How to Read Legislation

Does Using AI Breach Confidentiality? Have Bar Associations Really Banned It?
Paper-cut illustration of a professional using AI while checking contracts and information handling

A conversation with a lawyer

Last week, a lawyer told me:

“Bar associations generally prohibit AI because confidential information enters the provider's servers and may breach confidentiality.”

I replied, “If confidential information can be stored on cloud servers, why is AI different?”

The answer was, “Apparently because AI analyzes the information.”

That conversation prompted me to investigate. My conclusion is:I could not verify a general bar-association prohibition in publicly available materials.“Because it analyzes” is also not a workable dividing line. However, despite the imprecise wording, part of the concern is valid.

Let us examine it step by step.

Do bar associations really prohibit AI in principle?

Within the public materials searched, I found no association rule or resolution generally prohibiting lawyers' own AI use. I found the following.

Issuer Document Status
Tokyo Bar Association, Legal Practice Reform Committee, Second Subcommittee Guidelines for Appropriate Use of Generative AI Services in Legal Practice (effective March 27, 2025) Recommendations for members
Japan Federation of Bar Associations AI Strategy Working Group Points to Note on the Use of Generative AI in Legal Practice Internal material; full text unreviewed because access is members-only
President, Japan Federation of Bar Associations Presidential statement (September 1, 2026) Statement of the president's position; addresses a separate issue discussed below

The Tokyo Bar Association's journal describes its guidelines as follows.

Recommendations for members, neither mandatory compliance rules nor intended as direct standards for conduct investigations or discipline

The message is guidance for reference, not discipline for noncompliance.It is not a prohibition.

The JFBA working group's guidance exists, but as a nonlawyer I could not access its members-only text. I therefore make no claim about its contents, which may contain stricter provisions.

For the Dai-Ichi Tokyo, Daini Tokyo, Osaka, and Aichi bar associations, I found no independent guidance within the search conducted. That means not found, not a definitive statement that none exists.

Why does it circulate as a general prohibition?

The search revealed how the report may arise.

Searching “lawyer,” “AI,” and “prohibition” produces an entirely different issue:Article 72 of the Attorney Act.

A person who is not an attorney or legal professional corporation may not, for remuneration, engage as a business in legal services—including opinions, representation, arbitration, or settlement—or arranging such services concerning litigation, noncontentious matters, administrative appeals including requests for review, reinvestigation, or further review, and other general legal matters, except where this Act or another law provides otherwise.

This is a rule thatnonlawyers may not perform legal services for compensation. It is invoked concerning AI providers offering services resembling legal advice. The JFBA president's September 1, 2026 statement also concerned the Ministry of Justice's approach to Article 72.

This is a rule thatIt does not prohibit lawyers themselves from using AI as a tool.Yet both issues use the words lawyer, AI, and prohibition. Mixing them creates the summary that bar associations ban AI.

What does confidentiality prohibit?

First examine Article 23 of the Attorney Act.

An attorney or former attorney has the right and duty to maintain secrets learned in professional duties, except where otherwise provided by law.

Article 23 of the JFBA Basic Rules on the Duties of Practicing Attorneys is more specific.

Without justifiable grounds, an attorney must not disclose or use client secrets learned in professional duties.

For tax accountants, Article 38 of the Certified Public Tax Accountant Act provides:

Without justifiable grounds, a certified public tax accountant must not disclose or misappropriate secrets learned in tax-accounting duties. The same applies after ceasing to be a tax accountant.

Article 59 provides imprisonment for up to 2 years or a fine up to JPY1 million. Prosecution requires a complaint. Attorneys are subject to Penal Code Article 134, but its named professions are physicians, pharmacists, pharmaceutical sellers, midwives, attorneys, defense counsel, and notaries;tax accountants are not included.Their provisions are contained in the tax accountant legislation.

What exactly do these provisions prohibit?Disclosure, use, and misappropriation. None says information must not be analyzed.

Why can cloud services be used?

Why, then, can client information be placed on cloud servers? Data reaches the provider's infrastructure and, in that sense, leaves the office.

Article 25 of the Act on the Protection of Personal Information offers a clue.

When a personal information handling business operator entrusts all or part of personal-data handling, it must exercise necessary and appropriate supervision over the entrusted party to ensure secure management of that data.

The provision does not prohibit outsourcing itself.It permits entrusting data subject to supervision.

External services may be used under appropriate contracts and controls. However, fulfilling data-protection supervision duties and professional confidentiality are separate questions. Assess the secret, purpose, necessary scope, client contract or consent, provider terms, security, and actual operation. Neither external use alone nor the existence of an outsourcing contract determines the answer.

The same assessment applies to AI.Whether analysis occurs does not distinguish cloud services from AI.

Almost no cloud service avoids analysis entirely: malware scanning, search indexing, spam filtering, and deduplication all involve machine reading. If analysis were the dividing line, cloud services would also be unusable.

AI nevertheless has distinct risks

The argument so far challenges the explanation, not the concern itself.AI presents one risk absent from ordinary cloud storage:

if inputs train the model, they become incorporated into the model itself.This differs qualitatively from storing a file. Incorporated information may in principle influence responses to other users. This is probably what the lawyer meant by analysis.

Two further issues arise. First, provider staff reviewing interactions for safety may expand who can see them. Second,consumer free services and business contracts are easily confused.Some consumer services have settings allowing inputs to be used for training.

The question is not AI versus cloud, but which contract

This third issue is important. Reading providers' terms shows thatthe same provider's AI can have very different consumer and business conditions.

Anthropic's commercial terms for Claude state:

Anthropic may not train models on Customer Content from Services.

In substance, models must not be trained on customer content obtained through the services. Its consumer terms, by contrast, provide for training by default with an opt-out setting.

OpenAI's business terms contain a similarly structured provision.

OpenAI will not use Customer Content to develop or improve the Services, unless Customer explicitly agrees to such use.

Google's Workspace terms also provide similarly.

Google will not use Customer Data to train or fine-tune any of its generative artificial intelligence models supporting the Google Workspace Generative AI Services without Customer’s prior permission or instruction.

All three default to not training on customer data for business services. Consumer services either use data by default or require users to select a no-training setting.

The right question is therefore not simply whether AI may be used, but under which contract.

Three questions for explaining the position to clients

These three questions begin the service review. They are not new criteria; they resemble questions already asked when selecting cloud accounting, e-filing, and email providers.

Question Where to check
Will inputs be used to train models? Terms of service and data policies
Where is data retained, and for how long? Retention periods and deletion options
Which provider personnel may access it, and in what circumstances? Safety-review and abuse-investigation provisions and subcontractor terms

These are starting points for service selection, not sufficient grounds to enter client secrets. Rather than a blanket AI ban, assess contracts, terms, security, and operation according to the information and purpose.

I recommend three practical steps.

  • Use business contracts for office AI, not consumer accounts. Do not paste client information into free consumer services
  • Read and record the provider's terms against the three questions, including the date because terms change
  • Prepare a one-page office policy so you can answer clients asking whether AI is used

Finally, the limits of this article:I could not read the members-only JFBA guidance.It may go further than public materials. Lawyers should first consult their own association's documents. This article addresses only what is externally visible.

Sources

This paper is a general explanation based on laws and regulations as of September 12, 2026. Please always consult individually before implementation.

We take care in preparing this information. If you notice an error, please let us know through our contact page.

Related Articles

These articles explain related systems and practical issues.

Why Is Tax Law So Difficult? Reading Parenthetical Clauses with Color Coding

Back to the list