Major Labels to Send ‘Takedown Notices’ to Streaming Services for AI Soundalikes

The three major label groups have been in talks with the big music streaming services to find a way to get them to remove recordings with AI-generated vocals created to sound like popular artists, Billboard has learned. The idea under discussion with Spotify, Apple Music and Amazon Music would operate much like the one laid out by the Digital Millennium Copyright Act but would cite violations of rights of publicity, rather than copyright, according to sources at all three majors. Unlike the DMCA, however, this arrangement appears to be voluntary.

The 1998 DMCA gives online services that use, store or transmit copyrighted works a “safe harbor” from secondary liability for copyright infringement as long as they abide by a notice-and-takedown system that allows rightsholders to ask them to remove copyrighted content. That law would not apply to most AI-generated soundalike tracks because they do not infringe protected elements of copyrighted recordings or compositions but rather a trademark or a right of publicity, the protection celebrities may be able to receive to protect their names and likenesses from unauthorized commercial exploitation.

Songs that imitate the voices of big-name talent have become a trend over the past month, reaching widespread attention in mid-April when the track “Heart on My Sleeve,” which apparently used AI to mimic the style and tone of vocals by Drake and The Weeknd, was uploaded to streaming services and then swiftly removed. (The song did not credit those artists, although they were referred to in social media posts about it.)

Citing rights of publicity can be more complicated than copyright, because they are matters of state law in the U.S., backed by limited legal precedent. Rights vary by state, protections for deceased artists vary even more widely, and the use of soundalike vocals for creative purposes may in some cases be protected as free speech. Further complicating matters, these rights almost always belong to artists, not labels, which would presumably file notices on their behalf with authorization. Right now, however, this is the most obvious legal argument with which to keep AI-generated soundalikes off major streaming platforms.

In an April 26 earnings call, UMG CEO and chairman Lucian Grainge seemed to signal this approach to investors. “The recent explosive development in generative AI will… create rights issues with respect to existing copyright law, in the US and other countries, as well as laws governing trademark, name and likeness, voice impersonation, and right of publicity,” he said. “Further, we have provisions in our commercial contracts that provide additional protections.” It is not clear if takedowns issued by the majors would rely on these provisions, state law, goodwill, or some combination.

Some executives have raised concerns that AI soundalikes that imitate the voices of popular artists could result in consumer confusion. Still, a few artists like Grimes and Holly Herndon have embraced the technology, training their own AI voice models and making them available to the public.

Meanwhile, companies like Uberduck, Supertone, Lingyin Engine, and Covers.ai are marketing models with which to replicate voices. Covers.ai, which launched last week, has said that it received over 100,000 sign-ups in anticipation. Tencent Music Entertainment executives announced in November that with the company’s Lingyin Engine they had created and released over 1,000 songs containing synthetic AI voices already, one of which amassed 100 million streams.

This stance taken by the leading streaming services counters a recent announcement from the blockchain-based music platform Audius, which announced that artists can now “opt-in” to allow AI-generated works on their artist page. To organize this new music and avoid confusion, Audius would create a separate tab on the artists’ page especially for user-generated content.

Representatives for Universal, Sony, Warner, Spotify, Apple Music and Amazon Music did not respond to requests for comment.