Regulation AI Minute Newsroom 2026-08-18

Japan will ask AI companies to publish what they trained on. There is no penalty for saying no.

A Japanese government expert panel on intellectual property in the age of AI broadly approved a draft principle code for generative AI companies on 18 August. It asks firms to disclose publicly what data they trained their models on and how they collected it, to publish information about the models and the training process, and to take measures protecting intellectual property rights. The code is not binding. The government intends to run it on a comply-or-explain basis: a company may decline to follow it, provided it says publicly why it will not. The draft rests on Japan's AI technology law enacted in May 2025 and was written amid growing concern that text and images are being used for training without the consent of rights holders, and that the results may infringe copyright.

Why it mattersJapan has one of the most permissive copyright regimes in the world for AI training, and this is its government trying to buy legitimacy for that position without closing the door — transparency offered in place of permission. Comply-or-explain only works when explaining is embarrassing, which depends on whether anyone reads the disclosures and whether the largest labs, most of them foreign, sign up at all. But notice what is being asked for: a public account of where the training data came from. That is the exact fact that nearly every AI copyright lawsuit on earth is currently fighting to obtain through discovery. If companies begin publishing it voluntarily in Japan, plaintiffs in other countries will read those documents very carefully.

✓ Verified · 2 sources

Read in the app — free, in 9 languages

Japan will ask AI companies to publish what they trained on. There is no penalty for saying no.

Related stories