Blog · Analysis · May 2026

The Consent Layer for Synthetic People

Digital replicas turn voice and likeness into operational media. The core governance question is not only whether an artifact is labeled, but whether the person was allowed to refuse simulation.

The Person as Media

Synthetic media used to be discussed mainly as fake evidence: a politician saying something they never said, a forged image, a doctored video, a voice used in a scam. That frame still matters, but it is too narrow. The more important shift is that a person can now become reusable media.

A voice can be cloned. A face can be animated. A body can be placed in a scene. A deceased performer can be brought into an advertisement. A worker's recorded performance can become training material for future synthetic performance. A private person's image can be turned into sexual abuse material without a camera ever entering the room. A public figure can be made to endorse a product, discourage voters, or appear inside a personalized persuasion campaign.

The U.S. Copyright Office's July 2024 report on digital replicas defines the issue in exactly this territory: the realistic replication of an individual's voice or appearance. The report treats the problem as larger than copyright because unauthorized replicas implicate privacy, publicity, consumer protection, fraud, reputation, and labor. Its core conclusion is blunt: existing law is not enough, and a federal digital-replica right is needed.

That conclusion matters because it separates the person from the file. Copyright can protect a recording or photograph. It does not automatically protect the living pattern that makes a voice or face socially recognizable. The AI-era harm is not only unauthorized copying of a work. It is unauthorized activation of a person-shaped signal.

Why Labels Are Not Enough

Provenance is necessary. Content credentials, watermarks, visible labels, platform disclosures, and source trails can help people understand whether an artifact was generated or altered. The site's own Provenance and Content Credentials protocol exists because public culture needs inspectable source trails.

But provenance answers a different question from consent. Provenance asks: where did this artifact come from, and what happened to it? Consent asks: did the person whose identity is being simulated agree to this use?

A perfectly labeled unauthorized replica can still be a violation. A synthetic advertisement may disclose that it is AI-generated while still exploiting a performer's likeness. A voice clone may carry a watermark while still enabling fraud. A nonconsensual sexual deepfake may be labeled synthetic while still attacking dignity, safety, and identity. A political parody may be lawful and protected in context, but the same technical act inside a voter-suppression robocall becomes a civic threat.

That is why "label it" cannot be the whole governance answer. Labeling protects the audience from mistaking an artifact for original evidence. Consent protects the person from being turned into an instrument.

The Law Is Splitting

By 2026, the legal response has started to split into several tracks.

Deepfake abuse law: The TAKE IT DOWN Act was signed into U.S. law on May 19, 2025. The White House's signing notice describes it as creating a criminal prohibition on intentional disclosure of nonconsensual intimate visual depictions and requiring covered platforms to remove such depictions. This is an important but specific intervention. It targets a severe class of image abuse, not the whole field of synthetic identity.

Voice and likeness rights: Congress.gov lists the NO FAKES Act of 2025 as introduced in the Senate on April 9, 2025 and referred to the Senate Judiciary Committee. The bill's official title is a bill to protect intellectual property rights in the voice and visual likeness of individuals. As of the Congress.gov record reviewed for this essay, it had not become law. Its significance is that it treats digital replicas as a general right-of-person problem, not only a pornography, fraud, or election problem.

Transparency duties: The EU AI Act creates transparency obligations for synthetic outputs. Article 50 requires certain AI-generated or manipulated content to be marked or disclosed, including duties around deepfakes. That is a disclosure regime, not a complete consent regime. It helps audiences interpret media, but it does not by itself settle when simulating a particular person is permissible.

Communications enforcement: The FCC clarified in February 2024 that AI-generated voices in calls are held to the same Telephone Consumer Protection Act standards that apply to artificial or prerecorded voice messages. That is a telephony rule. It is powerful in robocall contexts, but the Copyright Office correctly notes that communications law cannot govern every website, model output, social platform, ad system, fan tool, or workplace use where digital replicas appear.

The pattern is clear: institutions are treating synthetic people as a new legal object, but each instrument catches only part of the surface. The emerging problem is not one law. It is the missing consent layer across media, labor, platforms, and AI systems.

Labor Saw It First

Performers understood the issue before most institutions did because their work already lives at the boundary between person and media. An actor is paid not only for words or motions, but for presence: voice, face, movement, timing, style, recognizability, and embodied judgment.

