The story circulating about Tokyo publishers banning automated tools inside manga studios doesn’t hold up: no such contract text exists in public sources. What exists is an interpretive Japanese government paper from March 2024, an operational checklist from July, a joint statement by 17 publishers in October 2025, and a publishers’ working group opened in March 2026 aiming at clauses in author contracts. Meanwhile studios are short of human assistants, not overstaffed.
A manga studio looks nothing like the myth of the solitary artist. There is the mangaka, who draws the characters and decides how each page is framed, and there are the assistants: one does backgrounds, one fills the blacks and the screentones, one cleans up linework. It is a hand-built assembly line running against a weekly or monthly deadline. And it is precisely the kind of work an automatic image generator looks designed to hollow out: backgrounds repeat, textures are procedural, colouring is arguably just a function of the greyscale image underneath.
The artificial intelligence at stake here is of two different kinds, and keeping them apart is the whole point. First, generative image models — neural networks trained on vast collections of illustrations that turn a text prompt or a rough sketch into a finished new drawing — and their sharper edge for a working artist: lightweight fine-tuning (LoRA) done on a few dozen works by one specific author in order to reproduce that author’s style. Second, far less spectacular tools already in daily use: machine learning models that colourise a black-and-white drawing, or assist the writing stage. Japan’s current public argument is almost entirely about the first kind. An assistant’s job is mostly the second.
The contract everyone cites and nobody has read
Better to say it up front, because it is the premise that circulates most and stands up least: there is no publicly available contract from a Tokyo publisher governing the use of automated tools inside an author’s studio. It isn’t secret — the work has only just begun. On 30 March 2026 the Japan Book Publishers Association, the trade body for Japanese book publishers with 381 member companies, set up a working group on generative AI, joined by roughly forty of them. The stated aim is guidelines for publishers by autumn, with a view to later inserting clauses into contracts with authors.
Two things in that sentence matter more than they seem. First, the subject is the use of AI by the author — the person who signs. Not the internal organisation of the studio, not the collaborators. Second, we are at the stage where guidelines get drafted, not applied. Nobody working in a studio today faces a prohibition. They face a vacuum, which is psychologically different and in some ways more uncomfortable.
What Japan’s culture agency actually said
There is an institutional reference point, but it is more modest than the retelling. On 15 March 2024 the legal subcommittee of Japan’s Council for Cultural Affairs approved a document titled AIと著作権に関する考え方について — roughly, “On the approach to AI and copyright” — following a public consultation between 23 January and 12 February and technical approval on 29 February. It is an interpretive document: it explains how existing law should be read, in particular Article 30-4 of the Japanese Copyright Act, which permits uses of protected works where the purpose is not “enjoyment” of the work itself. That article is the legal ground on which model training rests in Japan.
It is not new legislation, it is not sector-specific, and it does not deal with manga as such. On 31 July 2024 the agency followed up with an operational checklist, which flags two concrete technical situations as potentially problematic: RAG systems drawing on databases of protected works, and additional training with LoRA on a small number of works by a single author. That is where the legal argument touches the craft. A recognisable style, built over a decade of serialisation, becomes reproducible from a dataset of a few dozen pages.
31 October 2025
The industry’s answer came from exactly there. On 31 October 2025 seventeen publishers — Kodansha and KADOKAWA among them — together with the Association of Japanese Animations and the Japan Cartoonists Association signed a joint statement on creation and rights in the age of generative AI. The trigger was the flood of videos closely resembling well-known works produced with OpenAI’s Sora 2. The statement objects to the opt-out mechanism adopted for that service and argues that a rightsholder’s consent is required at the training stage too. The same day Shueisha issued its own, sharper statement, calling for measures beyond opt-out and for legislative action by the state.
A joint statement is not a law and binds nobody. But it marks a gap: the reading of Article 30-4 offered by the culture agency and the position taken by publishers and authors do not line up. With no settled case law — 2024 sources note its absence — that gap stays open, and it lands on whoever signs the contracts.
The labour market points the other way
Here is the part that dismantles the popular version. According to people inside the industry, the problem in Japanese studios is not that assistants are being replaced by a model: it is that assistants cannot be found. More authors are looking for assistants than there are assistants available. Takahiro Ozawa, half of the manga duo Ume, has argued publicly that AI will not take work away from assistants or editors precisely because of that labour shortage.
The history of the tools confirms how ambivalent the sector is. On 29 November 2022 Celsys announced a “generative AI palette” built on Stable Diffusion inside Clip Studio Paint — the drawing software used by comics professionals worldwide. It withdrew the feature on 2 December, three days later, after user protests, and reaffirmed that position in February 2024: the program does not ship an image generator today. Yet the same company has long used non-generative AI — an automatic colourisation preview available since 2018, built on a proprietary dataset gathered through a 2016 campaign — and in September 2023 it invested in the AI developer ax Inc.
That isn’t hypocrisy. It is an operational distinction the public debate struggles to make. A model that colours a drawing you made yourself, trained on material contributed with consent, does not raise the same problem as a model that outputs a page in your own hand from a prompt. And accepted tools already exist in the pipeline: Comic-Copilot, the “AI editor” run by the Shonen Jump+ editorial team at Shueisha with Kensuke Furukawa’s company Alu, or automatic colourisation on pixiv Sketch.
What is actually at stake
Put the pieces together and the stake is not the hand that inks the backgrounds. It is the chain of accountability. A publisher releasing a series needs to be able to say, when challenged, who produced each image and with which tools — because the exposure is not aesthetic, it is infringement. That is why the clauses under discussion aim at the author: he or she is the single point of attribution. And if a contract one day requires a declaration about the use of generative tools, that declaration will cover work done by assistants too — not because anything is forbidden to them, but because the mangaka signs for everything that leaves the studio.
One detail is worth more than most of the statements. In the July 2024 checklist, the critical case is not the giant model trained on billions of images. It is the LoRA fine-tuned on a handful of works by one specific author — the cheapest, fastest, easiest tool of the lot, the one anyone, inside a studio or outside it, can train in an afternoon on a colleague’s back catalogue.