Can Your Voice Be Used to Train AI Without Your Permission?

A series of class actions in the United States is testing whether a publicly available recording of someone's voice can lawfully be used to build a voice AI model. For a software business selling into Europe, the more immediately relevant question is what European law already requires of a product that records calls, transcribes meetings, verifies speakers or generates synthetic speech, and how much of that depends on consent the business may never have collected.

What Are The US Cases About?

In May 2026, nine class actions were filed in the US District Court for the Northern District of Illinois against Adobe, Alphabet and Google, Amazon, Apple, ElevenLabs, Meta, Microsoft, NVIDIA and Samsung. The plaintiffs are Illinois journalists, podcasters, audiobook narrators and voice actors, including retired Chicago broadcasters Carol Marin and Phil Rogers. They allege that recordings of their voices were used without consent to extract voiceprints and train the foundational voice models behind products such as text-to-speech services and AI audio generation tools.

The claims rest largely on the Illinois Biometric Information Privacy Act, which expressly names voiceprints as biometric identifiers and requires written notice, a stated purpose, a retention schedule and written consent before collection. By late August 2026 the defendants had filed motions to dismiss, arguing among other things that the plaintiffs had not shown that their particular recordings were used, and Apple asked for the cases to be consolidated before a single judge. Nothing has been decided, and the outcome is likely to turn on what discovery reveals about the contents of specific training datasets.

Is A Voice Personal Data In The EU?

Yes, a recording of an identifiable person's voice is personal data under the GDPR, so processing it requires a lawful basis, a defined purpose, a retention period and transparency towards the speaker. That applies to any recording, including one obtained lawfully and one that has been publicly available for years.

The analysis changes where a business goes a step further and derives a voiceprint, meaning a mathematical representation of vocal characteristics used to identify a specific person. This is biometric data processed for the purpose of uniquely identifying a natural person, which falls under Article 9 of the GDPR and is prohibited unless a specific condition applies, most commonly explicit consent. The distinction has commercial consequences as well as technical ones. Transcribing a meeting and verifying a speaker are not the same activity in regulatory terms, even though both involve audio, and a product that begins as the first and later adds the second will have changed its compliance position without necessarily changing anything a customer would notice.

In France there is a further layer that sits outside data protection law altogether. Article 9 of the French Civil Code protects private life, and French case law has long recognised a person's image and voice as attributes of personality that cannot be exploited without their agreement. A French claimant whose voice had been cloned could therefore bring a personality rights claim independently of any GDPR complaint, and the two follow different procedures and offer different remedies. This is a point that US businesses often overlook, because there is no direct equivalent in the federal framework they are used to working with.

Does It Help That The Recordings Were Already Public?

No. There is no general principle in EU data protection law that publicly accessible personal data is free to reuse. Article 9(2)(e) does permit processing of special category data that the data subject has manifestly made public, but this is a narrow route. Publishing a podcast is a decision to share a recording rather than an evident decision to make one's biometric data available for model training.

Businesses also carry transparency obligations when they collect personal data from a source other than the individual concerned. Article 14 requires you to inform them, subject to a limited exemption where doing so would involve disproportionate effort, in which case you must take other measures to protect their interests, including making the information publicly available. Collecting audio at scale rarely removes that obligation, and usually makes it more difficult to satisfy, which is an argument for designing around it at the outset.

The CNIL has gone further than most European regulators in this area. It published guidance on using legitimate interest as the lawful basis for developing AI systems, alongside a companion document on collecting data through web scraping. The CNIL accepted that scraping publicly accessible data for AI development is not prohibited, but set out what it expects to see alongside it. That includes a documented assessment running through all three limbs of the legitimate interest test, a default list of sites the business will not scrape at all because of the nature of what they hold, observance of the technical means by which site operators signal an objection, discarding sensitive material that is not relevant to the purpose at the point of collection, and a route by which individuals can object before their data is taken rather than only afterwards. 

What Have Regulators Said About Training On Personal Data Generally?

