The Japanese government is considering imposing safety obligations on AI-related businesses and creating new penalties. Kyodo News reported the direction on November 11, saying it was confirmed in a closed working group that is discussing AI legislation. The government plans to publish the overall picture of its institutional changes, including legal amendments, by the end of the year. Japan's earlier AI legislation carried no penalties, so this marks a change of course.
What the proposal contains
According to Kyodo, the framework under consideration has three layers:
- AI-related businesses would be required to take on an obligation to ensure safety;
- if an incident occurs and a company refuses to consult with the government, the state could issue a legally binding order when necessary;
- violations of such an order would be subject to newly created penalties, narrowed to incidents caused by high-performance AI.
The government's stated rationale is to respond to serious threats such as the misuse of AI in cyberattacks and AI going out of control, and to make sure it has options when dealing with malicious operators who refuse to cooperate. The report cites two pieces of background: a U.S. report said advanced AI had experienced unplanned, unauthorized access during development, and a number of recent cyberattacks are suspected of involving the misuse of AI.
Who the rules cover has not been decided. The AI business chain is finely divided, running from research and the development of foundation models to services offered to end users, and the government says it needs to work out further how far down that chain the rules should reach. The report also notes that if large U.S. IT companies were brought in, questions would be raised about whether Japanese users could actually be protected. The report does not describe the specific form or amount of the penalties.
From "name and shame" to penalties
Japan's AI legislation over the past two years can be traced through several milestones.
In early 2025, Jiji Press reported on the draft AI bill of the time: no penalties and no improvement orders, with action against malicious operators going no further than publishing their names.
The AI Promotion Act passed in 2025 kept that approach. The law focuses on promoting research, development and adoption. The government can investigate, issue guidance and, where necessary, disclose the names of operators, but there are no penalties.
The latest change adds an enforceable component within the framework of the same law, but only for high-performance AI. The government has no plan to apply it across the board to every AI service. It has not formally explained why the scope is so narrow. From the background listed in the reports, the scenarios singled out are high-consequence ones such as cyberattacks and loss of control, and ordinary problems with generated content are not at the core of this discussion.
Set against its neighbors, the EU AI Act allows fines of up to 7% of global annual turnover for violating companies, while South Korea's AI Basic Act, in force since January this year, caps fines at 30 million won. If Japan follows the current direction, it would land between the two: far narrower than the EU, focused only on high-performance models, but with an added layer of orders and penalties beyond the earlier approach of simply publishing names.
What to watch at year-end
When the overall framework is published at the end of the year, several questions should get answers: how high-performance AI will be defined, whether by compute, parameter count or capability test results; whether the obligations fall on model developers, on service providers, or on both; how foreign companies will be brought in, and whether penalties can be enforced against firms with no physical presence in Japan; and when a bill to amend the law will be submitted to the Diet.
The government has not publicly addressed any of these questions yet, and the Kyodo report does not say when a bill might be submitted. For now, Japanese AI companies and overseas model providers that serve users in Japan can only wait for the year-end plan and see which tier they fall into.
Sources: Kyodo News reporting on the AI legislation working group, Jiji Press, CocoLoop, Mainichi Shimbun. The framework of duties and penalties follows Kyodo's account; the EU and South Korean fine ceilings follow the legal texts of those jurisdictions.