When a Japanese company uses or develops generative AI, copyright questions are decided first under the Japanese Copyright Act, not by the outcome of lawsuits abroad. Using works to train AI is, as a rule, permitted without the copyright holder's consent (Copyright Act, Art. 30-4), but there are exceptions where the purpose includes reproducing the training data's expression, or where the use would unreasonably prejudice the copyright holder.
When AI output is used outside the company, infringement is judged by the ordinary test of similarity and reliance, and an employee can infringe even without knowing the original work. Courts in the US and UK issued several decisions in 2025, but each rests on its own country's law and its own record, and none of them carries over to Japan as it stands.
Situations where this comes up
- Your marketing team creates an advertising visual with an image generator, and before launch someone points out that it looks like a well-known illustration.
- You want to add retrieval-augmented generation (RAG: a system that searches external data and builds answers from it) to your service, and are unsure whether you may ingest third-party articles.
- You learn that a business partner is using a model fine-tuned only on one illustrator's works.
- After reading about US lawsuits, your board asks whether the company's use of generative AI is safe.
Training: Article 30-4 is the starting point
The rule
A work may be used, by any method and to the extent necessary, where the purpose is not to enjoy, or have others enjoy, the thoughts or sentiments expressed in it (Copyright Act, Art. 30-4). "Enjoyment" here means experiencing the work as a reader or viewer. Information analysis (extracting elements from large amounts of information and comparing, classifying or otherwise analysing them) is listed in the statute as a typical example (Art. 30-4(ii)).
The Legal System Subcommittee of the Copyright Subdivision of the Council for Cultural Affairs published "General Understanding on AI and Copyright in Japan" on 15 March 2024 (the "Understanding"). It treats use for AI training as information analysis covered by Article 30-4. The statute draws no line between commercial and non-commercial purposes.
The Understanding itself is not legally binding; the courts have the final word.
Exception 1: an enjoyment purpose is mixed in
If a single act has several purposes and any one of them is enjoyment, Article 30-4 does not apply. The Understanding gives these examples:
- Deliberately overfitting a model so that it reproduces works in its training data.
- Fine-tuning on a small set of one creator's works with the aim of outputting the creative expression in those works.
- Building a RAG database with the aim of outputting the creative expression of the articles in it.
By contrast, a "style" as such is an idea, and similarity of style alone is not infringement. With a small body of works, however, the shared features may go beyond style into expression, so the purpose of fine-tuning can bring it within the exception.
Exception 2: unreasonable prejudice to the copyright holder
Article 30-4 contains a proviso: it does not apply where, in light of the nature and purpose of the work and the manner of use, the use would unreasonably prejudice the copyright holder's interests. The Understanding says this turns on whether the use competes with the market for the work or impedes its potential future markets.
The standard example is copying, without permission, a database that is compiled and sold for information analysis. The Understanding adds that where a site restricts AI-training crawlers through robots.txt (a file that tells crawlers which pages they may visit) or a login, and its record of building databases or licensing them shows that a training dataset is likely to be offered, circumventing that restriction to collect data may fall under the proviso. On the other hand, a mere "no AI training" notice from the rights holder does not by itself trigger the proviso.
In practice
Caution
Free Tool Related to This Article
Contract Risk Checker
Try our free simulator related to this topic.
Try for free →RAG and Article 47-5
RAG involves copying separate from model training: articles and other material are copied into a search database. According to the Understanding, if building that database is not aimed at outputting the creative expression of the source works, Article 30-4 may apply.
When an answer displays parts of existing works, Article 47-5(1) may apply as a "minor use" incidental to providing the results of computer information processing.
Caution
Output and use: similarity and reliance
The rule
AI output is treated like human work: it infringes if it is both similar to an existing work and created in reliance on it. Similarity means that the essential features of the existing work's expression can be directly perceived (Supreme Court, First Petty Bench, judgment of 28 June 2001, Case No. 1999 (Ju) 922). Reliance means that the new work was based on the existing one.
Three scenarios for reliance
The Understanding sorts reliance in generative AI into three situations:
- If the user knew the existing work and had the AI generate something similar, reliance is found. Typical cases are inputting someone else's image to generate an image ("image to image") and prompting with a work's title or a character's name.
