Japan AI Penalties: Binding Orders for Advanced AI Firms

Japan has had an AI law for about eighteen months, and in that time the government has had no way to punish anyone under it. That may change. According to Kyodo News, officials are now considering safety duties for AI businesses, with penalties for breaking them.

The idea came up at a closed-door working group on AI legislation on Wednesday, November 11. Kyodo’s report cites a source familiar with the matter, so nobody has put a document on the table for the public to read yet.

What would actually change

Here is the mechanism as Kyodo describes it. If an AI business doesn’t cooperate with the government after an incident, officials could issue a legally binding order. Ignoring that order would carry a penalty, and Kyodo says those penalties are expected to cover only incidents involving advanced AI.

My read: that’s a narrow trigger. Japan’s AI law, passed in May 2025, already lets the government investigate businesses and give guidance when someone behaves maliciously, per Kyodo. It has no penalty provisions, so the measures are nonbinding. My read: in practice that leaves the government able to ask, and the company able to say no.

A Japanese-language piece in Saga Shimbun, carrying Kyodo’s reporting, adds some colour. The target is the business that refuses to talk after a serious problem such as AI being misused in a cyberattack or getting out of control. Ordinary AI would stay under existing laws and guidelines.

The same report says Japanese units of Anthropic and OpenAI, plus NTT, told the meeting what they’re doing on safety and what they want from the rules. The legal database CODE, in the free lead of a paid article, says six companies at home and abroad were heard, including Google, and that the Cabinet Office secretariat presented a first draft. Per that lead, the draft would put the burden on companies that provide AI rather than people who use it.

What nobody has decided

Quite a lot, going by the sources.

Question	What the reporting says
Who is covered?	Kyodo says the government will examine which types of firms fall in: researchers, model developers or end-user services.
What counts as advanced AI?	No definition in Kyodo or Saga Shimbun. CODE’s lead says “a certain level” of performance.
How big are the penalties?	None of the three sources gives an amount.
Do US companies count?	Saga Shimbun says including big American tech firms would raise the question of whether enforcement could work.
When?	Kyodo says the working group aims to present an outline of the review, possible law changes included, by year-end.

An outline by December isn’t a bill. Nothing here has passed, and nothing takes effect.

Why the timing isn’t random

Kyodo ties the proposal to concern about advanced models behaving unpredictably and to a run of data leaks and unauthorized access cases suspected of involving AI misuse. We covered Japan’s earlier cyber emergency declaration, where officials answered with a rule about bank ID photos. My read: a penalty power is a different kind of response.

Saga Shimbun also points to the EU, which it says requires some AI providers to report serious incidents and can impose large fines. Australia is moving the same way: its government said on November 10 it would create a regulator and six duties for frontier AI firms, and that one also has no fine amounts yet.

My read

My read: this is a sign of how the soft-law approach is aging, more than a crackdown. A law that can only ask nicely works while companies want to cooperate. The first time one doesn’t, the government finds out it has no lever, and that appears to be the gap Tokyo is trying to close.

I’d also be careful about the headline. A source-based report on a closed meeting, a narrow scope and a year-end outline are all early signals. Plenty of proposals like this get softened or dropped before a bill exists. What would change my mind is a published draft with a definition of advanced AI and a fine amount in it.

The thing to watch is the working group’s outline, due by the end of the year according to Kyodo.

Sources

Note on sources: the main report is Kyodo’s, which cites an unnamed source. The Cabinet Office meeting materials were not available to me, and CODE’s full article is paywalled, so only its public summary is used. Check the Cabinet Office site for the official documents when they appear.