Skip to main content
    Back to Blog
    17 min readLast reviewed: Reviewed by Zachary Brenner, Co-Founder & CEO
    Legal
    Confidentiality
    ABA
    ChatGPT
    AI Compliance

    Can Lawyers Use ChatGPT? Confidentiality, Privilege, and AI in 2026

    T

    Team Sonomos · Sonomos

    The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.

    Short answer: Yes, lawyers can use ChatGPT — and many do — but doing so without violating the duty of confidentiality, attorney-client privilege, or your jurisdiction's competence rules requires a deliberate workflow. As of 2026, the ABA and a growing number of state bars have issued guidance making clear that entering client information into an AI tool whose terms allow the vendor to share it or train on it can breach the duty of confidentiality. The good news: with the right product tier, the right contractual terms, and a habit of redacting at the source, generative AI is genuinely useful in the practice of law.

    This guide explains the rules that matter, the specific risks unique to legal work, and the workflow patterns that let lawyers benefit from AI without exposing the firm or the client.

    What the rules actually say

    The American Bar Association's Model Rules of Professional Conduct, which most US jurisdictions track closely, are explicit on three obligations that apply directly to AI use:

    • Rule 1.1 — Competence. A lawyer must provide competent representation. Comment [8] to the rule (the "duty of technological competence") directs lawyers to keep abreast of "the benefits and risks associated with relevant technology"; it has been adopted by 40 states plus D.C. and Puerto Rico.
    • Rule 1.6 — Confidentiality. A lawyer "shall not reveal information relating to the representation of a client unless the client gives informed consent, the disclosure is impliedly authorized in order to carry out the representation, or the disclosure is permitted by" Rule 1.6(b). Rule 1.6(c) requires "reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client." Comment [18] lists the factors that bear on whether efforts are reasonable, including the sensitivity of the information and the likelihood of disclosure.
    • Rule 5.3 — Responsibilities Regarding Nonlawyer Assistance. Lawyers with supervisory authority have a duty to make reasonable efforts to ensure that a nonlawyer's conduct conforms to the lawyer's professional obligations — a duty the ABA has said extends to generative AI providers and tools.

    On July 29, 2024 the ABA's Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, "Generative Artificial Intelligence Tools," confirming that all three rules apply to GenAI use. State guidance now exists alongside it: the State Bar of California's Practical Guidance (first issued November 2023 and replaced by a 2026 edition), Florida Bar Ethics Opinion 24-1 (January 2024), the New Jersey Supreme Court's preliminary guidelines (January 2024), D.C. Bar Opinion 388 (April 2024), the Pennsylvania and Philadelphia bars' Joint Formal Opinion 2024-200, the New York City Bar's Formal Opinion 2024-5 (August 2024), and Texas Opinion 705 (February 2025). The California and Florida documents preceded the ABA opinion, which cites both.

    The opinions are not identical. They differ on consent: the ABA requires informed consent before information relating to the representation is entered into a self-learning tool; Florida recommends consent whenever confidential information would be disclosed to a third-party tool; New York City and California tie the requirement to whether the tool shares inputs or lacks adequate security; Texas says consent "may" be needed; D.C. cautions against asking clients to consent to wide-ranging disclosures. All agree that no consent is needed when no information relating to the representation is entered. Beyond consent, they converge on the following:

    • Lawyers must understand the tool well enough to advise the client about it (Rule 1.1).
    • Outputs must be independently verified for accuracy (Rule 1.1, with several courts now sanctioning lawyers for unverified AI-generated citations).
    • Fee arrangements should account for AI use (Rule 1.5): hourly billers may charge only for time actually spent, and not for time spent learning the tool.

    The phrase "reasonable efforts" in Rule 1.6(c) is doing real work here. It is the same standard that applies to email, cloud storage, and outsourced document review — and the opinions draw on their earlier cloud-computing and outsourcing guidance when applying it to AI.