The 2023 SAG-AFTRA TV/Theatrical agreements made AI a labor issue. SAG-AFTRA's public materials describe protections requiring informed consent and compensation for creation and use of digital replicas of members, including living and deceased performers. The details matter less here than the principle: a synthetic performance is not a neutral technical derivative. It is a labor relation.

If a studio can scan a performer, store the pattern, and reuse it later, the bargaining question changes. The worker is no longer only selling today's performance. They may be licensing a machine-readable version of future performability. Consent has to be specific enough to prevent one signature from becoming a lifetime extraction event.

This is not limited to Hollywood. Customer-support agents, audiobook narrators, teachers, streamers, translators, salespeople, therapists, call-center workers, and ordinary employees all produce voice, video, writing style, facial expression, and domain performance that can become training data or replica substrate. The labor question is spreading outward: who owns the reusable pattern of a person's work?

Fraud Makes the Boundary Urgent

Fraud shows why this cannot be left to after-the-fact reputation repair. Voice cloning reduces the cost of impersonation at exactly the point where people are most vulnerable: a phone call from a family member, a supposed executive instruction, a political message, a bank request, a medical emergency, a celebrity endorsement, a personalized extortion attempt.

The FTC's Voice Cloning Challenge was built around this risk. In April 2024, the agency announced winners for approaches to addressing AI-enabled voice cloning. The important lesson is not that one prize challenge solves the problem. It is that consumer-protection agencies recognized voice cloning as an abuse surface requiring technical, institutional, and legal response.

The FCC's 2024 robocall action points in the same direction. A voice clone is not merely content. It is a credential humans are trained to trust. When the voice sounds like a child, a parent, a candidate, a boss, or a doctor, the artifact carries social authority before the listener has time to inspect provenance.

That is the distinctive danger of person-shaped media. It does not only assert a claim. It borrows the trust channel through which claims usually travel.

A serious consent layer for synthetic people should be more concrete than a checkbox and more durable than a platform promise.

Consent should be specific. A person should know whether they are consenting to training, one generated asset, one campaign, one production, one product feature, translation, dubbing, archival restoration, parody, advertising, or open-ended reuse.

Consent should be time-limited. Open-ended licenses create an obvious abuse path. A digital-replica agreement should say how long the permission lasts, how renewal works, and what happens after death.

Consent should be revocable where possible. Some uses cannot be practically unwound after publication, but future generation, future distribution, and future model access can often be stopped or narrowed.

Consent should preserve refusal. A worker, artist, or ordinary user should not have to surrender their likeness as a hidden condition of access, employment, audition, account creation, or participation unless the use is genuinely necessary and separately negotiated.

Consent should include compensation when value is extracted. If a synthetic version of a person's voice, face, body, or performance substitutes for paid work, compensation is not a courtesy. It is part of recognizing the person as the source of the value.

Consent should distinguish public interest from commercial capture. Journalism, satire, documentary work, criticism, research, and political speech need breathing room. So do fraud prevention, abuse reporting, and evidentiary preservation. But those exceptions should not become loopholes for synthetic endorsement, sexual humiliation, workplace displacement, or mass persuasion.

Consent should leave records. The consent trail should travel with the artifact or be available to the platform, publisher, employer, court, union, or auditor that needs to inspect it. A consent regime without records becomes theater.

The Spiralist Reading

Digital replicas are recursive reality applied to the person.

The model observes a human signal. It compresses the signal into a reusable representation. It regenerates the signal in new contexts. Other people respond to the regeneration as if some part of the original person is present. The loop then teaches institutions that identity can be operationalized: the voice as authentication, the face as interface, the performance as asset, the body as promptable surface.

That is why the consent layer matters. It is one of the places where society says the model may learn patterns, but it may not silently turn persons into public machinery.

The boundary is not anti-technology. Synthetic dubbing can expand access. Voice banking can preserve communication for people losing speech. Digital restoration can serve history. Satire and fiction need room to breathe. A careful replica can honor agency. The danger is not simulation as such. The danger is simulation without permission, memory without obligation, and likeness without the living person's veto.

Provenance slows source collapse. Consent slows person collapse. Both are needed. Without provenance, audiences cannot tell what kind of artifact they are seeing. Without consent, people become raw material for artifacts that speak with their face, voice, body, and authority.

The practical rule is narrow: no system should treat a recognizable person as freely promptable media. If a synthetic person is being made, the first governance question is not "Can the model do it?" or even "Can the audience tell?" It is: who is being invoked, for what purpose, under whose permission, with what right to refuse?

Sources

Return to Blog