- If the user did not know the work but it was in the AI's training data, reliance is normally presumed. It may be rebutted if it can be shown that the system is technically designed so that the expression of training works is not output.
- If the user did not know the work and it was not in the training data, any resemblance is coincidence and there is no reliance.
In practice
Important
Liability if infringement is found
| Item | Details |
|---|---|
| Injunction | An injunction to stop or prevent infringement may be granted without intent or negligence (Copyright Act, Art. 112) |
| Damages | Damages require intent or negligence (Civil Code, Art. 709), and the Copyright Act provides presumptions for calculating damages (Art. 114) |
| Criminal penalty | Intentional infringement is punishable by imprisonment of up to 10 years, a fine of up to 10 million yen, or both (Art. 119(1)) |
| Corporate fine | Where an employee acts in the course of business, the company may be fined up to 300 million yen (Art. 124(1)(i)) |
The Understanding notes that even a user with no intent or negligence may have to return the unjust enrichment, measured as a reasonable licence fee. It also says that developers and service providers are more likely to be treated as the infringing party where the system frequently produces infringing output, or where they know similar output is highly likely yet take no measures to prevent it. A court will not normally order destruction of a trained model, but it may do so where the model is in a state that produces output similar to its training works with high probability.
In litigation, the court may order a party to produce documents needed to prove infringement (Art. 114-3). Disclosure of training data may well become an issue.
Is there copyright in what AI produces?
AI has no legal personality, so it cannot be an author. Whether AI-assisted material is a copyright work depends on whether a human made a creative contribution, assessed case by case.
The Understanding says that detailed instructions specifying expression, and repeated attempts in which the user reviews the output and revises the instructions, make protection more likely. The mere number of attempts, or long instructions that only convey ideas, do not count. For commercial use of AI output, see "Generative AI and copyright: who owns AI output and when it infringes".
In patents, the Intellectual Property High Court upheld the dismissal of a national-phase application that named an AI as the inventor (judgment of 30 January 2025, Case No. 2024 (Gyo-Ko) 10006; a petition for acceptance of final appeal was not accepted on 4 March 2026).
Litigation abroad: what the decisions actually say
Foreign cases are decided under their own countries' laws. Below are only those whose content we could confirm from decisions and orders published by the courts.
- Authors v. Anthropic (US District Court, Northern District of California, No. 24-cv-05417): An order of 23 June 2025 held that training large language models on books was fair use (the US copyright exception). It held, however, that building a central library from books taken from pirate sites could not be treated the same as training copies, and sent that issue to trial.
- Authors v. Meta (same court, No. 23-cv-03417): An order of 25 June 2025 granted Meta summary judgment on fair use as to the books of the 13 plaintiff authors, on the record before the court. The court stated expressly that this does not mean such training is lawful; it reflected the plaintiffs' failure to present evidence that AI output would harm the market for their works.
- Thomson Reuters v. ROSS Intelligence (US District Court, District of Delaware, No. 1:20-cv-613): A memorandum opinion of 11 February 2025 rejected fair use for using legal headnotes to build an AI search tool. The court noted that the AI at issue was not generative AI.
- Getty Images v. Stability AI (High Court of England and Wales, [2025] EWHC 2863 (Ch)): Before the judgment of 4 November 2025, the training claim had been dropped because there was no evidence that training took place in the UK. The court dismissed the secondary infringement claim, holding that an AI model that does not store or reproduce the works is not an "infringing copy", and upheld the trade mark claim only for certain watermark examples. On 16 December 2025, permission to appeal was granted on the secondary infringement point ([2025] EWHC 3343 (Ch)).
- AI as author (US Court of Appeals for the D.C. Circuit, No. 23-5233): On 18 March 2025, the court upheld the Copyright Office's refusal to register a work attributed to an AI, holding that the Copyright Act requires a human author. The Supreme Court denied certiorari on 2 March 2026 (No. 25-449).
Many other cases, such as The New York Times Company's suit against Microsoft and OpenAI (US District Court, Southern District of New York, No. 1:23-cv-11195, filed 27 December 2023), appear in court records, but we have not confirmed any final outcome. This article does not forecast results.
