Digital Replicas: The Copyright Office’s Part 1 Report, the Law That Has Followed It, and a Compliance Guide for Copyright Owners Creating Content with Large Language Models

The first three blog entries in this series addressed whether AI-generated content can be copyrighted, whether training on copyrighted works is fair use, and what the Bartz v. Anthropic settlement means for both questions. This entry addresses a different risk—one that does not turn on copyright at all, and that is now the most common way a business using generative AI ends up on the wrong side of a demand letter. It is the risk of creating a digital replica: an AI-generated image, video or voice that realistically but falsely depicts a real person.

The United States Copyright Office addressed this subject first, in Part 1 of its report on Copyright and Artificial Intelligence, released July 31, 2024. The report is a careful survey of the harms digital replicas cause, the patchwork of state and federal law that currently governs them, and the federal statute the Office believes Congress should enact. Two years later, Congress has not yet enacted it—but the NO FAKES Act that would do so cleared the Senate Judiciary Committee unanimously in June 2026, the federal TAKE IT DOWN Act now criminalizes the worst category of replicas, Florida and a dozen other states have passed their own statutes, and the first federal court decisions applying existing law to AI voice clones have been issued. This post summarizes the report, updates the legal landscape, and closes with practical recommendations for copyright owners and content creators who use large language models and image, video and voice generators in their work.

For an overview of our copyright and entertainment practices, see Copyright Matters and Entertainment Law.

The Part 1 Report

What a Digital Replica Is

The Office defines a digital replica as “a video, image, or audio recording that has been digitally created or manipulated to realistically but falsely depict an individual.” Two features of the definition matter. First, it is technology-neutral: a replica can be produced with generative AI or with conventional editing software, and the report uses “digital replica” and “deepfake” interchangeably. Second, it turns on realism. The concern is not caricature, impression or obvious parody, but depictions “so realistic that they are difficult to distinguish from authentic depictions.” A cartoon of a celebrity is not a digital replica, but a photorealistic video of that celebrity saying something she never said is.

The Harms the Office Identified

The report catalogs four categories of harm, and each has a business analogue. The first is sexual exploitation: the Office cited research that explicit content made up 98 percent of deepfake videos online and that 99 percent of the individuals depicted were women, and it noted the growing problem of students creating explicit images of classmates. The second is fraud, including a case in which a finance employee was induced to transfer $25.6 million after a video call with a replica of the company’s chief financial officer, and voice-clone scams in which “family members” call demanding money. The third is electoral and political manipulation, including a robocall using a replica of President Biden’s voice to discourage primary voting. The fourth—and the one most directly relevant to the creative industries — is displacement of performers and artists. The list is extensive. There are film productions that are already using replica extras in place of background actors, voice actors replaced by clones of their own voices, and musicians whose voices were used to release “new” songs they never recorded. Essentially, “copyrighted works are [being] used to produce digital and the replicas are often disseminated as part of larger copyrighted works”.

Why Existing Law Is Not Enough

The heart of the report is its analysis of the law as it stands, and its conclusion that existing remedies are “both inconsistent and insufficient.”

Copyright. The Office is direct on this point: “Copyright does not ... protect an individual’s identity in itself, even when incorporated into a work of authorship. A replica of their image or voice alone would not constitute copyright infringement.” A photograph is protected; the face in it is not. A sound recording is protected; the voice on it is not. If an AI system generates a new image or a new recording that depicts a person without copying any protected work, copyright provides no claim. The report does note two adjacent copyright provisions—Section 114(b), which permits sound-alike recordings that do not copy the original recording, and which the Office recommends clarifying so that it does not preempt state voice-replica laws, and the moral-rights concepts of attribution and integrity, which the Office observes are analogous to the non-commercial harms replicas cause.

State right of publicity and privacy. These are, the Office acknowledges, “the most directly relevant state law [remedy],” but they vary enormously. Some states have no statutory right, some protect only persons whose identity has commercial value, some do not protect the voice at all, or protect it only if it is “distinctive”, postmortem terms range “from as short as 20 years in Virginia, to 100 years in Indiana, and indefinitely in Tennessee”, and most statutes reach only advertising, merchandise and other commercial uses, leaving the non-commercial replicas that are “particularly prevalent in the internet environment” unaddressed. The result, the Office found, is that “[d]ifferent jurisdictional requirements create discrepancies as to who may seek relief.”

The Lanham Act. Section 43(a) false-endorsement claims exist, but they require commercial use and a likelihood that consumers will be confused about the person’s sponsorship or approval—a showing the Office notes “may be difficult for many individuals, including… artists and performers” to make.

