ElevenLabs

ElevenLabs was founded in 2022 by two developers from Poland - one a former Google engineer, the other coming from Palantir, a data analytics company. The story they tell themselves is simple: they watched American films badly dubbed in Poland and got fed up with the flat, foreign voices laid over the actors' lips. They decided to build a model that speaks like it's alive - with intonation, with breathing, with a pause in the right place, not just reading text out loud.

In practice the product is several things in one. You write text, the system reads it out in a chosen voice, and it sounds like a live person, not a GPS narrator. You give it a short recording of someone's voice, it learns it, and can then say any sentence in that voice. And the most impressive part - it takes a video in one language and returns it in another, keeping the speaker's voice, timbre and even emotion, instead of replacing it with a foreign dubbing actor.

Studios use it to dub films and shows in a fraction of the time and budget of a classic dubbing shop. Podcast and audiobook creators use it when they don't have the budget for a studio with an actor. Companies use it to put a voice assistant on the phone line instead of a live person. And this is where the company turns into infrastructure: it doesn't make content, it just supplies the voice that others build on.

Cloning is the part that should worry you. It doesn't need a studio, doesn't need weeks of recording - a short audio file, taken from anywhere, is enough to produce a synthetic voice that passes for real to a human ear. There have already been cases of fake calls from a 'relative in trouble' and fake voice messages from politicians, made with exactly these kinds of tools. ElevenLabs has filters and consent rules for the voice's owner, but once a technology is out, copies pass through looser clones with no filters at all.

On October 9, 2026 the company set off from the other side too. In ElevenAgents, its platform for agents, it released synthetic voice detection: the system listens to the first few seconds of a call and says whether a human or an AI-generated voice is speaking. What happens next is up to the company's rules; for example, a synthetic voice asking for a bank account change or a password reset gets stopped right at the start. All ElevenLabs audio carries a watermark, but audio from other sources often doesn't, so the detector doesn't rely on the mark. How often it gets it wrong, the announcement doesn't say. For now the feature is available only to the enterprise customers its Forward Deployed Engineering team works with; a broader group of enterprise customers is promised later in October. The same day ElevenLabs joined the Personal Agent Protocol as a design partner, the group writing how an agent tells a company on its own who it is and whom it represents.

A voice has always been proof - hear someone speak, and you believe it's them. ElevenLabs is the technology that, for the first time, makes that proof forgeable.

Here's where the class shows

As someone who's spent thirty years listening to what a brand sells and what trust sells, I look at ElevenLabs with two hands at once. With one hand, admiration, because the tool is genuinely useful and opens doors for people with no studio budget. With the other, concern, because we've spent decades building the instinct 'I know his voice, so it's him', and now that instinct no longer holds automatically.

The shock isn't in the company itself - it keeps rules. The shock is that the technology is now accessible enough for someone with no rules to pick it up. When you hear a voice on the phone from now on, it's worth asking something only the real person would know, before you believe it blindly.

The visual is generated code-art, no third-party images.