Who Owns a Performance? AI Actors, Digital Replicas, and the Future of Consent

When Robert Zemeckis had actors Tom Hanks and Robin Wright age and de-age across roughly six decades for his film “Here,” it no longer happened frame by frame in months of post-production. The VFX studio Metaphysic trained its own machine-learning models on the actors' earlier films and generated the altered faces in real time on set (WIRED). The VFX supervisor stressed that the rights to all source material had been secured and that artists were involved in order to stay true to the performance (Variety).

This is precisely the heart of a debate that has gripped Hollywood ever since the strike year of 2023. A performance is not merely image or sound. It consists of identity, human labor, and interpretation, and of data that can be commercially reused. As soon as face, voice, and movement exist as trainable material, the real question arises: who owns this performance if it can be reproduced and altered at will? The thesis of this article: technical editability does not establish blanket power of disposal. The fact that digital actors are possible says nothing about who is entitled to decide over them.

A conceptual compass, and why precision matters

The discussion suffers from conflation. “AI actor” is an umbrella term covering very different techniques that must be assessed differently in legal and ethical terms. What matters is a clean distinction, above all between established VFX and capture techniques and generative AI.

• Digital replica: In the California law cited here (AB 2602, AB 1836), the term is defined as a “computer-generated, highly realistic electronic representation” that is recognizable as the voice or likeness of an identifiable person and in which the person either did not actually perform at all or whose performance was “materially altered” (Davis+Gilbert LLP). Mere remastering or sampling is expressly excluded (ibid.). This legal definition applies to California law and is not universal; other jurisdictions and collective bargaining agreements construe the term more narrowly or more broadly.
• Digital double (digi-double): a VFX technical term, a “virtual replica of a real person in three dimensions,” whose physical features are typically captured via 3D scanning and photogrammetry (Eisko). A digital double is thus tied to a real person and primarily oriented toward visual accuracy; it is not identical to the labor-law “employment-based digital replica” of the collective agreements, even if both can be based on scans of the same performer.
• Digital human / fully computer-generated character: a figure “designed originally from the ground up” that is not a replica of an existing person (Eisko). Such figures can be created with generative AI, but need not be; classical sculpting and animation are equally sufficient (ibid.). “Fully computer-generated” therefore does not automatically mean “AI-generated.”
• SAG-AFTRA “Synthetic”: by contrast, a narrow, contractual term, an AI-generated asset that creates the impression of a natural performer, is not voiced by a real person, and does not depict an identifiable person (Weintraub Tobin “The Briefing”). A CGI character built purely by sculpting is therefore not necessarily a “Synthetic” within the meaning of the contract.
• Face replacement / face swap: replaces or blends a face; Metaphysic advertises “real-time photoreal faceswaps” (vendor statement) (Metaphysic Live).
• Voice cloning / speech-to-speech: generates a voice output from source material. The provider Respeecher describes its technique as language-agnostic and emphasizes “ethical use,” “strict consent mechanisms,” and compensation (vendor statement) (Respeecher).
• Performance capture and motion capture: record the facial expression and movement of a real performer, established capture techniques that are not, in themselves, generative AI.
• De-aging: can be done classically (frame-based VFX as in “The Irishman”) or generatively (real-time models as in “Here”) (Variety).
• AI-assisted alteration of a real performance: edits the actually delivered work, for instance an accent.

How narrow the line runs is shown by “The Brutalist.” Editor Dávid Jancsó used the AI of the company Respeecher to refine the Hungarian dialogue lines of Adrien Brody and Felicity Jones for a more authentic accent (The Guardian). Director Brady Corbet stressed that the performances were “entirely their own,” the goal being not replacement but authenticity in another language (ibid.). The process is an AI-assisted alteration of a real performance, not an independent replica. Anyone who conflates these categories, or blanket-labels classical VFX and capture techniques as “generative AI,” blurs precisely the lines at which decisions about consent and compensation are made.

Consent as a founding principle: what the collective agreements enshrine

The turning point was the SAG-AFTRA agreement of 2023. For the first time, informed consent (“informed consent”) and fair compensation for the creation and use of digital replicas became binding industry-wide, for living as well as deceased performers, whether generated on set or licensed from third parties (Screen International). Studios must adequately describe the intended use; background performers are protected against unauthorized use of their replicas (ibid.).

On June 5, 2026, SAG-AFTRA ratified a successor agreement with 91.42 percent approval; it runs from July 1, 2026 to June 30, 2030 and is set for four years for the first time (GreenSlate; Los Angeles Times). According to a legal analysis, the AI rules are anchored in two sections: Section 64 for replicas and digital alterations, Section 64.1 for generative AI and “Synthetics” (Weintraub Tobin).

The new cornerstones (as presented in this contract analysis):

• Lead time and form of consent: For the scan to create an “employment-based digital replica,” at least 48 hours' lead time is required; the consent must be “clear and conspicuous” and signed separately (Weintraub Tobin).
• No blanket consent: Required consents must contain a “reasonably specific description” of the intended use; sweeping, open-ended consents are excluded (ibid.).
• Compensation and residuals: If a replica replaces real work, the pro-rata daily rate or the minimum is paid, whichever is higher, and residuals accrue if the performance would have triggered them (ibid.).
• Foreign-language dubbing: As of July 1, 2027, using an AI voice replica for dubbing requires consent under Section 64 (ibid.).
• No strikebreaking by replica: A digital replica may not be used when a performer could refuse to work during a lawful strike (ibid.).