The FTC Act and the Communications Act. The Federal Trade Commission can reach deceptive practices and the Federal Communications Commission has ruled that AI-generated voices in robocalls are “artificial” voices under the Telephone Consumer Protection Act, but both address narrow slices of the problem and neither gives the depicted individual a private remedy.

The Federal Statute the Office Recommended

Having found the gaps, the Office recommended that Congress enact a new federal right, and it specified the elements in detail. The statute should protect all individuals, “not just celebrities, public figures, or those whose identities have commercial value,” because “[e]veryone is vulnerable to the harms that unauthorized digital replicas can cause.” It should be narrower than state name-image-likeness rights, targeting only realistic replicas. It should last at least for life, with any postmortem term “limited in duration.” Liability should attach to distribution or making available of an unauthorized replica, not to creation alone, and should require actual knowledge that the content is a replica of a particular person and that it is unauthorized. Traditional secondary-liability principles should apply, with a notice-and-takedown safe harbor for online platforms. Individuals should be able to license their replica rights, “subject to guardrails, but not to assign them outright,” with extra safeguards for minors. Free-speech concerns should be addressed through a balancing framework rather than categorical exemptions. Remedies should include injunctions, damages, and statutory damages or fee-shifting so that protection does not depend on the victim’s resources, with criminal liability “[i]n some circumstances.” And the federal law should be a floor, not a ceiling: the Office did “not recommend full federal preemption” of state law.

The report also considered, and declined to recommend, protection for artistic style — the ability of an AI system to produce work “in the style of” a living artist or author. The Office acknowledged “the seriousness of creators’ concerns” but concluded that style should not be included in new legislation “at this time,” pointing to existing remedies—unfair competition, false endorsement, and copyright where actual expression is copied—as the appropriate tools.

What Has Happened Since

Federal Legislation

The NO FAKES Act—the Nurture Originals, Foster Art, and Keep Entertainment Safe Act—is the vehicle for the Office’s recommendations. First introduced in 2023 and reintroduced in April 2025, it was reintroduced again in 2026 as S. 4591 and H.R. 8915 with bipartisan sponsors including Senators Blackburn, Coons, Tillis and Klobuchar and Representatives Salazar, Dean and Schiff, and with the support of the entertainment guilds, the studios, the record labels and several major AI developers. On June 18, 2026, the Senate Judiciary Committee advanced it by unanimous voice vote. The bill would create:

• A federal property right in an individual’s voice and visual likeness against unauthorized digital replicas;
• The right would survive death for up to 70 years;
• Licenses would be subject to limits, including heightened requirements for minors;
• Online platforms would be shielded from liability if they operate a notice-and-takedown system, with penalties of at least $25,000 for knowingly false counter-notices;
• News, documentary, biographical, sports, parody and similar uses would be excluded; and
• The bill would preempt future state digital-replica laws while preserving those enacted before January 2, 2025.

Several senators raised First Amendment concerns at the markup, and the bill’s floor timing is uncertain. If it passes in substantially its current form, it will track the Office’s recommendations closely, with the significant exception that it preempts future state law rather than setting a floor.

Congress did enact one narrower measure. The TAKE IT DOWN Act, signed May 19, 2025, makes it a federal crime to knowingly publish non-consensual intimate imagery, including AI-generated “digital forgeries,” and required covered platforms to establish a 48-hour notice-and-removal process by May 2026, enforced by the FTC. However, it only addresses the first of the Office’s four harms.

State Legislation, Including Florida

The states have moved faster. Tennessee’s ELVIS Act, effective July 1, 2024, was the first to add voice expressly to the right of publicity and to reach the distribution of tools whose primary purpose is producing unauthorized replicas. California enacted two statutes in September 2024 that require contracts for digital replicas of performers to be specific and to include professional representation, and that prohibit commercial use of replicas of deceased performers without estate consent. New York amended its labor law to similar effect for performers’ contracts. By mid 2026, more than half the states had enacted some form of digital-replica or deepfake legislation.

Florida’s approach has been targeted rather than comprehensive. Florida’s general right of publicity, Fla. Stat. § 540.08, has long prohibited the publication of a person’s name, photograph or “likeness” for trade, commercial or advertising purposes without express written or oral consent, with a 40-year postmortem right and statutory remedies, and now it applies with full force to an AI-generated likeness used in advertising. Since 2024 the Legislature has added three specific measures. First, Fla. Stat. § 836.13, which criminalizes the creation and distribution of “altered sexual depictions” of identifiable persons, provides civil damages of the greater of $10,000 or actual damages plus fees, and—under 2025 amendments known as Brooke’s Law—requires platforms to remove such content within 48 hours of notice or face liability under the Florida Deceptive and Unfair Trade Practices Act. Second, Fla. Stat. § 106.145, which requires a disclaimer on political advertisements that use generative AI to depict a real person doing something that did not occur. Last but not least, FDUTPA itself, which the Attorney General has applied to deceptive AI-generated content aimed at consumers. A broader “Artificial Intelligence Bill of Rights” bill that would have added name-image-likeness protections, SB 482, died in the 2026 session. Therefore, Florida has no general digital-replica right beyond § 540.08, which still requires a commercial use. Inevitably, Floridians harmed by non-commercial replicas outside the sexual-depiction context must rely on common-law privacy torts, defamation, or the federal law when it arrives.

