Japan made it legal to train AI on your work — without asking you

Slava Spitsyn · June 16, 2026 · 3 min read

In 2018, Japan rewrote its copyright law to make one thing unambiguous: training an AI model on someone's work does not require their permission. The country with the strongest creative exports on earth — anime, manga, games — handed its own creators one of the weakest legal shields against AI.

What's new

• Article 30-4 was added to Japan's Copyright Act in 2018 and took effect in January 2019. It permits the use of copyrighted works for purposes that are not about enjoying the expression of the work itself — explicitly including data analysis and machine learning.
• In plain terms: text, images, manga, and music can be ingested to train a model without asking the rightsholder and without paying them.
• The exception is broad. It applies regardless of the kind of work, and it does not carve out commercial model training.

How it works

Japanese copyright law draws a line between two kinds of use:

• "Enjoyment" use — reading the article, watching the film, listening to the track. This still needs permission.
• "Non-enjoyment" use — processing the work as data, where a human is never meant to perceive the expression. Article 30-4 places machine-learning training on this side of the line.

The reasoning is that a model "learning" statistical patterns from a work isn't consuming its creative expression the way a reader or viewer does. That framing is what opened the door — and it is far more permissive than most other jurisdictions.

Behind the news

The policy didn't go unchallenged. As generative models started producing work in the style of named artists, Japanese creators pushed back hard.

• In 2024, Japan's Agency for Cultural Affairs published guidance clarifying that Article 30-4 does not apply where the use would "unreasonably prejudice the interests of the copyright owner," and that infringement can still occur at the output stage if a generated work copies a protected one.
• But the default at the training stage stayed permissive. The clarification narrowed the edges; it did not close the door.

So the practical reality for a creator in Japan is this: by the time you could argue your interests were "unreasonably prejudiced," your work has already been ingested — quietly, at scale, with no notice.

Why it matters

This is the part that travels beyond Japan. Where legal remedies are uncertain or impractical, one remaining lever is technical control at the origin.

Recording the requester's confidence level, the active policy, and what the origin decided to serve. That record still does not reveal later training or use.

Origin controls can complement legal and contractual rights; they do not replace them.

How I see it

This is the shift I keep coming back to while building WARD. The instinct is to ask the law to protect the work. But the law is slow, fragmented, and in some places pointed the other way entirely.

The lever that stays in your hands is a declaration plus an origin decision: state, in machine-readable form, how AI may use your content, decide what to serve, and retain a minimal receipt of that decision. Spoofed or unidentified requesters remain an explicit uncertainty.

Japan is the clearest example of the gap, not the only one. Where the law goes quiet, the protocol has to speak.

→ getward.org