The most consequential guidance is the European Data Protection Board's Opinion 28/2024, adopted in December 2024 at the request of the Irish supervisory authority. Three of its conclusions have commercial significance.

The first concerns anonymity. A model trained on personal data does not become anonymous simply because the training has finished, and the EDPB declined to treat that as a general rule. Each claim has to be assessed individually, and the burden falls on the controller to demonstrate that there is no realistic prospect of personal data being recovered from the model, whether by interrogating it directly or by other reasonable means. A voice model that can produce a recognisable rendition of a particular person's voice would struggle to meet that standard.

The second is that legitimate interest can serve as a lawful basis for both development and deployment, but only following a documented three-step test covering the interest pursued, the necessity of the processing and the balance against the rights of the individuals concerned, with particular weight given to their reasonable expectations.

The third is likely to matter most to businesses buying AI capability rather than building it. Where personal data was processed without a lawful basis while the model was being built, that defect does not necessarily disappear once the model is in use, and it may undermine the lawfulness of the deployment as well. A customer's exposure can therefore run backwards into how its vendor assembled its training data. A supervisory authority may ask what assessment the deploying business carried out before putting the model into use, which makes this a due diligence question rather than purely a contractual one.

What Does The AI Act Add?

Since 2nd August 2026 the AI Act's transparency rules have applied. Providers of systems that generate synthetic audio must mark outputs in a machine-readable format so that they can be detected as artificially generated, and deployers producing deepfake content must disclose it. Enforcement is shared between the Commission's AI Office and national authorities.

These rules do not affect whether a business can train on someone's voice. They govern what it must do with the output. Where a product clones or synthesises a voice, marking and disclosure are operational requirements that belong in the product itself rather than in a policy document, and they need to hold up across the ways customers actually use the feature.

What Should You Review If You Sell Voice-Enabled Software In Europe?

The first question is whether the product creates voiceprints at all, because the answer determines whether it sits in Article 9 territory or in ordinary personal data territory, and the two carry very different consent requirements and very different product design implications. A number of teams discover that they are creating something that functions as a voiceprint without having framed it that way internally.

The second concerns suppliers, and specifically whether they use the audio you send them to improve their own models. This is the term most often missed in AI vendor agreements and the one most likely to surface during a European enterprise security review. Data processing agreements should state plainly whether training on customer data is permitted, and if it is, on what basis, with what controls and with what opt-out.

The third is at the start of the chain. Where a product records meetings or calls, participants need to be told before recording begins, and the business needs a lawful basis that will survive scrutiny. In an employment context that is usually legitimate interests rather than consent, because consent given by an employee to an employer is difficult to treat as freely given. A US SaaS provider selling a meeting recorder into France will generally find that its existing US consent flow, built around one-party consent rules, does not answer the European question at all.

Where training data has come from public sources, it also helps to be able to explain its origin. Provenance has become a routine question in European enterprise procurement, and it is considerably easier to document while a product is being built than to reconstruct once a customer asks.

The American cases may take years to resolve, and their reasoning will not transfer directly, since the EU has no equivalent of per-violation statutory damages for biometric breaches and relies instead on supervisory enforcement, compensation claims under Article 82 and collective redress. 

What does carry across is the commercial premise underlying them, which is that a voice is treated as belonging to the person it came from. Businesses that can show where their audio originated will be in a considerably stronger position than those that cannot. Where a product depends on voice data whose origins are uncertain, or where it is unclear whether the product is creating biometric data at all, those are questions worth resolving before an audit or a customer security review raises them.

How Can Gerrish Legal Help?

Gerrish Legal is a dynamic digital law firm. We pride ourselves on giving high-quality and expert legal advice to our valued clients. We specialise in many aspects of digital law such as GDPR, data privacy, digital and technology law, commercial law, and intellectual property. 

We give companies the support they need to successfully and confidently run their businesses whilst complying with legal regulations without the burdens of keeping up with ever-changing digital requirements. 

We are here to help you, get in contact with us today for more information.