InterviewKorean startup targets 'legal AI music' via copyright tracking

More than half of the new tracks uploaded to French streaming service Deezer each day are now made by artificial intelligence (AI). The company said such uploads peaked in June at about 90,000 a day, up from roughly 10,000 in January 2025.

The record industry's response is shifting from the courtroom to the negotiating table. Universal Music Group and Warner Music Group, which sued AI music generators Suno and Udio in 2024 alongside Sony Music, have since settled with Udio. Warner has also settled with Suno.

What the industry still lacks is a way to pay the people whose music feeds these systems.

Streaming runs on a simple formula, in which a song is played, counted and paid for, but nothing comparable exists for AI-generated songs that draw on existing ones. For now, platforms are simply keeping fully AI-generated tracks out of royalty payments.

These are not distant problems for Korea.

K-pop reaches much of its global audience through the same streaming services, and at home, AI-assisted tracks have recently climbed the charts on Melon and YouTube, even though major streaming platforms here do not ask distributors whether AI was used.

In August, the Korea Music Copyright Association, the country's largest songwriters' royalty collection body, introduced Korea's first standards for registering songs made with the help of AI, only to withdraw them 22 days later. There is still no concrete standard for how royalties from AI-assisted music should be collected and distributed.

Lee Jong-pil, CEO of Seoul-based music tech startup Neutune, says the technology to close that gap is already here. When AI produces a new song, it is now possible to work out which recordings went into it — one source track might have lent its explosive energy, another its guitar tone and a third its drum pattern, taken from between the 30- and 50-second marks.

"In my view, the technology is already settled," Lee told The Korea Times in an interview at the company's office in Gangnam District, Seoul, Oct.8. "The slow part is getting consent from rights holders, and it can't help but be slow."

The obstacle, he said, resembles the industry's shift from CDs to digital downloads, when rights built around physical records had to be renegotiated with creators. AI use needs a similar opt-in or blanket licensing framework.

Founded in 2020 by researchers from Korea Advanced Institute of Science and Technology's Graduate School of Culture Technology, Neutune develops AI attribution technology, which identifies the original works behind AI-generated music and measures how much each contributed. Its products include Music DNA, a service that scans new music for material drawn from existing works, and MixAudio, an AI remix platform with attribution built in.

Blocks, not tracks

Neutune's approach rests on breaking music into what it calls "blocks."

A song is split by instrument, such as bass and drums, and by section, such as the intro and verse, creating a grid of small units that can be pulled apart and recombined, much as video game soundtracks are assembled on the fly.

"It's like Lego," Lee said. "If you build a phone as one solid piece, it can only ever be that phone. With a Lego set, you can build a phone or something else entirely."

That level of detail matters for payment, because users no longer just "click a button" for a three-minute song, according to Lee. They generate a track, take it apart and rework individual parts.

“If all we can say about a newly generated song is that songs A, B and C were used in its making, accurate settlement isn't possible,” Lee said.

Licensing deals between labels and AI firms mostly cover whether a song was used to train a model or to make a remix, he noted. Checking whether a finished track copies existing works requires a separate monitoring system that scans new output against a vast music database, much as Shazam identifies songs. That is the market Music DNA targets.

The technology still has gaps. Copyright in the music itself is relatively easy to track, Lee said, but rights tied to a person, such as a singer's voice and likeness, are harder to protect and measure. He expects that area to be sorted out around next year.

Legal or illegal AI

One obstacle is that composers are still reluctant to admit using AI.

The reason, Lee said, is not AI itself but platforms like Suno, which record labels have sued over its training data.

“If [the platforms] were clean services that had resolved their copyright issues, there would be nothing to hide,” he said, likening it to using a cracked copy of Microsoft Word — something no one would proudly admit to.

"I don't think the issue is AI versus humans," Lee said. "It's legal AI versus illegal AI."

For the same reason, he sees requiring composers to declare their AI use as a stopgap.

Overseas, standards are being built so that music software automatically records how a work was generated, as is already common for images. Once that spreads to audio, he said, the information will simply be embedded in the file.

Lee believes Korea is well placed in AI music, since Korean consumers are among the world's most advanced when it comes to fandom and adopting AI.

Lee sees the same block-based logic shaping K-pop's next stage of growth.

The industry has grown by crafting every piece of content around its artists, an approach he said does not scale indefinitely. The next step, in his view, is a "remix market in which fans can use artists' work through a platform," sharing revenue or paying for it.

While many in the industry remain cautious about AI, Lee said some of K-pop's most prominent figures are already paying close attention.

"Whether it's for the K-pop industry or for us, I want what we're pursuing to grow here first," he said. "I just want to do well with K-pop."