Our FDUTPA page addresses the consumer-protection overlay in more detail.

The Courts

The first federal decision applying existing law to AI voice replicas confirms the Office’s diagnosis. In Lehrman v. Lovo, Inc., two voice actors alleged that they had been hired through Fiverr for what they were told were internal research recordings, and that the buyer—an AI voice company—had instead used the recordings to build commercial voice clones sold under invented names. The court dismissed the Lanham Act false-endorsement claims and nearly all of the copyright claims, holding that a voice is not a copyrightable work and that training on the recordings and generating new clone outputs did not reproduce any protected expression. What survived were the claims the Office would have predicted: breach of contract, New York’s statutory right of publicity (Civil Rights Law §§ 50–51), and New York’s consumer-protection statute—together with a single copyright claim for the defendant’s use of an actual recording in marketing materials. The lesson is that, as the law now stands, the strongest claims against a digital replica are contract and state publicity claims, not copyright.

Two older cases remain the foundation for voice claims and are worth knowing. In Midler v. Ford Motor Co., the court held that deliberately imitating a well-known singer’s distinctive voice in an advertisement violated California common law, and in Waits v. Frito-Lay, Inc., a jury awarded $2.6 million on similar facts, including a Lanham Act false-endorsement verdict. Both decisions predate AI by decades, and both apply directly to a voice clone used in a commercial.

Recommendations for Copyright Owners Creating Content with Large Language Models

Our clients increasingly use large language models and their image, video and voice counterparts to produce marketing, entertainment, educational and commercial content at scale. The following recommendations are designed to let them do so without creating an unauthorized digital replica, and to preserve their own rights when they license a real person’s likeness or voice. They should be read together with the recommendations in our earlier post on copyrightability, which address how to ensure the content itself is protectable.

1. Treat any depiction of a real person as a clearance issue

The threshold question for every AI-generated image, video or voice is whether it depicts, or could be understood to depict, an identifiable real person. If it does, the content requires the same clearance that a photograph or a recording of that person would require—regardless of the fact that no photograph or recording was used to make it. For instance, prompts that name a real person, describe a recognizable public figure, upload a person’s photograph or voice sample as a reference, or ask for content “in the voice of” or “looking like” someone should be treated as generating a replica and routed to review. This applies equally to employees, customers, influencers and executives of the company itself: an AI-generated video of your own CEO requires the CEO’s consent, and the consent should be in writing.

2. Obtain written, specific licenses and understand their limits

Where a real person’s likeness or voice will be replicated, obtain a written license that:

• Identifies the person;
• Describes the replica to be created;
• Specifies the permitted uses (media, territory, duration, and the specific projects or campaigns);
• States whether the replica may be modified or used to generate new content beyond the licensed material;
• Addresses the person’s right to review and approve outputs;
• Provides for termination and destruction of the model or voice profile; and
• Allocates compensation.

The performers’ unions have negotiated detailed consent and compensation terms for digital replicas of their members since the 2023 SAG-AFTRA agreement, and those terms are a useful benchmark even for non-union talent. Florida’s § 540.08 requires express consent for commercial use, California and New York now impose specific contractual requirements for performers, and the NO FAKES Act, if enacted, will limit licenses to specified uses and impose additional requirements for minors. A license that says only “all rights in perpetuity in all media now known or hereafter devised”, which is the traditional boilerplate, may not satisfy those statutes and should be replaced with replica-specific language.

Our entertainment law practice prepares these agreements.

3. Build your own replica rights into talent and employment agreements now

Businesses that employ on-camera talent, spokespersons, voice-over artists, podcast hosts or influencers should update their agreements to address whether the company may create and use a digital replica of the individual, on what terms, and for how long after the relationship ends. Absent such a provision, the company that has invested in a spokesperson’s public identity may find that it cannot use AI to produce new content with that identity, and the individual may find that the company claims a right it never bargained for.

The same is true in reverse. Creators who license their own voice or likeness to an AI platform should insist on the guardrails the Office recommended: a limited term, defined uses, no outright assignment, approval rights, and destruction on termination.

