AI and Client Confidentiality
On September 17, 2026, the Ethics Committee of the Israeli Bar Association published an updated opinion on the use of AI tools by lawyers (Opinion 60/24)¹. It is detailed, serious, and largely correct. It addresses professional responsibility, verification, confidentiality, informed consent, and autonomous action.
Its starting point is right. AI is an aid to the lawyer. Responsibility remains with the lawyer.
That is how I work. I give instructions, test the answers, challenge the assumptions, and decide what can be used. The system assists. I remain responsible for the client and the work.
But directing an assistant is one task. Knowing what happens to the information entrusted to it is another. The opinion establishes sound ethical principles. What it does not yet provide is the practical implementation guidance that working lawyers need to apply those principles every day.
Lawyers have always worked with assistants: trainees, secretaries, colleagues, and external specialists. Clients understand that their lawyer does not perform every operation personally.
AI occupies an assistant's place in that working process. It can organize documents, compare accounts, identify inconsistencies, and prepare material for examination. Its ability to produce a convincing answer does not give it authority to make a professional decision.
The analogy has limits. Supervising the assistant's output does not establish who can access the underlying information, where it is retained, or which contractual protections apply. The lawyer must examine both: the quality of the assistance and the conditions under which it is provided.
The obligation the opinion places on the lawyer therefore covers not only results — it covers disclosure, informed consent where required, and the assessment of the platform before any restricted information is supplied.
Some clients have an exceptional requirement: particular information must remain entirely off networked systems.
A client who requires that restriction should say so expressly, before supplying the material. "Keep this confidential" and "do not place this on any networked system" are different instructions. The first is an ordinary professional obligation. The second imposes a specific working constraint that the lawyer cannot reliably infer from silence.
This does not transfer the lawyer's confidentiality duties to the client. Nor does silence authorize unrestricted AI use. The lawyer must still assess the information, the proposed system, and any need for disclosure or consent.
A client who prohibits networked processing may require a different workflow, more time, or a different professional arrangement entirely.
The opinion establishes the right principles. The gap is between those principles and the decisions a lawyer must make in daily practice.
Many lawyers store client files on Dropbox or Google Drive. Both platforms now offer AI features built into the standard interface: summarizing documents, answering questions about files, identifying connections across a document set. These features are increasingly part of routine use.
When a lawyer uses those features, client content is processed by AI infrastructure — in Dropbox's case through a third-party provider arrangement, in Google's case through Google's own systems. The data leaves the document storage layer and enters an AI processing layer. Retention periods, contractual protections, and account-level settings vary significantly across products and account types and cannot be assumed to be short or uniform.
A lawyer who carefully avoids uploading a client file to a dedicated AI assistant may use an equivalent feature within an established cloud service without examining how that file is processed.
Familiarity can conceal a material change in the working arrangement. The relevant questions should follow the information wherever it goes: which service processes it, which account type and settings govern the operation, whether another provider receives it, what is retained, for how long, and under what exceptions. The location of the button does not answer those questions.
There is a further reason to examine the information flow carefully — one the opinion itself recognizes.
A folder may contain a contract, correspondence, financial records, and a draft pleading. An AI assistant reads across them and produces something new: a compact account of the dispute, the weak point in a defence, an unexplained payment, a contradiction between witnesses, a possible settlement strategy.
The opinion states this directly: where a system uses information for cross-processing or creating connections between data, the risk of unauthorized leakage increases — even without any deliberate intent on the user's part.
The provider may therefore process and retain not only the source documents, but also the analytical layer created from them. A short conversation can contain the lawyer's questions, provisional assessments, rejected arguments, and instructions about the next move. Calling it "a chat" understates what it may hold.
Lawyers should protect the analysis of a client's matter with the same care they apply to the documents themselves.
There is a widespread misconception: that opting out of model training means conversations are not retained. It does not.
These protections address different questions. "No training" concerns one particular use of the information. Retention concerns how long it remains on the provider's servers. Access controls concern who can reach it. Contractual commitments determine what the provider undertakes to do.
On Claude's standard consumer plans, conversations are retained in the conversation history until the user deletes them, subject to the provider's policies. Opting out of training is a separate action that does not trigger deletion. A session containing client information from months ago may still be on the provider's servers unless the lawyer deleted it.
The opinion's distinction between open and closed platforms must be applied to the actual service arrangement — the account type, version, settings, safeguards, and enforceable commitments. A product name or a "business" or "enterprise" label is not sufficient. Classification depends on what the arrangement actually provides.
For a lawyer using a service that retains conversations, a practical discipline is available:
Without this, conversations containing client material remain in the history with no automatic expiry.
The saved record should contain what is needed for continuity — not become another unnecessary copy of the entire confidential exchange.
Deletion reduces unnecessary retention. It does not undo the original disclosure, remove every possible copy, or make an unsuitable platform suitable. The assessment must begin before the information is supplied.
The opinion correctly identifies the problem and the principles. What working lawyers need next is implementation guidance they can act on.
That guidance should address dedicated AI assistants, AI features embedded in Dropbox, Google Drive, and similar office tools, connected document systems, and the materially different conditions attached to consumer, business, and API use across different providers. It should explain what contractual assurances to seek, how to assess account-level settings, and how to document that assessment.
It should also be designed for small practices. A solo practitioner or small firm cannot satisfy a requirement they cannot translate into a specific decision about a specific tool.
AI can reduce time spent sorting, comparing, searching, and drafting. It also creates work: preparing instructions, checking results, correcting errors, and managing information responsibly.
The lawyer still absorbs the material, directs the case, appears in court, negotiates, signs, and takes responsibility for the advice. In my experience, a matter that once required a hundred hours may now require thirty to fifty. That reduction is real, and clients are entitled to benefit from it.
The charging basis should be explained clearly. Where work is billed by the hour, the charge should reflect the time actually spent. Fixed fees involve a different assessment — scope, judgment, and responsibility undertaken. AI-related costs that form part of the service belong in that explanation.
Efficiency should improve the service and be reflected honestly in its price.
The alternatives extend beyond unrestricted AI use and an entirely paper office. A lawyer may use digital systems while restricting AI access, work with selected or appropriately anonymized material, or adopt a service arrangement offering suitable contractual protections. Exceptionally sensitive matters may require stronger limitations.
Where consent to AI use is required, it must concern the actual proposed arrangement and its material conditions — not a general clause.
Confidentiality belongs to every client. It does not depend on whether disclosure would cause inconvenience or catastrophe.
Keeping AI at assistant level is an essential professional discipline. The lawyer directs the work, tests the output, and decides what happens next.
That discipline must extend to the information itself: what the assistant receives, what the provider may do with it, and what remains after the task is finished.
The question is therefore more than whether I checked the answer.
It is whether I know what remains after I close the window.
¹ Israeli Bar Association Ethics Committee, Opinion 60/24 (Updated) — The Use of Artificial Intelligence Tools in the Work of Lawyers in Israel, September 17, 2026. Available at: israelbar.cld.bz/ai-based-tools-for-lawyers-opinion-60-24