Audio Data Is the Next AI Liability Frontier

The Voice Rights Wave

A coordinated legislative response to AI voice cloning and unauthorized audio training is now underway at both the federal and state levels. This is not one bill in one state. It is a pattern.

The federal NO FAKES Act creates a nationwide right to control the use of one's voice and likeness in AI-generated content. It establishes liability for platforms and developers that create or distribute unauthorized digital replicas, with both individual and class-action enforcement mechanisms.

Tennessee's ELVIS Act (Ensuring Likeness Voice and Image Security Act), enacted in 2024, was the first state law to explicitly protect against AI voice cloning. It extended the state's existing right of publicity to cover AI-generated voice replicas, making Tennessee the test case for how these claims will be litigated.

California AB 2602 went further. It voids broad AI consent boilerplate in entertainment contracts, establishing that consent to AI use of a performer's voice must be specific, informed, and separately negotiated. The law directly targets the practice of burying AI training rights in standard talent agreements.

Nevada NRS 597.790 defines “commercial use” to explicitly include the use of a person's voice, creating a statutory foundation for right-of-publicity claims against AI developers who train on voice recordings without consent.

These are not aspirational proposals. They are enacted statutes creating actionable liability. And they apply retroactively to audio data that was collected before the laws existed. If your organization trained an AI model on voice recordings two years ago without explicit consent for AI use, you now face potential liability under laws that did not exist when you collected the data.

Why Audio Is Different

The first wave of AI training data litigation centered on text and images. The New York Times sued OpenAI over text. Getty Images sued Stability AI over photographs. These cases are significant, but they involve relatively straightforward copyright claims. Audio introduces three additional layers of legal complexity that make the liability exposure qualitatively worse.

Biometric identity. A voice is a unique biological signature. Unlike text, which carries no inherent connection to a specific human body, a voice recording contains the biometric identity of the speaker. AI voice cloning does not merely reproduce copyrighted content; it reproduces a person. This triggers right-of-publicity claims, biometric privacy statutes (in states like Illinois under BIPA), and a distinct category of harm that copyright law was not designed to address. When an AI model is trained on a voice recording, the model does not just learn what was said. It learns who said it.

Copyright entanglement. Audio recordings frequently contain multiple copyrighted works layered on top of each other. A broadcast radio segment may include the host's spoken word (protected as a literary work and a sound recording), background music (protected under separate composition and recording copyrights licensed via ASCAP, BMI, or SESAC blanket licenses for over-the-air broadcast), and advertiser content (protected under separate commercial agreements). The blanket licenses that authorize over-the-air broadcast do not authorize reproduction of the same music in AI training datasets. Separating the licensable content from the unlicensable content requires technical infrastructure that most organizations do not have.

Consent complexity. A single audio recording may involve the rights of multiple parties: the speaker, the producer, the station or network that broadcast it, the music rights holders whose compositions appear in the background, and the advertisers whose content is interleaved. Obtaining consent for AI training use requires navigating all of these relationships simultaneously. This is fundamentally different from text, where a single author or publisher typically controls the relevant rights.

These three factors — biometric identity, copyright entanglement, and consent complexity — mean that audio data carries higher per-unit liability exposure than text or images. An organization that trains an AI model on 10,000 hours of broadcast audio is not just facing potential copyright claims. It is facing potential copyright claims, right-of-publicity claims, biometric privacy claims, and breach-of-license claims, multiplied across every speaker, every rights holder, and every piece of background content in those 10,000 hours.

The Liability Chain

Audio data moves through a supply chain with liability exposure at every node:

Originators — broadcasters, podcasters, voice actors, musicians — hold the underlying copyrights and publicity rights. They are the primary instigators of litigation. The American Federation of Musicians has sued major labels over AI licensing deals. SAG-AFTRA's 2023 contracts mandate explicit consent and additional compensation for digital replicas. Individual voice actors have filed class-action suits against AI text-to-speech companies.

Aggregators and data brokers scrape, clean, structure, and package audio datasets for licensing to AI developers. They face secondary infringement exposure and contractual liability. Standard Tech E&O policies frequently exclude intentional intellectual property misappropriation, leaving aggregators functionally uninsured under legacy frameworks.

AI developers who train models on the aggregated audio sit at the center of the liability chain. They face direct copyright infringement claims (up to $150,000 per infringed work in statutory damages), DMCA circumvention claims, and state-level right-of-publicity violations. Traditional cyber insurance does not cover deliberate mass data ingestion. If their Tech E&O policy carries an IP carve-out, they are uninsured for their core business activity.

Enterprise deployers who license generative audio AI models inherit the downstream risk. If a marketing firm generates a synthetic voice track that inadvertently replicates a recognizable voice or hallucinates a copyrighted melody, the deployer faces strict liability for direct infringement regardless of its knowledge of the training data. Sophisticated procurement departments now demand total indemnification from AI developers, but that indemnification is only as good as the developer's insurance — which, after the ISO exclusions, may not exist.

What the Courts Are Saying

Three lines of litigation are establishing the judicial framework for audio AI liability.

