AI Regulation Tracker  /  Copyright guidance

Japan's "free to train" copyright rule has limits: what Article 30-4 and the Bunkacho guidance actually allow

Japan's Copyright Act Article 30-4 lets AI developers copy protected works to train models, but the Agency for Cultural Affairs guidance carves out uses that unreasonably prejudice rights-holders and treats generation as a separate infringement question. IP lawyers advising a "train in Japan" strategy need the exclusions on the record.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

Japan is often described as the one major market where you can train an AI model on copyrighted works without a license. That reputation rests on Article 30-4 of the Copyright Act, and it is not wrong. But it is incomplete. The Agency for Cultural Affairs, known as Bunkacho, has set out how it reads the provision in a document titled "Approaches to AI and Copyright" (考え方), whose core text is dated March 15, 2024. For any lawyer building a cross-border training strategy around Japan, the exclusions in that guidance matter as much as the permission.

What Article 30-4 permits

Article 30-4 allows the use of a copyrighted work, including copying it, where the purpose is not to enjoy or cause others to enjoy the expression of the work. Machine learning generally fits: a model ingests text or images to learn statistical patterns, not to present the expression to a human audience. On that reading, the training-stage copies that AI development requires are, as a baseline, permitted without the rights-holder's consent. This is the legal foundation of the "train in Japan" narrative.

The limits the guidance spells out

The guidance does not stop at the permission. It reads Article 30-4 together with its own proviso, which withdraws the exception where the use would unreasonably prejudice the interests of the copyright holder. Bunkacho describes situations that can fall outside the safe zone. Deliberate overfitting, where a model is trained so that it reproduces specific protected works, points away from a non-enjoyment purpose. So does copying that circumvents an existing market for licensed training data, because that harms the rights-holder's ability to exploit the work. The point is that purpose and market effect are live questions, not settled by the fact that the copying happened at the training stage.

Two stages, two tests

The most important structural feature for practitioners is that Bunkacho analyzes AI copyright in two stages: development and training on one side, generation and use on the other. Article 30-4 speaks to the first. It does not immunize the second. If a model produces an output that is similar to and derived from a specific copyrighted work, that output can infringe under ordinary copyright principles, regardless of how the training was handled. A clean training process does not launder an infringing generation, and a lawful-looking output does not cure a training use that was aimed at reproduction. Each stage carries its own analysis.

The paid-dataset track

Alongside the exception, a licensed route is taking shape. A May 2025 Stakeholder Network summary, produced with rights-holder participation, and an emerging government-supported model point toward rights-holders preparing and selling training datasets under contract. That matters commercially. If a licensed-data market exists and is functioning, copying that bypasses it is exactly the kind of use the guidance flags as potentially unreasonable prejudice. The growth of paid datasets can therefore narrow the practical reach of the free-training exception over time.

The cross-border angle

For a US reader, Japan sits between the two poles of the current debate. It is not the United States, where the training question is being fought as fair use in litigation such as The New York Times v. OpenAI, and it is not the European Union, where the text-and-data-mining rules let rights-holders opt out of machine learning. Japan grants a statutory exception up front but conditions it on purpose and market effect. The guidance does not bind US firms directly, and it is not a court ruling. But it is the reference point Japanese counsel and counterparties will use, and any advice that sells Japan as a no-strings training haven is missing the exclusions that Bunkacho has now put in writing.

What it does not do

The guidance does not create new penalties, and it does not rewrite the statute. It is an interpretation, not binding law, and disputes will ultimately be decided by courts applying Article 30-4 and the infringement rules to specific facts. It also does not resolve every hard case; the line between permitted learning and unreasonable prejudice is drawn example by example. What it does is remove the excuse for treating Article 30-4 as a blanket license.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.