NYLJ “New York Court Tackles IP Issues Around AI Voice ‘Clones’”

By Steve Kramarsky

Late last month, two important decisions came down from courts in the Northern District of California regarding the unauthorized use of copyrighted material for the training of large language models (commonly referred to as AI training).

The two decisions are Judge William Alsup’s opinion in Bartz v. Anthropic PBC, No. C 24-05417 WHA, 2025 WL 1741691 (N.D. Cal. June 23, 2025), and Judge Vince Chhabria’s opinion two days later in Kadrey v. Meta Platforms, Inc., No. 23-CV-03417-VC, 2025 WL 1752484 (N.D. Cal. June 25, 2025).

Over the past few weeks, a great deal has been written about these two cases, but no real consensus has emerged as to the effect they will have on the broader AI litigation landscape, including the major AI cases pending in the Southern District of New York.

Earlier this month, in Lehrman v. Lovo, Inc., No. 24-CV-3770 (JPO), 2025 WL 1902547 (S.D.N.Y. July 10, 2025), Judge J. Paul Oetken addressed some similar issues, and his opinion offers an opportunity to evaluate the legal context and examine some underlying policy concerns that may lead to differing outcomes in the New York courts.

Background: The ‘Anthropic’ and ‘Meta’ Decisions from California

Oddly, for two such different decisions, the two California cases reach the same bottom-line result: on the facts and arguments presented, both courts granted partial summary judgment to the AI companies holding that their use of books to train large language models without permission from the authors constituted “fair use” under the copyright law.

Both courts also held, however, that the plaintiffs’ claims relating to the unauthorized downloading of their books from pirating websites could go forward. Given these identical outcomes (and nothing more), it is tempting to think that the cases establish a simple rule: as long as you pay for a copy of the training data and don’t steal it, training AI on copyrighted material is fair use.

The AI companies have characterized the decisions that way, and in some cases the press has followed suit. But a closer reading of the opinions reveals something far more complex, and suggests that relying on that proposition would be a mistake.

Alsup’s opinion in Anthropic adopts an expansive view of the fair use exception. He finds AI training to be a “spectacularly transformative” use of copyrighted works and, based on that finding and an analysis of the other statutory factors that in some ways derives from it, he holds that AI training will almost always constitute fair use rather than infringement.

Absent a claim that the AI model can be prompted to spit out infringing copies of the training data (which is at issue in some of the New York cases), Alsup’s opinion leaves little room for future plaintiffs to allege infringement.

In the Meta opinion, on the other hand, Chhabria writes that, “in most cases” the use of copyright- protected material to train AI models without paying the copyright holder will be illegal.

He grants partial summary judgment to Meta, because “plaintiffs made the wrong arguments and failed to develop a record in support of the right one,” but his opinion lays out his philosophical disagreement with Alsup on fair use and gives a roadmap for future plaintiffs to bring a more successful infringement case.

The essential difference between the opinions comes down to the fourth statutory factor in fair use analysis, which is “the effect of the use upon the potential market for or value of the copyrighted work”. 17 U.S.C. §107 (4).

Copyright law is designed to promote the creation of new works by protecting the rights of creators without overly limiting access to their works, and fair use analysis is one tool that attempts to strike that balance.

The fourth factor explicitly looks to economic effects: if an unauthorized copy is likely to harm the market for the original through direct competition it is probably not fair use; if it does not directly compete with the original, then it is more likely to be fair use.

In the AI context, this becomes a philosophical debate over the idea of market competition. Alsup views AI training as analogous to teaching schoolchildren how to read and write.

He writes: “Authors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works. This is not the kind of competitive or creative displacement that concerns the Copyright Act.” Bartz, 2025 WL 1741691, at *17.

Chhabria rejects that analysis, noting that AI is not the same as a human child and that the sheer volume of competing output that an AI can create will likely cause market harm, at least in creative fields where the specific author’s voice is less important (such as news reporting).

His analysis focuses not on specific works, but on whether the entire class of AI generated works might be “indirect substitutes” for some or all of the human-created works used to train the generative models.

Notwithstanding the outcome in the case before him, he writes: “[n]o matter how transformative LLM [large language model] training may be, it's hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works that could significantly harm the market for those books.” Kadrey, 2025 WL 1752484, at *23.

The Meta and Anthropic decisions are thoughtful, well-reasoned opinions from respected judges with a deep understanding of the copyright law and a fundamental disagreement about how it should work. New York courts have tended to be somewhat more protective of artists’ rights (at least in the internet context) than those in California.

