contract also restricts the use of AI to create fabricated performances using a performer's image without their participation.
The resurrection of deceased performers is a separate and less settled question. A studio can negotiate with an estate for the use of a dead actor's likeness. The legal basis is the right of publicity, which exists in about half of U.S. states and varies widely. California's right of publicity extends 70 years after death. New York's extends 40 years. Some states have no postmortem right. A studio planning to use a deceased performer's image must therefore navigate a state-by-state legal map or restrict distribution to states where the estate holds rights.
Music labels have begun licensing AI-generated vocals that mimic deceased artists. The practice is legal when the estate consents and illegal when it does not. The ethical debate centers on whether a performer can consent to uses they never imagined. SAG-AFTRA's position, stated in its 2023 contract, is that consent must be specific and revocable. No major label has yet faced a court test of a consent signed before AI-generated media became commercially viable.
Key facts
- Term 'deepfake' coined: 2017, on Reddit
- Core technology: Generative Adversarial Networks (GANs)
- China deepfake labeling regulation: Introduced 2023
- EU AI Act transparency provision: Passed 2023, phased enforcement
- U.S. federal deepfake law: None as of March 2024
- SAG-AFTRA AI likeness protections: Negotiated in 2023 strike contract
Major platform policies on AI-generated material (as of March 2024)
| Platform | Labeling requirement | Detection method | Enforcement |
|---|---|---|---|
| YouTube | Creator must label realistic AI material | Automated scans + creator self-reporting | Demonetization or takedown for failure to label |
| Meta (Facebook/Instagram) | AI info label applied automatically when detected | C2PA signal + internal classifiers | Removal if material violates harassment/bullying policy |
| TikTok | Label required for AI-generated realistic material | Recognizes watermarks from some tools (e.g., DALL-E 3) | Takedown for unlabeled fabricated material |
Frequently asked questions
Can I legally use AI to make a video of a living celebrity saying something?
Not without consent. The celebrity holds a right of publicity in most U.S. states and in many other countries. Using their name, image, or voice for commercial purposes without permission is usually illegal. Non-commercial uses, such as parody, may be protected by the First Amendment in the U.S., but the line is not firmly drawn by courts in the AI context.
Do deepfake detection tools work?
They work for some classes of deepfakes and fail for others. Detection models trained on older GAN-generated material often miss newer diffusion-model material. Compression, cropping, and screenshotting further reduce accuracy. No major platform or vendor claims perfect detection.
What happens if a studio uses AI to recreate a dead actor without the estate's permission?
The estate can sue for violation of the right of publicity, where that right exists. California's law protects deceased performers for 70 years. Other states provide shorter or no protection. A studio may therefore choose to distribute the material only in states where the estate has no legal standing, though major studios rarely take that risk.