For Synthetics, not recognizable as a specific person, a “principle in favor of human performances” applies, according to the analysis. A Synthetic may take on a human role only if it brings “significant additional value”; otherwise, notification to the union, negotiation, and, if necessary, an arbitration procedure are provided for (Weintraub Tobin). The SAG-AFTRA leadership describes this as a strong incentive in favor of human performers, not a complete ban (union position, Variety).

In parallel, video-game performers set standards: their Interactive Media Agreement, ratified in July 2025 with 95.04 percent, requires consent and transparency for AI replicas and allows consent for new content to be withdrawn during a strike (Reuters). The framing is important: this right of withdrawal concerns video games and is tied to a strike. No general, at-any-time right of withdrawal for every film or television use follows from the sources documented here. Whether consent once granted can be revoked depends on the specific contract; project, purpose, term, and territory must be examined case by case.

Producers do have legitimate interests here. Continuity across long shooting periods, reshoots and pick-ups, de-aging as in “Here” (WIRED), localization such as the accent work in “The Brutalist” (The Guardian), scheduling and risk planning, as well as a robust chain of title, are real practical production reasons for digital techniques. Precisely for that reason, specific consent, documented usage rights, and clear compensation are not an obstacle but a prerequisite; only those who hold cleanly described rights can reuse a replica with legal certainty. The California rule that voids clauses lacking a specific description of use thus protects both sides (Davis+Gilbert LLP).

The law catches up, but unevenly

Contracts bind only the contracting parties. That is why the legislature steps in. In California, two laws have applied since January 1, 2025 (Davis+Gilbert LLP): AB 2602 voids contractual clauses that permit a replica as a substitute for real work, contain no specific description of use, and where the person was represented by neither counsel nor a union (ibid.). AB 1836 prohibits the creation or distribution of digital replicas of deceased personalities in expressive works without prior consent and provides for penalties of at least 10,000 US dollars or the actual damages (ibid.; Senate Judiciary Committee).

At the federal level, the U.S. Copyright Office already recommended a new federal law in July 2024 in “Copyright and Artificial Intelligence, Part 1: Digital Replicas.” The report notes that existing laws do not offer adequate protection, that copyright does not protect a person's identity as such, and that a “patchwork” of regulations exists (U.S. Copyright Office). It recommends a right for all persons during their lifetime, informed consent, “clear and conspicuous” contractual clauses, and a ban on the full sale of rights (ibid.; Copyright Office NewsNet 1048).

At the federal level, the associated bill, the NO FAKES Act, came out of the Senate Judiciary Committee by unanimous voice vote, according to reporting (Deadline). The bill would give every person, not just celebrities, the right to authorize the use of their voice and likeness in digital reproductions; this right would be inheritable and licensable and would survive death by up to 70 years, with exceptions for news, documentaries, sports, biographical works, as well as commentary, criticism, and parody (ibid.). It is supported across party lines (among others Blackburn, Coons, Tillis, Klobuchar); a related bill is before the House of Representatives (ibid.). Crucial for the assessment: the NO FAKES Act is not a law in force. It has only cleared the committee stage; whether and when it will reach the Senate and the House of Representatives is open, and the 2024 version did not make it past the committee (ibid.). Penalty amounts cited in reporting, such as up to 750,000 US dollars per work for platforms, come from secondary trade press and are not to be understood as law in force (Music Business Worldwide). Legal statements therefore apply depending on jurisdiction and date; this text is no substitute for individual legal advice.

Counterarguments, limits, and risks

The advances are real, but they have gaps. First, enforcement: even supporters concede that while the new SAG-AFTRA framework documents consent and compensation, practical enforcement remains open (Weintraub Tobin). Second, the power to define: a co-chair of the union's technology committee criticized that in the end a studio lawyer defines what “significant additional value” means, and that when licensing performances to third parties, a meeting is provided for, but “no consent” and “no guaranteed compensation” for the individual performer (union position, Variety).

Third, territorial inequality: the California laws apply only in California; a nationwide standard does not yet exist (Davis+Gilbert LLP). Fourth, training data and provenance: regarding the AI figure “Tilly Norwood,” which according to its maker Particle 6 was not designed after real models, SAG-AFTRA president Sean Astin said such systems take the work of members built up over generations “without permission, without compensation and without acknowledgment” (union position, Los Angeles Times).

Fifth, the vendor perspective: technology providers emphasize legitimate benefits. “Here” would not have been possible in this form without the real-time models; the VFX supervisor highlighted that rights were secured and artists involved (Variety). Respeecher advertises strict consent mechanisms and compensation (vendor statement) (Respeecher). Such self-reporting is no substitute for independent scrutiny, but it does show that parts of the industry orient themselves toward consent standards.

Conclusion: control before compatibility

The technical question has long been answered: face, voice, movement, and acting can be reproduced, altered, and reused beyond death. The real questions of 2026 are different ones. Who grants informed, project-specific consent rather than a blanket authorization? Who is compensated when a replica replaces real work? And who prevents “feasible” from being confused with “permitted”?

Contracts, bills, and agency reports converge on the same principle: a performance combines identity worthy of protection, human labor, and contractually exploitable rights; it is not freely available raw material. In doing so, one must remain legally precise, US copyright does not protect a person's identity as such (U.S. Copyright Office); protection arises from a web of contract, collective bargaining law, and state laws that is only now growing together. Editability is not a title of ownership. Anyone who wants to dispose of a digital replica needs consent, clarity about purpose, time, and territory, and a fair share of the economic value that a person's data generates.