Award-winning journalists, book narrators, and podcasters have filed coordinated federal lawsuits against Amazon, Apple, Google, Meta, Microsoft, and Nvidia, alleging their voices were harvested without consent to build some of the most commercially valuable AI systems in history. Two Illinois statutes are the weapons of choice—and they may be exactly the right ones.
The Voices Behind the Complaint
Carol Marin has spent decades as one of Chicago’s most recognized broadcast journalists. Her reporting has earned her a Peabody Award—the most prestigious prize in broadcasting—along with multiple Emmy Awards and a reputation for the kind of documentary journalism that demands trust between a reporter and her audience. Her voice is, in a real sense, her professional identity.
Philip Rogers, a Murrow Award-winning broadcaster, has a similar relationship to his voice. The Edward R. Murrow Awards, given by the Radio Television Digital News Association, are among the highest honors in electronic journalism. They recognize not just accuracy and impact but the craft of audio storytelling—the way a broadcaster’s voice carries authority, inflection, and presence.
Alison Flowers is a podcaster and journalist whose work has been featured in national publications and whose podcast audience knows her voice as the vehicle for investigative long-form audio.
These are not anonymous plaintiffs. They are named professionals with recognizable, documented voices—people whose vocal identities have been cultivated over careers, whose voices are commercially valuable precisely because audiences recognize and trust them, and who have never licensed those voices to be fed into AI training pipelines operated by the largest technology companies in the world.
That is what the lawsuits filed in federal court in Illinois on May 13, 2026 allege happened anyway.
Six Defendants, Six Complaints, Two Core Claims
The complaints—one against each of Amazon, Apple, Google, Meta, Microsoft, and Nvidia—were filed as a coordinated wave in the Northern District of Illinois. The defendants collectively represent the full architecture of modern voice AI: Amazon’s Alexa and Polly; Apple’s Siri and on-device voice synthesis; Google’s Assistant, WaveNet, and Cloud Text-to-Speech; Meta’s voice cloning and audio AI research; Microsoft’s Azure Cognitive Services speech platform; and Nvidia’s Riva speech AI platform and the underlying hardware infrastructure that enables voice model training at scale.
The allegations in each complaint follow a common factual template, adapted to the specific defendant’s documented products and public disclosures. Two legal claims anchor all six cases.
Claim One: Illinois BIPA — Voiceprints as Biometric Identifiers
The first and most powerful claim invokes the Illinois Biometric Information Privacy Act, 740 ILCS 14/1 et seq.—the same statute that has generated billions of dollars in litigation against employers using fingerprint timekeeping and companies using facial recognition in consumer applications.
BIPA defines “biometric identifier” to include “a retina or iris scan, fingerprint, voiceprint, or scan of hand or face geometry.” The inclusion of voiceprints was not accidental. Illinois legislators in 2008 recognized that voice identification technology was already commercially viable—voiceprint systems were being used in telephone banking authentication and call center identification—and that the same privacy interests at stake in fingerprint and facial recognition applied to voice.
A voiceprint is not a recording of a voice. It is a mathematical representation of the unique acoustic and spectral characteristics of an individual’s voice—the particular combination of vocal tract anatomy, learned speech patterns, pitch distribution, and prosodic features that make one person’s voice distinguishable from every other person’s. Modern voice AI systems extract voiceprint-equivalent representations from audio recordings as a standard step in training: to build a voice synthesis model that sounds like a real person, or to build a speaker identification system, or to build a voice style transfer system, you need to extract and encode the distinctive features of that person’s voice. That process, the plaintiffs allege, is precisely what BIPA was designed to prohibit without consent.
BIPA’s requirements apply before biometric data collection: the entity collecting the data must inform the subject in writing that biometric data is being collected, state the purpose and duration of collection, and obtain a written release. No tech company sent Carol Marin a written notice before allegedly extracting voiceprint data from her recordings. None obtained her signature on a written release.
BIPA’s damages are $1,000 per negligent violation and $5,000 per intentional or reckless violation, plus attorneys’ fees, with no requirement to prove actual harm. The statute of limitations is five years for BIPA claims, meaning violations going back to 2021 are potentially recoverable. Given the scale of voice training datasets—which may include thousands of hours of audio from each professional narrator, broadcaster, or podcaster whose work was collected—the per-violation damages arithmetic can produce extraordinary aggregate figures even before class certification is considered.
