AI Regulation Tracker  /  Professional guidance

Japan's bar federation issues 5-point generative-AI cautions for lawyers, flags Article 72 line

The Japan Federation of Bar Associations has published operational cautions on generative AI for practicing lawyers, covering output verification, client confidentiality, and the unauthorized-practice boundary. It is guidance, not a binding rule, but the most detailed guidance the federation has issued on lawyers' use of generative AI.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

Japan's national bar body has told its members what many courts abroad are now enforcing the hard way: a lawyer who lets generative AI touch a matter still owns every word that leaves the office. In September 2025, the Japan Federation of Bar Associations, acting through its AI Strategy Working Group, published a set of five operational cautions on generative-AI use in legal practice. The document is short and practical, and it is aimed squarely at lawyers who are already pasting facts into chat tools.

What the guidance says

The five points track the failure modes that have generated the most trouble for lawyers using these tools. First, verify AI output, because generative models produce confident, well-formatted case law that does not exist. Second, protect client confidentiality when inputting information to an external AI service, since the act of entering data can amount to disclosure to a third party, the AI provider. Third, watch Attorneys Act Article 72, the provision that restricts non-lawyers from providing legal services, because AI-driven legal tools can drift across that line. The remaining cautions address conflicts of interest and general accuracy. Together they form a checklist a working lawyer can apply to a single matter, not a policy abstraction.

The confidentiality point deserves emphasis because it is the one most easily overlooked in the rush of a busy practice. Entering a client's facts, draft, or opposing party's name into a public generative-AI service can constitute disclosure to the provider, and depending on the service, that data may be retained or reused to train the underlying model. The safe default is to assume anything typed into an external tool has left the firm's control. Lawyers who want to use these tools without breaching confidentiality need to know how a given service handles inputs, whether an enterprise configuration disables data reuse, and what the client has and has not consented to.

What it does not do

The guidance is explicit about its own weight. It is not the JFBA's official binding position. It is guidance, the most authoritative statement the federation has issued on the subject to date, but it does not by itself create a new disciplinary rule or a bright-line penalty. That distinction matters. A Japanese lawyer who ignores the cautions is not automatically in breach of a code provision. What the document does is set a standard of reasonable practice that a disciplinary body, a court, or a client could later measure conduct against. The professional duties it draws on, competence, confidentiality, and good faith, already exist. The cautions simply apply them to a new tool.

The Article 72 line

The most distinctly Japanese element is the treatment of Article 72 of the Attorneys Act. That provision limits the provision of legal services by non-lawyers. The working group frames the concern for AI: a generative tool that answers legal questions for the public, or that a firm deploys as a client-facing legal service, can raise the question of whether legal services are being provided by something other than a licensed lawyer. For lawyers, the practical reading is that the tool is an aid to the lawyer's own judgment, not a substitute lawyer. Keeping a qualified human in the decision seat is both a competence point and an Article 72 point.

Why a US reader should care

For a US firm with a Tokyo office or Japanese clients, the guidance does not bind US-licensed lawyers, but it signals the standard local co-counsel and courts will expect. It also rhymes with what US state bars have already said. California and Florida have issued AI ethics opinions built on the same duties of confidentiality, competence, and candor. And the verification point is no longer theoretical anywhere: courts in the United States have imposed sanctions on lawyers who filed briefs citing cases invented by AI. The JFBA cautions read as a pre-emptive shield against that same outcome reaching Japanese courts. A lawyer who wants one rule that travels across borders can use the strictest version of each point and be safe in every jurisdiction.

The near-term takeaway is unglamorous. Verification is a task, not a setting. Confidentiality is a decision made before the prompt, not after. And the human lawyer, not the model, remains the one the bar and the client can hold to account.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.