    Three properties of legal work make casual AI use riskier than it looks:

    1. Privilege is binary. Attorney-client privilege protects confidential communications between lawyer and client made for the purpose of legal advice. Disclosure to a third party can waive privilege over what was disclosed; in February 2026 a federal court held in United States v. Heppner that privileged information a defendant typed into a consumer AI tool was waived "just as if he had shared it with any other third party." Whether a waiver extends to related, undisclosed material is governed by Federal Rule of Evidence 502(a) and is limited to intentional waivers where fairness requires it. Unlike confidentiality (which can sometimes be repaired), privilege waiver is hard to undo.
    2. The work product doctrine is fact-sensitive. Materials prepared in anticipation of litigation receive separate protection. Disclosure to an AI tool creates an argument that the materials are no longer "kept confidential" in the way the doctrine assumes. Courts have split: in 2026 a Texas Business Court, agreeing with two federal district courts, held that a party's ChatGPT conversations prepared for litigation remained work product because the platform is not an adversary, while the Heppner court denied work-product protection on different facts because the documents were not prepared by or at the behest of counsel.
    3. Conflict checks rely on metadata. Pasting a brief or a client list into a consumer AI service is, among other things, an uncontrolled export of conflict-relevant information. Future engagements at the firm — and any audit of the conflict-check process — depend on knowing where that information has been.

    Compare a doctor pasting a patient name into ChatGPT to a lawyer pasting an opposing-party draft motion. The doctor faces HIPAA exposure; the lawyer faces HIPAA-equivalent confidentiality duties plus the possibility of waiving privilege in pending litigation.

    The product-tier question

    For lawyers, the consumer-versus-enterprise distinction is more than a billing matter. As of November 2026, per the vendors' documentation, the relevant defaults are:

    • ChatGPT Free / Plus / Pro (consumer) — Conversations may be used to train models unless the user turns off "Improve the model for everyone"; no DPA; not eligible for OpenAI's BAA.
    • ChatGPT Business (formerly Team) — Not used for training by default; business terms apply; OpenAI's HIPAA guide excludes it from BAA coverage.
    • ChatGPT Enterprise / Edu — Not used for training by default; workspace admins control retention; enterprise agreement and DPA; SOC 2 Type 2; BAA available where OpenAI's HIPAA guide specifies.
    • OpenAI API — Not used for training by default since March 1, 2023; abuse-monitoring logs retained for up to 30 days by default; zero data retention subject to OpenAI's prior approval; DPA available.
    • Anthropic Claude — Consumer plans (Free, Pro, Max) are asked whether to allow their chats to be used for training; when that setting is on, data may be retained for up to five years. Claude for Work, Government, Education and the API are not used for training by default; DPAs and zero-retention arrangements are available for eligible commercial customers. What that does and does not make confidential is set out in Is Claude Confidential?.
    • Google Gemini (consumer accounts) — Human reviewers may read some chats, and reviewed chats are retained for up to three years; Google's own guidance says not to enter confidential information. Workspace and Vertex AI are governed by separate enterprise terms.
    • Microsoft 365 Copilot and Copilot Chat (enterprise data protection) — Prompts and responses are not used to train foundation models when the user is signed in with a work (Entra) account; covered by Microsoft's DPA. Microsoft retired the earlier "Commercial Data Protection" name in September 2024.

    Reasonable efforts under Rule 1.6(c) are much easier to demonstrate on a properly configured enterprise tier than on a consumer plan. No bar has declared consumer tools categorically off-limits, but D.C. Bar Opinion 388 and New York City Bar Opinion 2024-5 both direct lawyers to the vendor's terms — whether inputs are shared or used for training — which is where free and personal accounts tend to fail.

    A workflow that works in practice

    Below is a pattern that tracks the published guidance:

    1. Tier selection. Use a contracted enterprise plan (ChatGPT Enterprise, Claude for Work, Microsoft 365 Copilot, or comparable) under a DPA, with retention set to the shortest period the plan's admin controls allow. Block consumer endpoints at the firm's network or via an identity policy.
    2. Local redaction on the device. Run Sonomos Locke on devices used for client work. Locke runs on the user's device as a desktop app; its companion browser extension screens AI websites, and a local proxy screens command-line, script and SDK traffic. Locke 1.0 detects names, addresses, account numbers, and other identifiers locally and replaces them with one-way placeholders before the prompt leaves the device; reversible tokenization is planned shortly after 1.0. Coverage depends on the app, platform and settings. This reduces what reaches the AI endpoint and supports a defensible minimisation posture. It does not by itself take a document outside Rule 1.6, which covers all information relating to the representation regardless of its source — the terms of a client's agreement can be confidential with the names removed. The legal determination is the firm's or its counsel's.
    3. Document the policy. Update the engagement letter, the firm's information-security policy, and the conflict-check procedure to reflect AI use. ABA Formal Opinion 512 warns that boilerplate engagement-letter provisions purporting to authorize AI use are not sufficient to obtain informed consent; consent must be specific to the tool, the information at risk, and the ways it could be used against the client.
    4. Verify outputs. Treat every AI-generated citation, quotation, and rule statement as an unverified claim from a junior researcher. Multiple courts in 2023–2025 have sanctioned attorneys for filing briefs with hallucinated cases, beginning with the $5,000 penalty in Mata v. Avianca (S.D.N.Y. June 22, 2023); the verification step is no longer optional.
    5. Train and audit. Annual training, plus periodic walk-throughs of representative matters, tends to catch process drift faster than self-reporting alone.