Suno and Udio. Universal Music Group, Sony, and Warner Records filed coordinated lawsuits against two generative audio AI companies, alleging “willful copyright infringement on an almost unimaginable scale.” Both companies admitted under discovery pressure that they used open-source tools to scrape audio directly from YouTube for training. The labels amended their complaints to add DMCA circumvention claims. Warner and BMG reached bifurcated settlements, but Sony and UMG continue to litigate. Concurrently, class-action suits filed on behalf of thousands of independent artists seek maximum statutory damages and injunctions requiring deletion of contaminated models.

Lehrman v. Lovo. Two voice actors sued an AI text-to-speech company for creating unauthorized AI replicas of their voices. The court dismissed federal Lanham Act claims but allowed New York right-of-publicity claims to proceed. This established that voice cloning creates a distinct, actionable state-law tort separate from any copyright claim over the underlying recordings. For the insurance industry, this means AI voice cloning triggers liability under both copyright and right-of-publicity theories simultaneously.

Thomson Reuters v. Ross Intelligence. Although this case involved text rather than audio, the court's ruling that AI training is not protected by fair use sent shockwaves through the insurance industry. The court rejected the “intermediate copy” defense and found that a licensing market for AI training data exists, meaning unauthorized use harms that market. For actuaries, this ruling eliminated fair use as a viable risk mitigation strategy across all data modalities, including audio.

The Insurance Market Has Already Priced It

The insurance industry does not wait for regulatory clarity. It prices risk based on loss experience and legal trajectory. The insurance market's verdict on AI audio liability is already in.

Verisk ISO's exclusionary endorsements (CG 40 47, CG 40 48, CG 35 08) have stripped AI liability from standard commercial general liability, professional liability, and products/completed operations coverage. Approximately 95% of carriers are adopting these exclusions. The “arising out of” trigger language is interpreted broadly: if a generative AI model appears anywhere in the causal chain of a media production, the exclusion applies, regardless of human editorial review.

Specialty insurers have entered the gap with affirmative AI coverage — but they impose underwriting requirements that amount to a market mandate for data certification:

• Armilla (Lloyd's coverholder, up to $25 million) requires independent AI system certification based on empirical model evaluation before binding coverage.
• Munich Re's aiSure (up to €15 million) uses a parametric structure where claims are settled based on measurable performance data, requiring verifiable telemetry.
• Testudo (Lloyd's Lab, up to $9.25 million) covers errors from AI outputs, IP infringement, and defamation, but targets organizations with demonstrated governance.
• Relm Insurance (50-state MGA) offers wrap policies designed to cover the specific exclusions left by standard policies, with governance-tiered underwriting.

The message from the insurance market is unambiguous: self-attestation of training data practices is insufficient for risk pricing. If you want AI liability coverage, you need verifiable provenance. If you cannot demonstrate where your audio training data came from and how consent was obtained, you cannot get insured. If you cannot get insured, you cannot sell to enterprise customers or government agencies that require it.

The Certification Solution

The convergence of legislative liability, judicial precedent, and insurance market requirements points to a single conclusion: organizations that use audio data for AI training need independent, third-party certification of their data practices.

Certification resolves the specific challenges that make audio data liability so acute:

• Consent documentation. A certified dataset includes verified consent records for every audio segment, specifying which downstream uses are authorized and which are not. This creates a defensible record against right-of-publicity and biometric privacy claims.
• Content separation. Certification standards require demonstration that copyrighted musical content has been identified and excluded using automated methods with defined accuracy thresholds. This addresses the copyright entanglement problem specific to broadcast and podcast audio.
• Provenance chain. A tamper-evident audit trail from source ingestion through dataset finalization provides the verifiable provenance that specialty insurers require as a condition of coverage.
• Supply chain trust. Certification marks travel with the data, allowing deployers to rely on verified data quality without conducting their own audits of every upstream vendor.

Box Commons operates this architecture today. Our BC-Certified Audio Data Standard runs a live certification pipeline across a 22-station broadcast radio network, handling stream capture, playout log reconciliation, audio fingerprinting for music exclusion, and tamper-evident audit trails. The tools are open-source and independently verifiable.

The Market Window

The organizations that build consent and provenance infrastructure now will own the premium segment of the AI audio data market. The organizations that wait will face a progressively narrowing set of options as legislative, judicial, and insurance pressures converge.

Independent broadcasters, podcast networks, and faith-based media organizations are sitting on decades of exactly the kind of natural, diverse, conversational speech that AI developers need most. That audio is currently unprotected, undocumented, and available to anyone with a web scraper. The voice rights wave is closing that window.

When the window closes, the value of audio data will be determined by whether it carries verified consent, documented provenance, and independent certification — or whether it carries unquantifiable legal risk. The premium for clean audio will be set by the insurance market, not by goodwill.

The question for media companies is not whether audio data liability is real. The courts have answered that. The insurance market has answered that. The legislatures are answering it now. The question is whether your organization will be a beneficiary of the new market for certified audio — or a defendant in the lawsuits that create it.

Content Integrity Notice: This analysis was authored by the Box Commons Policy Working Group. Generative AI was used for research synthesis and drafting support. All policy positions, recommendations, and normative claims were formulated and reviewed by human authors.