Policy

Japan to Set Up an Expert Panel on Who Is Liable When AI Agents Cause Harm, Aiming to Conclude by Summer 2027

· 6 min read · Policy

Japan to Set Up an Expert Panel on Who Is Liable When AI Agents Cause Harm, Aiming to Conclude by Summer 2027

Japan’s government will soon set up an expert panel to work out who bears legal responsibility when AI agents, the services that shop, book travel or sign up for things on a user’s behalf, cause damage, the Asahi Shimbun reported on November 19. The panel will sit under the AI and Digital Reform Promotion Council, the AI policy control tower launched in October, and is expected to wrap up its discussion around summer 2027. According to the newspaper, it is the first time the government will examine the legal liability of AI agents in earnest.

The gap is easy to describe. If an agent orders more than its user intended and a shop or the user loses money, current law gives no clear answer on whether the user, the developer or the service operator is liable. The question is separate from, but runs alongside, the review of Japan’s AI Promotion Act. On November 18, a Cabinet Office working group circulated an outline that would require providers of high-performance AI to take safety measures and report to the state, with powers for the government to issue recommendations and orders. As reported by Nihon Jiji Shimbun, the outline favours “indirect” penalties that apply only when a company ignores a request for information or an order, rather than punishing development or use itself, and says foreign providers with meaningful use in Japan should be covered.

What it means in practice

Japan to Set Up an Expert Panel on Who Is Liable When AI Agents Cause Harm, Aiming to Conclude by Summer 2027 — contextual photo

Nothing is settled yet. The outline does not set capability thresholds, penalty amounts or a date for a bill, and the agent panel has only been announced. For companies already putting agents into customer-facing services, the direction is visible, though: officials have discussed securing safety at the level of the whole system, including authentication, permission controls, guardrails and audit logs, rather than judging the model alone. Firms that can show who authorised an agent’s action, and within what limits, will be better placed whatever the final rules say. The working group heard from companies earlier in November, including Anthropic, Google and SB Intuitions, with views ranging from support for clear rules on frontier developers to caution about rigid duties before release.

The move builds on the communications ministry’s revision of its AI business guidelines to cover agents and the review of the AI Act that first put penalties on the table. What to watch is who is appointed to the panel, whether the Cabinet Office’s interim summary at year-end adopts the indirect-penalty model, and how the role of the AI Safety Institute (AISI) is expanded.

← Back to AI4Japan