On 11 November 2026, Kyodo News reported that a closed-door Japanese government working group on AI legal systems has confirmed it is considering new safety duties and penalties for firms behind high-performance AI. Under the approach described, the state could issue binding orders to a developer that refuses to consult after a serious problem, with penalties for violating an order. Nothing has been enacted, and the scope of covered firms is undecided.
What was reported on 11 November 2026
According to Kyodo’s report as syndicated on OANDA Lab (published 11 November 2026), the working group discussed requiring AI-related businesses to ensure safety and creating new penalties to deal with risks from high-performance AI. The stated aim is to give the government a countermeasure against bad-faith businesses that will not cooperate when serious threats arise, such as misuse for cyberattacks or runaway AI behaviour.
The mechanism described has three steps: a problem occurs; the business declines to consult; the government then issues a binding order as needed, and a violation of that order would carry a penalty. Kyodo reports that penalties would be limited to incidents involving high-performance AI. The government plans to present an overall picture of its system review, including the content of proposed law amendments, by the end of the year.
How this differs from the 2025 AI Promotion Act
Japan’s AI Promotion Act, passed by the Diet in May 2025, is a promotion-focused law and does not set monetary penalties, as law-firm summaries such as White and Case’s analysis describe. Kyodo’s framing is that the existing framework has no penalty provisions, so there is no compulsory step if a business ignores a government request. For a fuller account of the current law, see our guide to the Japan AI Promotion Act and its lack of penalties, and for regional context the comparison of the Korean, Japanese and Taiwanese AI basic acts.
What is still undecided
The reports leave several points open, and the page should not be read as describing a settled rule:
- Who is covered. Kyodo notes that AI businesses range from research and foundation-model development to user-facing services, and that the boundaries will be discussed further.
- Foreign firms. Kyodo raises the question of how effectively Japanese users could be protected if large US technology companies are included.
- Legal form. The final shape depends on the amendment content the government says it will set out by year end; no text has been published in the reports above.
What it could mean for AI research in Japan (editorial interpretation)
The sources do not say how academic or public-sector AI research would be treated, so the following is interpretation, not reporting. Kyodo’s description of the scope question lists research alongside model development and services, which means university and institute labs working on frontier-scale models cannot assume they would be outside any definition until it is drafted. Research offices at Japanese institutions funded through bodies such as JST or JSPS may want to track the consultation text for definitions of high-performance AI, incident-reporting duties and any exemption for non-commercial research.
What it could mean for international collaborators (editorial interpretation)
Because the reports say the question of including large overseas firms is unresolved, collaborators outside Japan should treat any obligations that attach to joint projects, shared models or hosted compute as unknown. A practical step is to watch for the year-end overview and then check collaboration agreements for clauses on safety incidents, cooperation with regulators and model access. For how frontier-safety governance is developing elsewhere, see our frontier AI safety timeline.
Sources
- Kyodo News, via OANDA Lab, report on the government considering penalties for runaway high-performance AI, published 2026-11-11: https://www.oanda.jp/lab-education/?p=133389 (checked 2026-11-14).
- White and Case, “Japan’s first AI legislation becomes law”: whitecase.com (checked 2026-11-14).
