Overview
ElevenLabs sells synthetic speech. Its products cover text to speech, speech to speech, dubbing, voice cloning from a short recording, and a conversational agents platform that answers live phone calls. Mati Staniszewski is co-founder and chief executive.
The money arrived fast. A $500 million Series D led by Sequoia closed in February 2026 at an $11 billion valuation, against $3.3 billion in the January 2025 Series C. Revenue figures differ by source and none are audited: coverage put annual recurring revenue above $330 million during 2025, while Sacra estimated roughly $500 million by April 2026 and reckons the platform is used inside 41% of the Fortune 500. Revolut and Klarna both run ElevenLabs voice agents on customer phone support.
The voice library now holds more than 10,000 voices, and the company says it has paid out over $11 million in creator rewards to the people who supplied them. In November 2025 it launched an Iconic Marketplace, where rights holders license named voices and ElevenLabs brokers the request. Michael Caine and Liza Minnelli signed on as two of the few living participants, alongside estates of dead figures including Maya Angelou and Judy Garland. ElevenLabs released an AI-narrated version of The Odyssey in 2026 using a recreation of Caine’s voice.
What It Knows About You
The thing it takes is your voice, and voice is a biometric identifier in a way that a typed prompt is not. ElevenLabs is fairly direct about this. Its privacy policy says that for some services it may ask you to record yourself so it can confirm an uploaded voice is yours, and that handling your voice this way may mean handling biometric or special category data under some laws. Its EU-US Data Privacy Framework policy repeats the point.
What the policy does not give is a number. Retention is described as for as long as necessary to provide the services or as law requires, after which data is deleted or de-identified. There is no published retention period for a cloned voice model, which is the asset that matters.
Training defaults split by tier, and the split is sharp. On the Free, Creator, Pro and Scale plans you are opted into model improvement unless you go into the data use settings and turn it off, and the opt-out is prospective: audio already submitted can stay in training sets. Enterprise is the reverse. ElevenLabs says it does not train on Enterprise customer data by default, and sells SOC 2 Type II, HIPAA with a signed agreement, GDPR terms, European data residency and an optional Zero Retention Mode on that tier. If you care about any of this, the paid enterprise contract is where the controls actually live, and the consumer plan most people sign up on is not it.
One detail in the documentation deserves more attention than it gets: cloned voices cannot be exported as files. They stay in the ElevenLabs account.
The Real Risks
A live biometric claim. On May 11, 2026, seven Illinois residents sued the company in federal court in Chicago as Amer v. Eleven Labs Inc., part of a wave of filings that also named Adobe, Samsung and several large platforms. The complaint alleges ElevenLabs took voiceprints from third-party platforms without written notice or release, extracted biometric signatures from recordings made for journalism and performance, and built them into commercial models. Illinois law sets liquidated damages of $1,000 per negligent violation and $5,000 per reckless or intentional one. The American Bar Association’s class action newsletter called the May 2026 filings a potentially transformative turn in AI privacy litigation. No ruling has issued, so this is exposure rather than a finding, but it is the clearest live legal test of whether scraped voice counts as a regulated biometric.
Other people using it as you. In January 2024, New Hampshire voters got a robocall impersonating President Biden telling them not to vote in the primary; the state attorney general confirmed it was fake. Pindrop compared the audio against more than 120 synthesis systems and put the likelihood it came from ElevenLabs above 99%, and WIRED reported that two independent expert teams agreed. ElevenLabs banned the account and added automatic blocking of politicians’ voices. In December 2024, Recorded Future assessed that commercial voice tools, ElevenLabs among them, were very likely behind voiceovers in a Russian-linked influence campaign tied to the Social Design Agency, spanning over 1,190 videos; the researchers judged the actual impact minimal, and the network kept posting German-language material into the 2025 German election period.
Scrutiny of the guardrails themselves. The Senate Joint Economic Committee wrote to Staniszewski on April 16, 2026, prompted by an FBI figure of $893 million in AI voice scam losses and a New Hampshire grandparent scam using a cloned relative’s voice. The letter asked how the company tracks attempts to clone no-go voices, what its real-time detection success rate is, how it protects minors’ voices, and whether it can identify audio its own tools produced. ElevenLabs told Axios it has an extensive array of safeguards, including blocking celebrity and public-figure cloning plus automated and human review. It has not published a catch rate, and much of the stack is moderation after generation.
Paid voice work is being substituted now. This is the score that drives the grade. Routine narration, phone trees, e-learning and low-budget advertising are the exact jobs synthetic speech does acceptably well and far cheaper, and the company markets cloning to performers as a way to earn without recording. The licensing side is real and better than nothing: $11 million in creator rewards, consent-based marketplace deals, estates deciding terms. Set against a global voice industry, it is a rounding error, and the money concentrates on famous voices rather than the working narrators whose rates are collapsing.
On bias the honest answer is weaker than the brand of this site might suggest. The one controlled study we found, Michel et al. at ACM FAccT 2025, tested English accents across ElevenLabs and Speechify and found no statistically significant quality difference between accents for ElevenLabs. Stock voices scored below excellent for every accent, and listeners frequently heard African, Indian and Australian voices as American or British, which is its own kind of exclusion. Claims of accent bleed in non-English output come mostly from vendor blogs and single bug reports, so we have not scored them.
Alternatives
- Record it yourself. A $70 USB microphone and Audacity cost nothing per minute and upload nothing. For narration you will reuse, this is frequently faster than iterating on prompts, and the output is unambiguously yours.
- Hire a voice actor on a contract that names AI use. California now requires contracts to describe digital replica usage explicitly. A clause covering whether a recording can be used to train or clone is worth more than any platform setting.
- Operating system speech. macOS, Windows, iOS and Android all ship text to speech that runs on the device for accessibility and draft read-throughs. Quality is lower than ElevenLabs. Nothing leaves the machine.
- Other hosted vendors such as Amazon Polly, Azure AI Speech or Google Cloud Text-to-Speech, if what you need is stock synthetic narration under an existing cloud contract rather than a clone of a specific human.
- ElevenLabs Enterprise with Zero Retention Mode. If the capability is what you need, this is the tier where training is off by default and retention is contractual. Note the trade-off the company states itself: under Zero Retention Mode it cannot diagnose problems with your generations.
Our Verdict
D. ElevenLabs is the best product in its category and that is most of the problem. The technology works well enough that it displaces paid voice work today rather than eventually, and the consent apparatus bolted onto it, marketplace licensing and creator rewards, pays a small number of recognisable voices while the ordinary narration market reprices underneath them.
The privacy score of 7 rests on specifics rather than atmosphere: training on by default for every consumer tier with a prospective-only opt-out, no published retention period for cloned voice models, and an active federal biometric class action filed in May 2026. Lock-in at 7 rests on a single documented fact, that voice clones cannot be exported and remain in the account. Autonomy and bias sit at 4 because the evidence there is thin and we are not going to inflate it.
Four changes would move this rating. Publish retention periods for cloned voice models. Make training opt-in on consumer plans, as it already is for Enterprise. Let people export or verifiably delete a voice model they created. Publish how often no-go voice blocking actually stops an attempt, which is the question the Senate committee asked in April 2026 and the one the company has answered with adjectives. A dismissal or a defence verdict in the Illinois case would matter too, and so would a loss.