What this means for Japanese companies
As the UK case shows, where training or deployment takes place affects which law applies. A model trained in Japan may still be challenged under US, UK or EU law if offered to users there. The EU AI Act requires providers of general-purpose AI models to put in place a policy to comply with EU copyright law and to publish a sufficiently detailed summary of the content used for training (Regulation (EU) 2024/1689, Art. 53(1)(c) and (d), applicable from 2 August 2025 under Art. 113(b)).
What you can do now
- 1
List the services
List the generative AI services used in the company and check each one's terms: prohibited inputs, ownership of output, and whether any indemnity for infringement is offered.
- 2
Set prompt rules
Set rules, either a ban or prior approval, for inputting others' works, prompting with titles or character names, and naming a particular creator's style.
- 3
Check before publishing
Before any output leaves the company, check it against existing works using image and text search.
- 4
Keep generation records
Keep records of prompts, the service and model used, the date of generation, and later edits.
- 5
Plan the complaint response
Decide in advance how to respond to a complaint from a rights holder (taking content down, checking the facts, replying, preventing recurrence) and who is responsible.
- 6
Document training data
If you develop or fine-tune models, document where training data comes from, the collection policy, how robots.txt is handled, and the measures that prevent similar output.
- 7
Check foreign rules
If you offer AI services to users abroad, check the copyright and AI rules of each target country.
When to talk to a lawyer
- You receive a warning letter, a takedown request or a demand for licence fees from a rights holder.
- Someone inside or outside the company says that planned or published output resembles an existing work.
- You plan to fine-tune a model or build a RAG database using a particular author's or publisher's material.
- You need to negotiate warranty or indemnity clauses on infringement with an AI vendor.
- You are launching a generative AI service for users in the US, UK or EU.
Our firm, led by attorney Yusuke Abe (admitted in Japan and New York), advises on copyright and contracts for adopting and developing generative AI, in Japanese and English. Please contact us through our contact page.
FREE CONSULTATION
Corporate Law: your first 30 minutes are free
- ・Handled by Yusuke Abe, admitted in Japan and New York
- ・Online (Zoom etc.) or at our office in Otemachi, Tokyo
- ・In English or Japanese
Frequently asked questions
Is permission from copyright holders not needed for AI training in Japan?
As a rule, no. Information analysis that does not aim at enjoyment may be done without permission (Copyright Act, Art. 30-4). Exceptions may apply, however, such as fine-tuning aimed at outputting the training data's expression, or copying without permission a database sold for training.
An employee unknowingly generated and published an image similar to an existing work. What now?
If the work was in the AI's training data, reliance may be presumed even though the employee did not know it, and the image may infringe. Without intent or negligence there should be no damages or criminal liability, but an injunction and return of a reasonable licence fee may still be sought. Take the image down first and preserve the generation records.
Is it illegal to generate images in a particular artist's style?
A style as such is an idea, and similarity of style alone is not infringement. Output in which the concrete expression of a work can be perceived may infringe, however. Prompting with an artist's name, or fine-tuning only on that artist's works, is a high-risk practice.
Has any Japanese court ruled on copyright infringement by generative AI?
In a search of the court website's case database on 26 October 2026, we found no judgment that squarely decided copyright infringement by AI training or AI output. The database contains only some of the judgments. For now, the main guides are existing case law and the Agency for Cultural Affairs' Understanding.
Does our company own copyright in images or text made with AI?
AI cannot be an author. Where a person contributed creatively, for example by specifying expression in detail or revising instructions while reviewing the output, the result may be protected. If you license such material to others, be ready to explain that AI was used and why the material qualifies as a work.
Sources: Copyright Act and Civil Code (e-Gov Laws, text as of 26 October 2026); Legal System Subcommittee, "General Understanding on AI and Copyright in Japan" (15 March 2024); Copyright Division, Agency for Cultural Affairs, "Checklist and Guidance on AI and Copyright" (31 July 2024); court website case database; decisions and orders of the foreign courts cited; EUR-Lex (Regulation (EU) 2024/1689).