Claim Two: Illinois Right of Publicity — Voice as Commercial Property
The second major claim arises under Illinois’s Right of Publicity Act, 765 ILCS 1075/1 et seq., which protects against the unauthorized use of an individual’s “identity”—defined to include “voice”—for commercial purposes.
Illinois’s right of publicity statute is among the broader such laws in the country. It prohibits any person from using an individual’s identity for commercial purposes during their lifetime and for 50 years after death, without written consent. The “commercial purpose” element is easily satisfied in the context of AI training: Amazon, Apple, Google, Meta, Microsoft, and Nvidia are building products they sell commercially. Those products—voice assistants, text-to-speech APIs, voice cloning services, voice authentication systems—are monetized at significant scale. The allegation is that the voices of professional narrators and broadcasters were used as training data to build those commercial products without the required written consent.
Unlike BIPA, the right of publicity claim does not depend on the technical question of whether a “voiceprint” was extracted. It depends on the simpler question of whether the plaintiff’s voice was used for commercial purposes without consent—a question that may be easier to establish once discovery produces evidence of what training datasets the defendants used.
The right of publicity claim also supports a disgorgement theory: if a company built a valuable commercial product on the unauthorized use of identifiable voices, the injured parties may be entitled not just to statutory damages but to a share of the unjust enrichment that resulted. This theory is particularly potent in the context of commercial voice AI, where the market for text-to-speech and voice synthesis services runs into billions of dollars annually.
The Provenance Problem: Filing on “Information and Belief”
One of the most legally significant features of all six complaints is that the plaintiffs acknowledge, explicitly, that they are alleging their voices were used in training “based on information and belief”—because none of the defendant companies have publicly disclosed the specific sources of their voice training data.
This is not a weakness in the complaints. It is a litigation strategy.
The opacity of AI training data provenance is itself well-documented. The complaints against Meta, for instance, allege that while the company has not disclosed its complete training data, the plaintiffs’ vocal profiles match the “profile” of training data that Meta described in public stock filings, investor presentations, and other public documents. This approach—inferring data usage from public disclosures about the scope and characteristics of training datasets, rather than from direct knowledge of training data contents—has been used with increasing sophistication in AI training data litigation.
The practical consequence is that the critical evidentiary question—whether Carol Marin’s recordings actually appear in Amazon’s or Apple’s or Google’s voice training datasets—will be resolved through discovery, not pleading. The plaintiffs need only allege sufficient facts to survive a motion to dismiss; the burden of establishing exactly what training data was used falls on the defendants’ document production and the plaintiffs’ forensic analysis of that production.
This creates a significant tactical problem for the defendants: to defend against the BIPA and right of publicity claims on the merits, they will likely need to produce detailed documentation of their voice training datasets and methodology—precisely the information they have kept confidential and that would, if disclosed, reveal the full scope of any alleged violations. The choice between defending on the merits and protecting training data confidentiality is uncomfortable.
Why Illinois? The Strategic Jurisdiction Choice
The plaintiffs’ choice to file in Illinois—rather than California, where most of the defendant companies are headquartered, or in any of several other potentially available jurisdictions—reflects sophisticated jurisdictional strategy.
Illinois is unique in the strength and specificity of its biometric privacy regime. BIPA has no equivalent in California, Texas, Washington, or any other state that has enacted biometric privacy laws in terms of the breadth of its statutory damages structure and the private right of action it creates. Chicago is also the home district for several of the named plaintiffs (Carol Marin and Philip Rogers are Chicago broadcast journalists), providing both personal jurisdiction and a plausible venue argument.
Federal court in the Northern District of Illinois gives the plaintiffs access to a judicial ecosystem with substantial experience in class action litigation and in BIPA specifically. The Northern District has been the venue for some of the most significant BIPA cases in the statute’s history, including early facial recognition cases and the ongoing wave of workplace biometric litigation. Judges in that district are familiar with BIPA’s technical requirements and less likely to be swayed by defendants’ arguments that the statute was never intended to apply to AI training.
The decision to name Illinois-law claims also creates a class definition opportunity: the plaintiffs can define a class of Illinois residents whose biometric data was collected without consent—a class that could include many other voice professionals, audiobook narrators, news broadcasters, and podcast hosts whose work was recorded in Illinois and potentially incorporated into AI training datasets.
