Japan's JASRAC won't collect for fully AI-made songs, and wants the training exception in Article 30-4 changed

Japan's main music rights society has drawn a line under fully AI-made songs. According to NHK World, JASRAC's new AI policy says lyrics and music created by AI from a simple instruction show no human creative input. They are not copyright works, and their "authors" can't ask JASRAC to collect fees and royalties on them. For songs that are only partly AI-generated, JASRAC will manage the human-created part and nothing else.

The same policy pushes for a change in the law. innovaTopia reports that JASRAC launched a special page on 11 June 2026 asking for Article 30-4 of the Copyright Act to be revised, so that rights holders get a chance to choose whether their work is used to train AI.

What does the guideline change for songwriters?

Three things, based on NHK's report:

• Fully AI-generated songs are out. If a track came from a prompt and little else, JASRAC won't register or license it.
• Mixed songs are split. Where AI wrote some of a song and a person wrote the rest, JASRAC manages only the person's share.
• Disclosure is on the member. Creators registering a work must say which parts are AI-generated, and anyone who makes a false claim bears the legal responsibility.

None of this is JASRAC inventing new law. It follows the Japan Copyright Office's 2024 "General Understanding" on AI and copyright, which says material AI generates on its own, or from a simple prompt such as "Draw a cat", isn't a "creatively produced expression of thoughts or sentiments" and so isn't a work. Where a person uses AI as a tool, with creative intention and creative contribution, the result can be a work and the person is its author. The office says this is judged case by case. Detailed instructions that specify the expression count for more than long prompts that only describe an idea. Repeated attempts count only when the person is checking outputs and changing the instructions.

What is Article 30-4, and what does JASRAC want?

Article 30-4 is Japan's broad data-analysis exception. The Copyright Office's overview says it allows works to be used without permission, in principle, for purposes other than "enjoyment" of their thoughts or sentiments, and it names AI development and collecting training data as examples. Listening to music is the office's example of enjoyment. Training on it isn't.

The exception has limits. It doesn't apply when a purpose of enjoyment coexists, for example fine-tuning on one writer's songs in order to output their expression. It also doesn't apply when use would "unreasonably prejudice" the copyright holder. The office's main example of that is copying a database that's sold for data analysis, including dodging technical blocks such as robots.txt to get at it. What the overview doesn't describe is any general way for an individual songwriter to say no.

That is JASRAC's complaint. According to innovaTopia, its core ask is to secure rights holders an "opportunity to choose" whether their works become AI training material. JASRAC says it isn't seeking a blanket ban or opposing AI development, and it argues that Japan is the only G7 country with this kind of framework. The same report gives the results of a JASRAC survey of music creators, run January to March 2026: 54% opposed or leaned against their works being used for AI training, and 20% supported it or leaned towards supporting it. The report doesn't say how many people were surveyed.

How does this compare with the EU and the UK?

The EU has an opt-out. Article 4 of the 2019 Digital Single Market Directive creates a text and data mining exception, but it applies only if rights holders haven't expressly reserved their rights. For content published online, the reservation has to be in an appropriate form, such as machine-readable means.

The UK looked at copying that model and then backed away. In its Report on Copyright and Artificial Intelligence, published on 18 March 2026, the government said a broad exception with an opt-out "is no longer the government's preferred way forward". We covered that at the time. The same report proposes removing the UK's special protection for wholly computer-generated works, while keeping copyright for works made with AI assistance.

So on outputs, the UK is proposing to move towards where Japan already is: a person has to be the author. On inputs, Japan gives songwriters the least say of the three.

What we think

JASRAC is right on both counts, but the guideline and the lobbying need each other more than JASRAC's framing suggests.

Start with the money. For mixed songs, "manage only the human part" means human lyrics on AI-composed music earn only the lyric share, and an instrumental use of that track, such as background music in a shop, has no human part to pay for. That is our reading of the policy, not a JASRAC worked example, but it's the logical result. It is good for human writers as a group: a flood of prompt-made tracks can't register with JASRAC and dilute the pool that everyone is paid from. It is also a real disclosure burden on individual members. They now have to account for their own process, under legal liability, when the rules on how much human input is enough are still being worked out case by case.

Now the training side. Under the current law, a Japanese songwriter's catalogue can be used to build the model, and the model's simple-prompt output then competes with that songwriter while paying nobody. CISAC's 2024 study put 24% of music creators' revenues at risk by 2028, citing both unpaid training and "the substitution effect" of AI output. Denying copyright to AI output deals with half of that. Article 30-4 leaves the other half open.

We'd push JASRAC further than an opt-out. An opt-out on its own puts the work on each creator. The EU version needs a machine-readable reservation, and most songwriters don't run websites that crawlers read. A collecting society is already the body that holds the mandate and the repertoire data. The useful change to Article 30-4 is one that lets JASRAC reserve rights on its members' behalf and license training collectively, so the "choice" comes with a price attached rather than just a refusal. The one place Japanese law already lets a block count is for databases sold for analysis. That suits labels and data vendors, not the lyricist with one song in the register.

The survey is also a warning to JASRAC. A majority object, but one in five creators it surveyed support or lean towards training use. Any collective scheme needs a real per-member choice in both directions, or it repeats the problem it's fixing.

Sources

• JASRAC's AI policy, as reported by NHK World, June 2026
• innovaTopia on JASRAC's special page and Article 30-4 request, 12 June 2026
• "General Understanding on AI and Copyright in Japan": Overview, Japan Copyright Office, May 2024
• Report on Copyright and Artificial Intelligence, UK government, 18 March 2026
• Directive (EU) 2019/790, Article 4, legislation.gov.uk
• Global economic study shows human creators' future at risk from generative AI, CISAC, December 2024