How to Give Attorneys AI Legal Research Tools Without a Confidentiality Gap

Where does the confidentiality gap come from?

The gap opens when an attorney uses whatever AI legal research tool is closest at hand, types in facts from a matter, and that tool's terms let the provider keep, review or train on what was entered. A new tool does not change the duty. In California, Business and Professions Code section 6068(e)(1) requires an attorney "to maintain inviolate the confidence, and at every peril to himself or herself to preserve the secrets, of his or her client."

Research feels like a safe use because the questions sound abstract. In practice, a good research prompt carries facts: a party name, a deal term, an unusual procedural posture, a theory of the case. Those details are what make the answer useful, and they are also what make the prompt confidential. The usual cause is a dozen attorneys each choosing their own tool, on their own account, under terms nobody at the firm has read.

What do the ethics rules say about AI and client information?

The ethics rules apply duties attorneys already have (confidentiality, competence and supervision) to a new kind of vendor. ABA Formal Opinion 512, issued July 29, 2024, walks through competence, confidentiality, communication, supervision and fees as they apply to generative AI. It calls for informed client consent before client confidences go into a self-learning AI tool, and it says boilerplate consent language in an engagement letter is not adequate.

The State Bar of California's Practical Guidance for the Use of Generative Artificial Intelligence is more direct. It says a lawyer must not input confidential client information into any generative AI solution that lacks adequate confidentiality and security protections. It tells lawyers to review the terms of use to determine how a product uses inputs, to anonymize client information and avoid details that identify the client, and to consult IT professionals or cybersecurity experts about security and data retention. It also says managerial and supervisory lawyers should set clear policies on permissible uses and provide training.

Competence covers the technology itself. Comment 1 to California Rule 1.1 says the duty includes keeping abreast of "the benefits and risks associated with relevant technology." And one federal court has already looked at a public AI tool through a privilege lens: in United States v. Heppner (S.D.N.Y. Feb. 17, 2026), the court held that documents a defendant generated on a public AI platform on his own initiative were not privileged, pointing to the platform's policy allowing it to retain and disclose inputs. That ruling binds only that case, but it shows how a court reads consumer terms.

What makes AI legal research tools safe enough for client work?

An AI legal research tool is safe enough for client work when its contract keeps inputs confidential and the firm controls who signs in. Consumer and free tiers are written for individuals, and their terms often allow the provider broader use of what users type. Business and enterprise plans, including the AI features inside legal research subscriptions, usually offer stronger commitments, but terms vary by plan and change over time. Read the current agreement for the plan the firm will actually buy, and check for these points:

• A written commitment that the provider will not train its models on firm inputs or outputs.
• A defined retention period for prompts and files, and a way to delete them.
• Limits on human review of inputs by the provider, with any exceptions spelled out.
• No disclosure of inputs to third parties except subprocessors the firm can see listed.
• An independent security report, such as a SOC 2 Type II report, available on request.
• Sign-in through the firm's identity provider with multifactor authentication.
• An admin console with audit logs the firm can export.

Record what you found and the date you reviewed it. When a client or carrier asks how the firm vetted its AI tools, that record is the answer.

How do you roll out AI for legal research step by step?

Roll out AI for legal research by starting narrow, configuring the tool before launch and giving attorneys one approved path so they stop improvising.

1. Choose one approved tool for each use case (research, drafting, summarizing) and document the terms review.
2. Sign the business agreement and turn off any data sharing or model improvement settings in the admin console.
3. Require sign-in with firm credentials and multifactor authentication, and block personal AI accounts on firm devices.
4. Block or warn on unapproved AI sites with web filtering, and set browser data loss prevention rules for client names, matter numbers and documents labeled confidential.
5. Decide, with the firm's ethics counsel, which matters or uses need client consent, and build that into intake. Opinion 512 says boilerplate engagement letter language will not do.
6. Train attorneys and staff on what may go into the tool, how to strip details that identify the client, and how to verify every citation and quotation before relying on it.
7. Turn on prompt logging and review a sample each month for confidential data and for use outside the approved purpose.
8. Re-read the terms at every renewal. Providers revise them, and a change in retention or training terms changes your risk.

Do attorneys still need to verify AI research results?

Yes. The California guidance warns that generative AI output can be false, inaccurate or biased, and says a lawyer must critically review, validate and correct both the input and the output. A research tool that cites real-looking cases is still a starting point. Treat verification as its own control, train it alongside confidentiality, and make the supervising attorney responsible for checking that it happens.

What should the firm do this quarter?

Close the gap in the order that removes the most exposure first.

• Ask attorneys and staff which AI tools they already use for research, and on which accounts.
• Pick the approved tools, review their terms against the list above and sign business agreements.
• Configure firm sign-in, multifactor authentication, logging and data loss prevention before announcing the tools.
• Publish a one-page law firm AI policy naming the approved tools and the information that never goes into any AI tool.
• Review consent practice with ethics counsel and update intake where needed.
• Schedule a short training and a monthly log review, and put a terms review on the renewal calendar.

How s90 helps

s90 sets up AI platform policy enforcement so approved research tools work only through firm sign-in, while unapproved AI sites are blocked or flagged. We configure data loss prevention for matter documents and client identifiers, and prompt monitoring so the firm can see what is being sent, by whom and to which tool. If you want a second set of eyes on an AI tool's terms or settings before rollout, we are glad to walk through it with you.