Holly+ gave a voice clone a constitution. The consent tools that followed settled for an opt-out box

Most AI voice clones of musicians arrived without asking. Holly+ was the rare one that started from the artist's side. On 13 July 2021 Holly Herndon announced a "digital vocal twin" that anyone could use: upload up to five minutes of polyphonic audio and get it back sung in her processed voice. Her studio still files it as a 2021 "protocol", a machine-learning instrument "with a shared governance model for artist identity". Five years later, that governance idea matters more than the tool, because most of the consent technology built since went the other way.

What was Holly+, technically?

The model was built by Never Before Heard Sounds, the company founded by Chris Deaner and Yotam Mann, and trained on several hours of Herndon's isolated vocal stems, according to her launch post. It kept the pitch and rhythm of whatever you uploaded and added the timbre and texture it had learned from her voice. Uploads were not retained; outputs were archived so people could share them. Users could do what they liked with the results, with credit appreciated.

That openness was the point. Herndon wasn't trying to stop people making music with her voice. She was trying to decide who could call that music an official Holly Herndon work.

How was the voice governed?

The launch post set out a plan for a Holly+ DAO on the OpenLaw Tribute framework. Herndon would airdrop ERC-20 "VOICE" tokens to collectors, friends, family and selected artists. Token holders would vote on new tools and on which submitted works counted as approved uses of the voice. Approved works would be minted as NFTs, with initial and secondary sales split 50% to the creator, 40% to DAO members and 10% to Herndon.

Her reasoning was about identity, not just income. She cited Midler v. Ford and Waits v. Frito-Lay, two US cases that protected singers' voices from commercial imitation, and noted that no case about a deep-learning voice model had yet reached the courts. She also argued that shared stewardship would protect the voice's reputation after her death better than handing the rights to someone less familiar with her work.

The terms have not stayed fixed in public. In a 2024 Art Basel interview, Mat Dryhurst described the model as shared under a protocol that returns 50% of profits from media made with the voice to Herndon, a different split from the 2021 plan. He also noted that a voice can't be copyrighted, though publicity and personality rights apply. We couldn't find a public document that reconciles the two descriptions, or a public record of the DAO's votes and payouts, and holly.plus currently shows nothing but the project's name.

What did Spawning add?

In September 2022 Herndon and Dryhurst co-founded Spawning with Jordan Meyer and Patrick Hoepner, Inverse reported. Its first tools were Have I Been Trained, which let artists search the 5.8 billion images in the LAION-5B dataset for their own work, and Source+, a beta standard for opting in or out of AI training. Dryhurst told Inverse, "I believe more will ultimately opt in than out."

Stability AI said it would honour opt-outs made through the tool. By March 2023 more than 80 million artworks had been opted out, but Sayash Kapoor and Arvind Narayanan found that only about 40,000 came from individual artists; the rest were bulk opt-outs from platforms such as Shutterstock and ArtStation. They also pointed out that no law required developers to respect opt-outs. TechCrunch reported the total had passed 1 billion by late April 2023, that Spawning had raised a $3 million seed round led by True Ventures, and that its planned API would cover audio as well as images. In the 2024 Art Basel interview Herndon described Spawning's next step: base models trained on public-domain data, then fine-tuned on datasets owned by people who can earn from them.

When we checked spawning.ai in October 2026, it showed only a Source.Plus header and an "Under Maintenance" notice.

How does it compare with Grimes and Imogen Heap?

Two other artists built their own answers. In spring 2023 Grimes launched Elf.Tech with the studio CreateSafe, Music Business Worldwide reported, offering a "GrimesAI-1" voiceprint in exchange for 50% of master royalties on tracks distributed through her team. Imogen Heap turned down commercial offers over their terms and trained her own model, ai.mogen, with an audio engineer on recordings from her career, MusicTech reported. She told the magazine she would release it eventually, "but I don't want my voice to say hateful things."

Grimes ran a royalty deal, and Heap kept the gate closed herself. Holly+ was the only one of the three to try handing the gate to a group of people.

What we think

Holly+ got the order right. Consent came first, the terms were published on day one, and anyone could use the voice while official status had to be earned by a vote. Spawning's opt-out tools were a pragmatic response to models already trained on scraped work, but they flip that order: the default is yes, and the artist has to find their work and object. The numbers show who can bear that cost. Platforms filed tens of millions of opt-outs; individual artists filed tens of thousands. A working musician with no catalogue team is never going to opt out at scale.

The second lesson is fragility. Both projects depended on a small team keeping a website and a governance process alive, and neither the Holly+ DAO's record nor Spawning's public tools are easy to find today. Consent infrastructure that lives on one startup's server isn't infrastructure. That's why the most useful legacy of Holly+ may be legal rather than technical. The NO FAKES Act would make a voice licensable but not assignable during the artist's life, and Denmark's bill would make whoever posts a clone prove consent. Both put the burden where Holly+ did, on the person using the voice.

The losers in an opt-out world are mid-career artists, the very group Spawning said it was built for. The winners are artists with enough leverage to set terms themselves, as Herndon, Grimes and Heap each did. Everyone else needs the default changed by law, and needs tools that leave them in control of what they make. Veena, for example, says everything you make with it "stays yours". Until that default shifts, the opt-out box is a fallback, not an answer.

Sources

• Holly Herndon, Holly+ (13 Jul 2021)
• Herndon Dryhurst Studio, Holly+
• Holly+, holly.plus
• Art Basel, AI choirs: Holly Herndon and Mat Dryhurst on data training as art-making (2 Oct 2024)
• Inverse, This couple is launching an organization to protect artists in the AI era (14 Sep 2022)
• Sayash Kapoor and Arvind Narayanan, Artists can now opt out of generative AI. It's not enough. (9 Mar 2023)
• TechCrunch, Spawning lays out plans for letting creators opt out of generative AI training (3 May 2023)
• Spawning, spawning.ai
• Music Business Worldwide, Grimes launches AI project in partnership with CreateSafe (3 May 2023)
• MusicTech, Imogen Heap on making an AI voice model (25 Jun 2024)
• Veena Studio, veena.studio