Your voice might not be yours, and I think builders should care
Recently I watched a video where a podcaster, Chris Williamson, found that a popular AI voice sounded exactly like him.
His friends thought it was him. The company, ElevenLabs, said the voice came from a real voice actor and that their own test showed the two voices were quite different. The voice was later taken down, and nobody ever explained why.
I couldn’t stop thinking about it, so I looked for similar cases. There are a lot.
In Japan, an anime voice actor named Kenjiro Tsuda took TikTok to court over an AI copy of his voice. The court ruled that a voice can be protected like a portrait, but it didn’t order the videos removed, and it didn’t say every AI voice resembling a real person is illegal. In New York, two voice actors say a company got their voice samples by saying they were for research, then sold AI versions of their voices. And in Illinois, a group of journalists, podcasters and voice actors sued nine tech companies, arguing their voices were used to train AI without permission. That last one is still unresolved, and I couldn’t find a public response from ElevenLabs.
And in 2024, Scarlett Johansson said OpenAI’s “Sky” voice sounded eerily like her, months after she had turned down the company’s request to voice ChatGPT. OpenAI paused the voice and said it belonged to a different actor.
Why does this keep happening?
I’ll be honest: I don’t know enough to say which reason matters most, and I don’t think there is one. I think all of these are real:
- Companies move fast, and getting permission feels like something that
- slows you down.
- The law is old and different everywhere. One of the cases people still lean on is from 1988.
- It’s genuinely hard to prove a voice was copied. Williamson couldn’t, and the company couldn’t prove otherwise.
- Voices are valuable and cheap to collect, so there’s money in not asking.
None of these is an excuse. They just stack up, and the person whose voice it is ends up with no clear way to get an answer.
What I think builders should do differently
I’m still learning, so take this as my view from the builder’s side, not as an expert.
- Ask first, by default. If a voice is going into a product, there should be a clear yes from the person. “It was publicly available” isn’t consent.
- Know where your data came from. If someone asks, “was my voice used?”, you should be able to answer. Williamson couldn’t get a clear answer partly because nobody could trace things.
- Make it easy to raise a concern. A support email and a test score you can’t interpret isn’t enough. People need a real process and an explanation, including when something is taken down.
- Don’t ship popular default voices without knowing their origin. The Archer voice was one of the most used on the platform, and the question of where it came from still hangs over it.
Where I land
It’s complicated, and I think the law should catch up, but quickly. Tsuda’s ruling shows courts are starting to take voices seriously, though it’s only one country and one case, and the American cases haven’t been decided yet. Until the rules are clearer, I think the people building these tools shouldn’t wait to be forced. If your product can imitate a human voice, treating consent as a core feature is just basic respect for the people behind the voices.