Japan eyes fines for US AI labs that stonewall its safety orders

Japan's AI Promotion Act has a hole in it. The law, which took full effect in September 2025, lets the government investigate and issue guidance when AI misuse harms someone's rights — but it punishes nothing if a company simply refuses to cooperate. Officials are now considering adding penalties to close that gap, with a full reform blueprint promised before the end of 2026.

The push targets foreign developers hardest. A group inside the ruling Liberal Democratic Party, led by former digital minister Masaaki Taira, warned in a draft proposal that other governments can extract information from AI companies while Japan may not — pointing out the European Union already fines firms that refuse to hand it over. Japan's sharpest worry is advanced AI it neither builds nor controls, and it wants safety duties plus penalties covering models that go out of control or get used in cyberattacks.

Anthropic is already in the room. The Financial Services Agency reportedly assembled a 36-member group of government and industry figures in May 2026 to study cyber threats tied to Anthropic's Mythos model, with the Japanese branches of Anthropic and OpenAI taking part. Japan's updated AI Basic Plan, approved in July 2026, flagged high-performance models such as Claude Mythos as a concern — and the U.S. Commerce Department had limited access to Anthropic's Mythos 5 and Fable 5 earlier in 2026 before restoring it.

The review formally kicked off at an expert panel meeting on October 26, according to reports. Digital minister Toshiharu Furukawa told the panel Japan needs to work more closely with partner countries on AI safety, and a group of academics held its first meeting two days later. The panel hopes to settle on a direction by the end of the year, but some participants are urging caution because not every country is tightening AI rules.

Japan deliberately left penalties out of its first AI law, passed in May 2025, as part of a bid to become the “most AI-friendly country” and avoid scaring off cautious adopters. The law leaned on social pressure and voluntary compliance, with analysts assuming a company like OpenAI would follow local rules to keep market access. South Korea went the other way: its AI Basic Act took effect in January, allowing fines of up to 30 million won (about $21,000) for refusing inspections or ignoring corrective orders — a sum too small to move a large U.S. AI company.