In November 2026, the Japanese government confirmed a major policy shift at a closed-door working-level task force meeting: it is studying amendments to the current AI Promotion Act (Act No. 53 of 2025), planning to introduce mandatory safety obligations for operators of “high-performance AI with significant risks” and to impose fines on operators that refuse to comply with rectification orders. The government plans to complete the overall institutional restructuring plan for the legal amendment within the year, including a framework document setting out the specific revisions.
To understand the weight of this shift, one must first see the actual state of the current law. The AI Promotion Act fully took effect on September 1, 2025, and its core design philosophy is “innovation first”: the government’s enforcement toolbox contains only administrative guidance (gyōsei shidō)—a written recommendation with no legal force, which companies are entirely free to ignore without any legal consequences. The only “punitive” mechanism is public naming and shaming, in which regulators make the names of non-compliant companies public and rely on public pressure. The law “sets principles and lets the government investigate and respond with guidance and advice when AI infringes on citizens’ rights, without attaching any penalties.”
The limitations of this framework were subjected to a real stress test in January 2026. Japan’s Cabinet Office summoned X’s Japanese subsidiary over non-consensual images generated by the Grok model. The government’s entire leverage consisted of issuing written inquiries and threatening to issue administrative guidance. With no compulsory order power in hand, the summons ultimately produced neither a public compliance record nor any verifiable follow-up action. The case demonstrated in the bluntest terms the enforcement ceiling of a soft-law framework when facing an uncooperative party.
The mechanism of the proposed amendment is already taking shape: the government would gain the power to issue legally binding rectification orders to specified companies, and if the recipient refuses to comply, penalty provisions would be triggered. The triggers would focus on two scenarios: high-performance AI becoming uncontrollable, or being used for cyberterrorism. Regulatory coverage would extend from foundation model developers to service providers, explicitly including major U.S. IT companies. The three core elements include: how to define the specific threshold for “high-performance AI,” the form and maximum amount of fines, and the boundaries of official discretion when issuing rectification orders.
The invitation list for the closed-door meeting was itself a signal of intent: those invited to participate in policy discussions included Anthropic (developer of Claude), OpenAI Japan (the Japanese legal entity of ChatGPT), and Japanese telecom giant NTT. The joint presence of these three sketched the basic outline of the new regulatory coverage—multinational generative AI foundation model providers and domestic infrastructure operators are both in view. For Anthropic and OpenAI Japan, attending was both a gesture of cooperation and a sign that, once the amendment takes effect, they will be among the first entities subject to the new obligations.
From an industry-chain perspective, the impact of this shift is uneven across levels. Foundation model developers face the heaviest compliance costs: under the current “zero-obligation” framework, they need not set any statutory baseline for model safety capabilities in Japan; if the new framework takes effect, they will be required to establish auditable safety control systems and provide substantive cooperation when regulators request it. Whether application-layer service providers fall within the “high-performance AI” category depends on the final definition of the “high-performance” threshold—currently the biggest compliance uncertainty. Japanese domestic companies such as NTT, meanwhile, gain a relative advantage beyond compliance pressure: a regular channel of dialogue with regulators is itself a competitive moat.
Horizontally, Japan’s move aligns its regulatory path with the EU and Australia. The EU AI Act is known for risk tiers and fine caps based on a percentage of global turnover, representing the “heavyweight” path of mandatory regulation; Australia is likewise exploring legal obligations for high-risk AI systems. Within the G7 framework, Japan has been the major member that adhered to the soft-law path the longest, and the 2025 law’s deliberate avoidance of fines was a product of that stance. Now that this last “soft-law piece” is coming loose, it means the AI regulatory landscape within the G7 is converging faster, increasing pressure on companies to design global compliance systems to the strictest standards.
The direct trigger for this policy shift is the escalation of the security-threat narrative—from the earlier version’s “infringement of citizens’ rights” to national-security-level risk descriptions such as “cyberterrorism” and “AI out of control.” The choice of this narrative framework is strategically significant: it moves the justification for regulatory power from consumer protection (where political resistance is greater) to national security (where political consensus is easier to form), thereby clearing legislative obstacles for introducing mandatory provisions.
As a forward-looking judgment, completing the framework within the year does not mean the amended law will take effect in the same window. Japan’s legislative process usually requires cabinet approval, parliamentary deliberation, and formal promulgation, and the path from framework to implementation often spans one to two legislative sessions. There are three key signals to watch: first, whether the institutional restructuring document to be published by year-end includes a quantitative definition of “high-performance AI” (such as parameter count, compute thresholds, or capability baselines); second, whether participating companies such as Anthropic and OpenAI Japan begin to mention Japanese legal obligations in earnings reports or compliance disclosures; third, whether formal investigations or rectification-order cases based on the new legal framework emerge later in 2026. If the plan is slow to be published within the year, or if the published content remains mainly phrased as “soft goals,” that can be taken as a signal that the amendment process is being slowed by industry lobbying.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接