Consistent with that policy, New York state law offers some protections that judges can turn to where the federal intellectual property laws fall short. In Lehrman v. Lovo, Inc., No. 24-CV-3770 (JPO), 2025 WL 1902547 (S.D.N.Y. July 10, 2025), Oetken did just that, in the shadow of the California fair use decisions.

New York Weighs In: ‘Lehrman v. Lovo’

Plaintiffs in Lehrman, are two professional voice actors who brought a putative class action against defendant Lovo, Inc, an AI voiceover company, alleging that Lovo used AI to synthesize and sell unauthorized “clones” of their voices.

In Lehrman, plaintiffs asserted claims for violations of New York civil rights and consumer protection laws, the Lanham Act (federal trademark claims), and the Copyright Act, along with common-law contract, fraud, conversion, unjust enrichment, and unfair competition claims.

In short, plaintiffs raised every possible claim under New York state law and the federal intellectual property laws (trademark and copyright) relating to their allegation that it was improper for Lovo to make and sell “clones” of their voices without permission.

Lovo moved to dismiss. The court ultimately held that federal trademark and copyright law were not a good fit for plaintiffs’ claims, and those claims would be dismissed.

But the court also held that “claims for misappropriation of a voice” could be brought under various New York state laws designed to offer those protections and “tailored to balance the unique interests at stake.” It therefore permitted those claims to go forward. Lehrman, 2025 WL 1902547, at *1.

The underlying facts in Lehman are relatively simple. Plaintiffs are voice actors who are hired to read scripts and create recordings, which their clients use as voice-over in media such as television, movies, and video games.

They are generally paid a fee for their work which includes some combination of upfront fees, royalties, and residuals. Defendant Lovo sells a text-to-speech subscription service that allows clients to generate voice-over narrations at “a fraction of the cost of the traditional model.”

It produces its audio using an AI model known as “Generator” or “Genny”, which it claims was “created using ‘1000s of voices.’” According to Lovo, Genny is capable of creating a voice “clone,” which “refers to a virtual copy of a real person's voice.

Rather than using machine learning to synthesize an original AI voice, voice cloning technology replicates an existing human voice.” Lovo advertises its services by emphasizing how similar its cloned voices are to the originals from which they are derived.

In 2019 and 2020, Lovo solicited plaintiffs (through anonymous users on the online freelance marketplace Fiverr) to provide voice recordings.

In both cases, plaintiffs were assured that the recordings would not be used publicly: plaintiff Lehman was told the recordings were for “internal research,” and plaintiff Sage was told that the recordings were “test scripts for radio ads” which would “not be disclosed externally.”

Plaintiffs learned that their voices had been used in unanticipated ways when they heard an episode of a podcast narrated by an AI voice which they allege sounded identical to Lehrman's voice.

Upon looking into the issue, plaintiffs determined that Lovo was marketing two cloned voices under its subscription service under the stage names “Kyle Snow” (allegedly cloned from Lehrman’s voice), and “Sally Coleman” (allegedly cloned from Sage’s voice).

Lovo heavily promoted the Kyle Snow voice on its website and featured it in its software and tutorials. In marketing the Sally Coleman voice, Lovo allegedly used “side-by-side” comparisons of Sage’s original audio recordings with the cloned version.

When counsel for plaintiffs reached out to Lovo, Lovo confirmed that the “fictitious characters” Kyle Snow and Sally Coleman had been created using the voice recordings from Lehrman and Sage and agreed to take them down, but denied any wrongdoing. Plaintiffs then brought their action.

The Claims in ‘Lehrman’

The court in Lehrman clearly believes that plaintiffs have suffered some commercial harm. The AI generated “clones” of their voices, made without their consent, are being explicitly and aggressively marketed as substitutes for their personal labor. The question for the court is what legal regime best addresses that harm.

First, the court finds that plaintiffs’ state law contract claims survive dismissal. This is a relatively simple analysis: assuming for purposes of the motion to dismiss that Lovo’s agents agreed to limit use of the recordings in the ways plaintiffs allege, there is a claim here that Lovo breached that agreement.

Plaintiffs’ Lanham Act claims are more complicated, and the court spends considerable time on its analysis of them. Plaintiffs assert claims for “unfair competition and false affiliation” under the Lanham act, which the court reads as essentially ordinary trademark infringement claims under Section 43(a)(1)(a). 15 U.S.C. §1125(a).

Broadly speaking, a trademark infringement claim of this kind requires (1) misuse of the plaintiff’s distinctive “mark or dress” and (2) a likelihood of confusion between the plaintiff’s good or service and that of the defendant.

The “mark” need not be a traditional trademark such as a logo or slogan. Courts have recognized misuse of a person’s image or likeness as a basis for Lanham Act claim, and the court notes that other circuits have recognized such claims for voices as well.

