Korea’s book trade spent last week asking to be given a collecting society.
Hong Young-wan, who chairs the Korean Publishers Society, told the Seoul Economic Daily today that his organisation and the Korean Publishers Association are jointly building a copyright trust body for publishing: a single counterparty to broker access to published works on behalf of authors and houses, “modeled on music copyright,” with results expected around 2027 (Seoul Economic Daily, 20 October). Four days earlier the same society issued a national appeal, No Books, No AI, signed by 158 organisations and 321 individuals, whose second demand is that an absence of objection from a rights holder “must not be treated as consent to use” (Seoul Economic Daily, 16 October). Both reports are machine-translated from Korean, which I note because I quote them.
The model is named in the demand, so it is worth looking at what the model has been doing.
KOMCA, the Korea Music Copyright Association, collected ₩437bn in 2024 and places its registrations in more than a hundred countries. From 24 March 2025 it would not register a song made with any machine involvement whatsoever: applicants confirmed their work was “created solely through human creative contributions,” and a chord suggestion built into the software disqualified it. On 3 August this year that ended. An AI-assisted song now registers and earns royalties provided the human role in the melody, lyrics, structure or arrangement was “substantial and leading” (Music Business Worldwide; Korea Herald).
Here is what an applicant now supplies:
which parts of the track involved AI; the specific AI tools used; a description of their own creative contribution; confirmation that the work was not generated from a prompt alone.
Read that as an examination and it collapses, because there is nobody on the other side of the desk. KOMCA said so itself when it imposed the ban: no technology can reliably establish whether AI was used in a piece of music. The four items are not a test. They are a question put to the one party who knows the answer, with consequences attached to the wrong one — withheld royalties, recovery of money paid, termination of the trust agreement, and a planned amendment allowing up to three times anything wrongly received. The body stopped examining the work and started pricing the reply.
Which brings me to the only figure anyone has about how well that works. In March the Board of Audit and Inspection sampled 8,540 registered songs belonging to 29 people, drawn from 81 who had each newly entrusted over two hundred works to KOMCA during 2024, and reported that 5,200 of them — 60.9% — were likely composed with machine help. It reached that figure by running the sample through an AI identification program (Seoul Economic Daily, 24 March).
So the finding that the register cannot be trusted is the output of precisely the kind of tool the register had already said cannot be trusted. The audit’s prose says likely; its headline says 60% of registered songs were composed by AI. I cannot tell which of those is nearer the truth, and neither, as far as I can establish, can anyone in Seoul. The trade is about to copy a procedure whose accuracy has never been measured by anything that procedure would accept as evidence.
Nobody has misbehaved here. A collecting society is a nineteenth-century answer to a distribution problem — too many users, too many small payments — and the art world runs the same machinery for the resale right, where a central office pays a percentage to a name on a roll without forming any opinion about the picture. Identification was assumed to be the easy part, and it was easy for a century and a half, so the bodies built then have no organ for it. Asked a question of a kind it cannot hold, the apparatus does the one available thing: it moves the question to the applicant and gets on with the paying, which is its actual function and which it performs well.
Publishing will be worse off in one specific way. KOMCA can at least ask for a project file, a revision history, a prompt log, because the music trade happens to work in software that keeps a record by accident. The book trade has produced manuscripts for five centuries without retaining a draft history, and no author’s contract I am aware of requires one. The principles eleven Korean organisations agreed in September say that “AI is merely a tool and does not have the status of an author,” and are explicitly voluntary (Korea Times, 30 September). None of it contains a way to look at a book and know.
My guess is that the Korean body arrives on schedule and functions: collects, splits, pays, balances. And that within two years somebody runs a text detector across a sample of what it has registered, publishes a percentage, and the percentage is quoted everywhere without its verb.
It will work. It will collect and distribute against the forms it is sent, and the accounts will reconcile to the last won. The question of whether a person wrote the book gets answered inside it once, in a box, by the party being paid. A trade that has read the audit and wants the body anyway wants it for the collecting — a respectable thing to want, and not what the appeal says it is for.
Sources
- Seoul Economic Daily, “Publishers Push Copyright Trust Body to Stop AI ‘Book Theft’,” 20 October 2026
- Seoul Economic Daily, “Korean Publishers Demand AI Firms Pay for Books Used in Training,” 16 October 2026
- Music Business Worldwide, “South Korea’s KOMCA ends its ban on AI-assisted songs,” 5 August 2026
- Korea Herald, “AI-assisted music now eligible for copyright registration in South Korea,” 5 August 2026
- Seoul Economic Daily, “Loopholes in Copyright Standards: ‘60% of Registered Songs Likely Composed by AI’,” 24 March 2026
- Korea Times, “Publishing industry lays down AI rules to counter ‘click-to-publish’ books,” 30 September 2026
- The Bookseller, “Over 250 UK publishers opt in to collective AI licensing scheme,” 24 June 2026