4. Do not replicate the dead without the estate

Postmortem rights are where the state patchwork is most treacherous. Florida’s right runs 40 years after death, California’s 70, Tennessee’s indefinitely, and the NO FAKES Act proposes up to 70 years. A replica of a deceased performer, author, athlete or public figure used in commercial content requires clearance from the estate or its licensing agent in every state where the content will be seen, and the safest practice is to clear under the most protective regime.

5. Handle “style” with care even though the Office declined to protect it

The Office did not recommend statutory protection for artistic style, and a prompt asking for text “in the style of” a living author, or an image “in the style of” a living illustrator, does not by itself create a digital replica. But three cautions apply. First, if the output reproduces actual protected expression from the artist’s works, it infringes copyright. Second, if the output is marketed in a way that suggests the artist’s involvement or endorsement, it invites a Lanham Act false-endorsement claim and, in Florida, a FDUTPA claim. Lastly, if the artist’s name is used in advertising the output, § 540.08 is implicated.

The practical rule is that style may be a creative reference internally, but the artist’s name should not appear in the product, its marketing or its metadata.

6. Label, disclose and preserve provenance

An accurate disclosure that content is AI-generated is not a defense to an unauthorized-replica claim. Florida’s § 836.13 says so expressly, and the Office’s recommended statute does not contemplate one. However, disclosure is required in specific contexts (Florida political advertising, FTC endorsement rules where a replica functions as a testimonial, several states’ consumer-facing chatbot laws) since it reduces the risk of a deception-based claim, and it is increasingly expected by platforms and audiences. Adopt the content-credential standards the major platforms now support, embed provenance metadata in AI-generated assets, and keep internal records of how each asset was created, what references were used, and what licenses cover it. Those records are also what will allow the company to prove human authorship if it later seeks to register the content.

7. Include a replica policy in your AI-use guidelines

Every organization creating content with AI should have a short written policy, reviewed by counsel, that: prohibits generating realistic depictions of real people without documented consent and uploading third-party photographs, recordings or personal data as generation references without rights, requires legal review of any content depicting a public figure, a competitor’s personnel, or a minor, specifies the disclosure and labeling practices to be followed, and identifies who is responsible for clearance. The Office’s recommended statute—and the NO FAKES Act—would attach liability to distribution with knowledge, which means the company that publishes content its employees generated is the party at risk, and a documented policy is the company’s evidence of what it knew and did.

8. Monitor for replicas of your own people and brands, and be ready to act

The same tools that create replicas are being used to create replicas of company executives (for fraud), spokespersons (for counterfeit endorsements) and brand ambassadors. Monitor for them as you would monitor for trademark misuse—see our earlier post on trademark enforcement for the mechanics—and be prepared to use every available channel: the platform takedown processes now mandated by the TAKE IT DOWN Act and Brooke’s Law for sexual deepfakes, the platforms’ own impersonation and synthetic-media policies for everything else, Section 43(a) and § 540.08 for commercial uses, defamation and FDUTPA where the content is false or deceptive, and the DMCA where an actual photograph or recording was copied. Where a replica is used to defraud the company or its customers, involve law enforcement and the FTC.

Our Litigation & Enforcement team handles these matters.

9. Remember that the replica question is separate from the copyright question

Finally, clearing a person’s likeness does not make the resulting content copyrightable, and copyrighting the content does not clear the likeness. A licensed digital replica generated entirely from a prompt is still, under the Copyright Office’s Part 2 report, unprotectable AI output and a fully human-authored video that includes an unlicensed replica is still actionable. The two analyses must be run in parallel, and the documentation for each—the human creative contribution on one side, the consent and license on the other—should be kept together in the asset’s file.

Conclusion

The Copyright Office’s Part 1 report was the first authoritative statement that the harms of digital replicas are real, that copyright cannot address them, and that a new federal right is needed. Two years later, the diagnosis has been confirmed by the first court decision on AI voice clones, the prescription is one committee vote closer to enactment, and the states— including Florida—have filled some of the gaps with targeted statutes.

For businesses creating content with large language models, the practical rule is simple to state: a realistic depiction of a real person is a licensed asset, not a prompt, and it should be cleared, documented and labeled as one. The businesses that build that discipline now will be positioned for the federal right when it arrives, and protected under the state law that already applies.

If your organization is creating AI-generated content that depicts real people, licensing a performer’s or executive’s voice or likeness, or responding to an unauthorized replica of your own personnel or brand, our attorneys can help.

Please visit our Copyright Matters, Entertainment Law and Litigation & Enforcement pages, or contact us at 305-374-8303.