The Broader Context: AI Training Data Litigation at Scale
The voice actor lawsuits join a rapidly expanding body of AI training data litigation that has targeted copyright, privacy, and publicity rights simultaneously.
On the copyright front, authors, news organizations, and music companies have filed dozens of cases alleging that LLMs and image generation models were trained on copyrighted content without authorization. The New York Times v. Microsoft/OpenAI case, filed in late 2023, remains the highest-profile copyright dispute in AI training litigation, with the Times alleging that its journalism was used to train GPT-4 and other models. Similar cases have been filed by a coalition of authors (including George R.R. Martin and John Grisham), by music publishers against AI music generation companies, and by visual artists against image generation platforms.
On the privacy and identity front, the voice actor lawsuits represent something distinct: they are not copyright cases (professional narrators and broadcasters may not own the copyright to their recordings, which are often owned by publishers, broadcasters, or studios), but they don’t need to be. BIPA and the right of publicity provide independent legal hooks that don’t depend on copyright ownership. A narrator who recorded an audiobook for Penguin Random House doesn’t own the copyright to that recording—but they own their voice, and BIPA protects their voiceprint regardless of who owns the copyright to the underlying recording.
This is a significant development in the AI litigation landscape. It means that the voice AI cases are not subject to the same fair use defenses that dominate copyright-based AI training litigation. Copyright defendants have argued, with some success, that training an AI on copyrighted content is transformative use or that the outputs are sufficiently distinct to avoid infringement. BIPA does not have a fair use exception. The right of publicity does not have a transformative use exception that could plausibly cover systematic extraction of voice data for commercial AI training. The plaintiffs in these cases have chosen their legal theories with care.
The Defendants’ Exposure, Company by Company
Amazon
Amazon’s exposure in this litigation reflects the scale of its voice AI business. Alexa, deployed in hundreds of millions of devices, is trained on enormous volumes of voice data. Amazon Polly, its commercial text-to-speech service, generates lifelike synthetic voices for applications ranging from e-commerce to audiobook narration. The complaint against Amazon alleges that it “built a global voice-AI business on the voices of real people”—a characterization that Amazon’s own marketing materials arguably support, given the company’s public statements about the quality and naturalness of its AI-generated voices.
Amazon also operates Audible, the dominant audiobook platform, which holds one of the largest libraries of professional narration recordings in existence. The question of whether Audible’s audio content was used to train Amazon’s voice AI systems—and whether any such use required consent from the narrators under BIPA and right of publicity law—is among the most consequential factual questions in the litigation.
Apple
Apple’s complaint centers on the allegation that it “built its commercial voice synthesis technology by extracting voiceprints from recordings of real people.” Apple’s voice AI products include Siri, the Personal Voice accessibility feature (which creates a synthetic voice that sounds like the user), and text-to-speech capabilities across its device ecosystem. Apple has been relatively opaque about the specific data used to train its voice systems, making the provenance discovery question particularly acute in its case.
Apple’s Personal Voice feature—which explicitly creates a personalized voice clone from user recordings—illustrates the company’s deep investment in individualized voice synthesis technology. The training approaches and data pipelines that enable Personal Voice may have required large-scale voice data collection and analysis.
Google’s voice AI portfolio is among the broadest of any defendant. WaveNet, developed by DeepMind and integrated into Google’s products, was a landmark advance in neural text-to-speech synthesis when it was introduced in 2016 and has been continuously refined since. Google Cloud Text-to-Speech is a commercial API offering high-quality synthetic voices to enterprise customers. Google Assistant deploys voice synthesis at massive consumer scale. Google’s YouTube platform, meanwhile, holds one of the largest repositories of professionally produced audio content—news broadcasts, podcasts, educational content, documentary narration—in existence.
The YouTube dimension is particularly significant for the broadcasters and podcasters named as plaintiffs. YouTube’s terms of service grant Google a broad license to use uploaded content for platform purposes, but that license language was written before the era of AI training data extraction and may not encompass the use of voice recordings as biometric training data under Illinois law, regardless of what the platform’s terms say.
Meta
The Meta complaint raises the training data provenance issue most directly, alleging that while Meta has not disclosed the provenance of its training data, the plaintiffs’ voice profiles match the characteristics Meta has publicly described. Meta’s voice AI investments include voice cloning research, speech synthesis in its Ray-Ban smart glasses platform, and audio AI capabilities embedded in its broader generative AI strategy. Meta’s public filings and technical papers describe the characteristics of training datasets used for voice-related models, and the plaintiffs have apparently used those descriptions as the basis for inferring that their own recordings were included.