Here the court holds that “there can exist a trademark-like interest in one’s image, likeness, persona, and identity. Given this, the court can discern no basis for categorically excluding voices, as opposed to images, from such protection.”

However, the court also notes that such “personal marks” have limitations not present in traditional trademarks.

Notably, it finds that the plaintiffs’ voices as used here were not “source identifying” marks: “plaintiffs’ voices may be protectable to the extent that they are being used primarily to identify the source of particular sound recordings, but are not protectable to the extent that they primarily function as content in those sound recordings.”

Because plaintiffs “have not alleged that their voices are primarily significant as brands rather than as services to which brands might be attached,” the court granted dismissal of the Section 43(a) (1)(A) claim.

The court next looks to plaintiffs’ Copyright Act claims. Here again, the court undertakes a detailed and lengthy analysis.

The Copyright law does not protect voices, but it does protect sound recordings so the issue is whether defendants used the actual recordings made by plaintiffs in any way that constitutes infringement.

Here, the court finds that Lovo’s use of copies of Sage’s recordings in its “side-by-side” comparisons constitute direct infringement, and those claims can go forward. On the AI training issue, the court finds that plaintiffs failed to “explain what training is or how it works, even at a very high level of generality.”

It therefore grants the motion to dismiss on the AI training issue, but invites plaintiffs to amend their pleading, noting that it may be “straightforward” for them to do so. In a footnote citing the Anthropic decision, the court notes that Lovo has not established a fair use defense as to AI training but may be able to do so upon amendment.

Finally, as to the AI outputs, the voice clones, the court notes that copyright statute covering sound recordings explicitly does not protect new recordings that “imitate or simulate those in the copyrighted sound recording” (17 U.S.C. §114(b)). Plaintiffs’ copyright claims as to the voice clones therefore are therefore dismissed.

Finally, the court turns to what is clearly the best fit for plaintiffs’ actual claims here: New York’s Civil Rights law. NYCRL Section 50 prohibits the use “for advertising purposes, or for the purposes of trade, the name, portrait, picture, likeness, or voice of any living person without having first obtained the written consent of such person,” and Section 51 provides a private right of action to any person whose voice is so used.

The court first addresses a statute of limitations issue, noting that Lovo’s ongoing use of voice clones of plaintiffs constitutes “republication” sufficient to refresh the limitations period.

It next holds that “digital replicas” of voices are covered by the statute, noting that, in the image context, not only digital avatars but also cartoons, composites, and “any recognizable likeness” have been held to be covered.

Finding that plaintiffs have adequately alleged that the clones are “recognizable” as their voices, that use occurred in New York, and that Lovo’s use of the voice clones constitutes “advertising and trade,” the court permits the Civil Rights law claims to go forward.

The court also addresses plaintiffs’ claims under the New York Consumer Protection law (which survive on allegations that Lovo misled its subscribers) and fraud and other common law causes of action (which fail as redundant).

In summary, the vast majority of plaintiffs’ federal intellectual property claims relating to the voice clones (whether under trademark or copyright theories) are dismissed, with only the AI training claim left open for substantial further pleadings. The specific New York state law protections, however, are more robust, and those claims survive.

What’s Next: Inputs and Outputs

Unlike the California cases, which focus on the “input” phase of the generative AI process (training the model), Lehrman is focused on the “outputs” of the models (the cloned voices themselves).

In the Anthropic case, Alsup suggests (in dicta) that the only possible claim for infringement based on output would be one in which the AI model can be prompted to produce a specific, infringing copy of the training data.

Conversely, in the Meta case, Chhabria chides plaintiffs for ignoring the market harm caused by AI’s output in the aggregate: “the potentially winning argument [] that Meta has copied their works to create a product that will likely flood the market with similar works, causing market dilution.” Kadrey, 2025 WL 1752484, at *2.

But although those two judges have strong, differing views on the “output” question, neither ruled specifically on the issue because it was not before them.

In Lehrman, Oetken plainly has these concerns in mind. AI output (the voice clones) is squarely at issue in the case, and the court wrestles at great length with the need to protect the livelihoods of creators without disrupting the balance between protection and access struck by the intellectual property laws.

Because the case involves sound recordings (which have idiosyncratic protection under the Copyright Act), it is not a perfect model for how New York courts will look at these issues going forward, but the court’s discussion of artist’s individual rights, and the policies underlying the federal trademark and copyright regimes is extremely instructive.

Here, as is often the case in New York, the court finds that the artists are protected, even with the failure of their broadest copyright and trademark claims.

This article first appeared in the New York Law Journal on July 21, 2025. Stephen M. Kramarsky, a member of Dewey Pegno & Kramarsky, focuses on complex commercial and intellectual property litigation.