Inside Melon's Top 100: an AI-composed Track Whose Lyricist Got Zero Royalties

Source: Hankyoreh — 'Human lyrics, AI composition' Joseon Hip-hop's 'Forest of the Mind' entered Melon's Top 100; lyricist earns zero royalties

An AI-composed track entered Melon's Top 100, South Korea's leading music streaming platform, in early September.

Its lyrics were written by a person; its composition and vocals were generated by Suno. The lyricist did not receive author's royalties on the song, and was compensated only through neighboring rights as a sound-recording producer because the composition and vocal performance were AI-generated.

Services like Suno let a user submit lyrics or a description of the desired genre and mood, and generate a full track with vocals and instrumentation. "Write a jazz piano piece for a cool, bright autumn day" is a typical prompt. Creators sometimes use the output as-is, and sometimes rework the lyrics or arrangement afterward. The producers behind this track have said that on later releases, they took a more hands-on role in arrangement and post-production.

This is not a technical gap but an attribution gap. Payment and distribution require a determination of which portion of a work constitutes human authorship, and that standard the Korean market does not yet have. The rules differ sharply by country, and whether you can copyright AI music at all depends on where you are.

Case study: outside Korea, AI-generated music is already more established. "Breaking Rust," a virtual country artist, topped Billboard's Country Digital Song Sales chart last year. "I'm Obliv," a UK-based AI music creator, signed with a US record label after one of his Suno tracks passed three million plays; he writes his own lyrics and iterates on prompts, producing up to 100 versions of a single song — a case where human judgment stays central even as AI does the generation.

Distinguishing AI-made music from human-made music is also becoming harder as the technology improves. In one study last year, roughly 9,000 people across eight countries were played two AI-generated tracks and one human-made track; only 3% correctly identified all three.

How the industry frames authorship is shifting too. In some markets, writing effective prompts is increasingly treated as a creative skill in its own right — not unlike how the arrival of computer-based music production let people make music without mastering a traditional instrument.

The Korea Music Copyright Association (KOMCA) proposed registration criteria for AI-assisted works last month — specifically, a framework for trust-registering the portions where a human substantively and directively created the work — then withdrew them 22 days later after the Ministry of Culture, Sports and Tourism conveyed its view that broader public consultation on cultural and artistic matters should come first. A ministry copyright official told Hankyoreh that "recognizing AI-assisted portions as copyrightable work, and KOMCA's own criteria for collecting and distributing royalties on trust-registered works, are separate questions" that required wider input from members and the music distribution industry before KOMCA revised its rules. The ministry is expected to hold a public forum as early as this month.

A track's composition and vocals being machine-generated does not by itself erase every claim. The written lyrics are a separate, human-authored element, and neighboring rights can attach to the sound recording produced by a human actor.

The open question is where to draw the line: which part is human authorship, and who gets paid for it. Until a standard answers that, attribution lags the technology that creates the works.

Why detection and marking are separate problems

Two different things get confused in this debate, and it is worth keeping them apart.

Detection works on a track that carries no mark at all. It looks for statistical traces left by generation models in the audio signal itself. It is probabilistic, it degrades as generation models improve, and it is a screening signal rather than proof. MuseBlossom operates MusicScanner for AI-generated audio detection of this kind.

Marking happens at creation. An imperceptible audio watermark or a C2PA Content Credential records what made the file and what was done to it, so the question never has to be inferred later. It is cheap and certain, and it only ever covers content that passed through a system that marked it. The difference between the two approaches is worked through in C2PA vs watermarking.

Neither one settles authorship. Detection tells you a model was involved; it does not tell you which share of the work a person contributed, and no technical signal can. That remains a question for the standard KOMCA and the ministry are still negotiating.

Note: detection of AI involvement identifies similarity or provenance; it is not itself a legal determination of copyright ownership or the share of human contribution.

Experience how AI-involvement detection works: https://go.musicscanner.app/story

Original 'Forest of the Mind' MV — publisher's own public upload