Microsoft
Microsoft’s exposure centers on Azure Cognitive Services’ speech platform, which offers text-to-speech, speech-to-text, and voice customization capabilities to enterprise customers, as well as voice synthesis features integrated into Microsoft 365, Teams, and other products. Microsoft is also a major investor in OpenAI, whose voice capabilities—including the ChatGPT voice feature and the Whisper speech recognition model—represent some of the most widely deployed voice AI in existence. The boundary between Microsoft’s independent voice AI liability and its exposure through OpenAI products is a question that may be litigated both in this case and separately.
Nvidia
Nvidia is the most structurally distinct defendant. It does not operate a consumer voice assistant or sell a text-to-speech API. What Nvidia offers, through its Riva platform and related products, is the infrastructure and tooling that enables others to train voice AI models. Riva provides pre-trained voice models, fine-tuning pipelines, and deployment infrastructure. The claim against Nvidia is that by providing the platform on which voice models are trained—using data that allegedly includes the plaintiffs’ voices without consent—it participated in the BIPA and right of publicity violations as a party that enabled and profited from the unauthorized collection.
This is a novel theory in the AI training data context, and its success or failure in the Nvidia case will have significant implications for the liability exposure of AI infrastructure providers more broadly. If Nvidia can be held liable under BIPA for providing the platform on which voice models are trained, the same theory potentially applies to cloud providers, hardware manufacturers, and other infrastructure layers.
What the Litigation Could Resolve—and What It Can’t
The voice actor lawsuits, if they proceed to class certification and eventually to trial or significant settlement, could establish several important precedents:
That BIPA applies to AI training data collection from audio recordings. The application of BIPA’s voiceprint provisions to AI training has not been definitively adjudicated. A court ruling that training an AI voice model on professionally recorded audio constitutes the collection of a voiceprint under BIPA would be a landmark decision with implications far beyond these specific defendants.
That right of publicity claims survive AI fair use defenses. The scope of fair use and transformative use defenses in the context of AI voice training is genuinely unsettled. These cases will force courts to determine whether the commercial use of a broadcaster’s voice recordings to train a synthetic voice system—without consent, and for the purpose of generating revenue from voice AI products—falls within the right of publicity or outside it.
What discovery of AI training datasets looks like. The procedural question of how plaintiffs can obtain and examine AI training data in discovery is unresolved. The defendants will almost certainly assert trade secret protections over their training data. Courts will need to develop frameworks for allowing plaintiffs to verify the presence or absence of their data in training sets while protecting legitimate confidentiality interests.
What the lawsuits cannot resolve—regardless of outcome—is the broader policy question of how the law should treat the relationship between professional voice work, AI training, and consent. Even a complete plaintiff victory in all six cases would not create a statutory framework governing AI voice training nationally. That work belongs to legislators, and the pace of AI voice technology development is already well ahead of any legislative response currently in progress.
The Larger Stakes: Voice as the Last Personal Frontier
There is something qualitatively different about voice in the AI data rights conversation that the framing of BIPA and right of publicity claims only partially captures.
A voice is not just a data point. It is the primary medium through which a broadcaster like Carol Marin or Philip Rogers has built their professional credibility over decades—the particular combination of authority, warmth, precision, and presence that audiences trust and return to. When that voice is extracted from recordings, processed into training data, and used to build a synthetic voice system that can generate similar-sounding audio on demand, the injury is not just technical or legal. It is the appropriation of something that took a career to build, repurposed without permission into a commercial product that can now produce that broadcaster’s “voice” for any purpose, at any scale, at effectively zero marginal cost.
The right of publicity was created precisely to address this kind of appropriation—the commercial use of a person’s identity for someone else’s gain without consent or compensation. The voice actor lawsuits are, in that sense, a straightforward application of a well-established legal principle to a new technological context.
What makes them landmark cases is not their novelty but their scale, their defendants, and their timing. Six of the most powerful technology companies in the world are being asked, in federal court, under some of the strongest consumer protection statutes in the United States, to answer for what they did with the voices of the people who trusted those voices to journalism, to audiobooks, to podcasting—and who never agreed to have those voices become the raw material for commercial AI.
The answer they give, through litigation or settlement or legislation, will shape the terms under which voice AI develops for the next generation.
