Every rights conversation now has a new party at the table. It does not sign contracts or attend book fairs, but it has read — or wants to read — almost everything. The large language models behind today's AI were trained on vast bodies of text and image, and a great deal of that material was made by the authors and artists that agencies like ours represent. What that means for copyright is still being worked out, in courtrooms and negotiating rooms at the same time. It is worth setting out plainly where things stand.

Start with the question that dominated the headlines: is it lawful to train an AI on copyrighted work? Across 2025 and into 2026, courts in the United States began to draw a line — but not the one either side expected. In the most closely watched book case, a group of authors sued Anthropic; the court found that training a model on works the company had lawfully acquired could be a transformative, fair use, but that downloading and storing pirated copies to do it was not. The case then settled for roughly 1.5 billion dollars — an amount tied directly to the pirated sourcing, with reported payouts of around 3,000 dollars per work. A parallel suit against Meta reasoned similarly: training on the books could be defended as fair use, while claims tied to how the files were obtained were allowed to proceed. Around these sit a whole field of disputes — from the authors' class actions naming writers such as Sarah Silverman, to Getty Images against Stability AI over image training, to The New York Times against OpenAI. These rulings are narrow, several are under appeal, and courts elsewhere may reason differently. But the signal is clear: the decisive question turned out to be not whether a machine may learn from a work, but how that work was obtained. Provenance is now the threshold that matters.

That threshold does not sit at the same height everywhere — and nowhere is the contrast sharper than within Asia. Japan's copyright law contains a broad exception permitting works to be used for information analysis, which extends to much AI training and has made the country notably permissive. Korea has moved the other way. After a wave of creator protest, the Korea Copyright Commission issued guidance in late 2023 urging AI developers to seek permission from rights holders before using their works for training, to state the purpose, scope, and duration of that use, and to pay fair compensation. Two neighbours, two philosophies — which is precisely why a rights position has to be read market by market rather than assumed from a single headline.

A second question runs the other way. If a machine generates a story or an image on its own, can anyone own it? Here the answer has been consistent: copyright requires a human author, and courts have repeatedly declined to recognise an AI system as one — a position the top of the US judiciary let stand in 2026. The most instructive case for our field involved a comic. When an artist sought to register a graphic novel whose images were produced with an AI image generator, the US Copyright Office split the work in two: the human-written text and the human selection and arrangement of the panels were protected, but the individual AI-generated images were not. Purely machine-made output, with no meaningful human hand, tends to fall outside copyright altogether. The human contribution has not become less important in the age of AI. If anything, it has become the thing the law is looking for.

None of this is abstract for the creators we represent, because in Asia the tension arrived first not in court but on the page — in front of readers. In May 2023, readers of the Naver Webtoon series The Knight King Returns with the Gods flooded it with one-star ratings after spotting that portions of the art had been refined with AI; the same series drew accusations that AI had been used to trace another studio's designs. The studio eventually pledged to stop, and re-drew episodes. The following month, webtoon artists mounted an online protest against a platform contract clause that required them to consent to their uploaded work being used as AI training data — an "AI-training right" quietly folded into the terms of publishing. This is exactly the kind of clause a rights agency exists to catch. It shows how a new use can hide inside old paperwork, and how covert use, once discovered, damages a title's standing with the very readers it depends on.

Regulation, meanwhile, has moved faster than the courts — and it has moved toward disclosure. Under the European Union's AI rules taking full effect in August 2026, providers of general-purpose AI models must publish summaries of the data they trained on and set out how they comply with copyright: the first binding attempt to give rights holders real visibility into what a model has read, reaching any catalogue whose rights touch the European market. Korea, for its part, has passed an AI framework law requiring AI-generated content to be labelled, with machine-readable watermarks permitted for visual formats such as webtoons, after a grace period. The commercial logic is already visible: with a large share of Korean webtoons' overseas revenue coming from English-language readers, studios have found that "100% hand-drawn" can become a selling point, and that hiding AI use costs more trust than disclosing it ever would.

While the courts sort out principle and regulators sort out disclosure, the market has been busy finding a price. The clear trend of the past two years is a shift from litigation toward licensing: rather than keep defending the claim that scraping was always fair, several AI companies have begun paying publishers for the right to use their text. The deals are uneven, still concentrated among large rights holders with well-cleared, long backlists, and by no means universal — much of the industry remains in court, and music-rights bodies such as Korea's KOMCA have stood up their own AI task forces. But a market for "training rights" is visibly forming, and that changes the calculation for everyone who owns a catalogue.

So where does a rights agency stand? Four principles hold regardless of how the cases finally land. First, sourcing matters: any right we grant should never become a route to launder unlicensed or pirated use. Second, a new use needs its own named permission — an AI-training or AI-adaptation right is a distinct slice of the bundle, not something that rides silently inside a translation licence or a platform's standard terms. Third, disclosure is now part of the deal: where a work was made with AI assistance, saying so is becoming both a legal duty and a matter of reader trust. Fourth, the author's moral rights — attribution and the integrity of the work — do not dissolve because a machine is involved. A well-drafted contract already anticipates the machine at the table.

The law will keep moving. Trials are on the calendar into next year, including what is reported to be the first jury trial in an AI-copyright case, and the picture will look different again by the time they conclude. What will not change is that someone has to read the rights carefully on the author's behalf, use by use, market by market. That is our work — machine or no machine.

Copyright & Technology offers general information about publishing rights and copyright and is not legal advice. The legal position described here is evolving and varies by jurisdiction. For decisions about a specific work or contract, consult qualified counsel in the relevant market.