    Examples: what is and is not safe

    ScenarioAssessmentWhy
    Paste a client's NDA into ChatGPT Plus to "make it more readable"RiskyConsumer tier, no DPA, no client consent
    Paste the same NDA after replacing names and amounts with placeholders into ChatGPT PlusLower riskLess reaches the endpoint, but non-identifying deal terms can still be information relating to the representation, and consumer terms still apply
    Use ChatGPT Enterprise under a DPA, with admin-controlled retention, to summarize a deposition transcript with the client's name in itAcceptable, with caveatsReasonable efforts are in place; verify output
    Use a local-first redaction tool that masks client identifiers, then paste into Claude for WorkLower riskReduces what reaches the endpoint; Claude for Work is not used for training by default
    Use any tool to draft a brief and file it without verifying citationsNot acceptableRule 1.1 violation regardless of tier
    Discuss strategy in a public ChatGPT thread that may be indexedNot acceptableSharing a conversation publicly is a disclosure to anyone who finds the link; Rule 1.6(c) requires reasonable efforts to prevent unauthorized access

    Frequently asked questions

    Does using ChatGPT waive attorney-client privilege?

    It can. In United States v. Heppner (S.D.N.Y. Feb. 17, 2026), the first ruling on the question, the court held that a defendant's exchanges with a consumer AI tool (Claude) were not privileged — the tool is not an attorney, and the vendor's consumer privacy policy defeated any expectation of confidentiality — and that privileged information he typed in was waived. The court left open whether counsel-directed use could be treated like a lawyer's agent. On work product, the picture is different: in Tate Group Automotive v. Legacy Automotive Capital (Tex. Bus. Ct., June 3, 2026) the court, agreeing with two federal district courts it quoted, held that a party's ChatGPT conversations were protected work product because disclosure to the platform is not disclosure to an adversary. No court has yet ruled on an enterprise tool used under a confidentiality agreement. The Sedona Conference formed a drafting team in June 2026 with a final commentary targeted for 2027; state ethics opinions address Rule 1.6, not evidentiary privilege.

    ABA Formal Opinion 512 requires informed consent before information relating to the representation is entered into a self-learning tool, and separately asks under Rule 1.4 whether the client should be told about AI use at all — always, if the client asks or the engagement terms require it. It also says boilerplate engagement-letter language is not sufficient to obtain that consent. Several state opinions (Florida, New York City) recommend consent whenever confidential information will reach a third-party tool, and all agree none is needed when no client information is entered. Reducing what is entered, including through local-first redaction, narrows the question; it does not answer it.

    Are there any AI uses that don't implicate Rule 1.6 at all?

    Yes, when the input contains no information relating to the representation of a client. Drafting marketing copy, brainstorming research questions in the abstract, summarizing public statutes — all generally fine. Once the prompt includes any information the client expects to be confidential, the rule attaches.

    Has anyone been disciplined for AI use yet?

    Yes. Reported discipline and sanctions so far have centered on fabricated citations (Rule 1.1, Rule 3.3), beginning with Mata v. Avianca in 2023. Regulators have signalled that confidentiality lapses through personal AI accounts are the next concern, but published confidentiality cases remain scarce. Plan for both failure modes.

    Is "I redacted by hand before pasting" a defensible position?

    Better than nothing, but hard to sustain. Manual redaction misses entries under deadline pressure and creates an audit trail of "I was careful most of the time." An automated, local detector with one-way masking applies the same rules every time and leaves a record of what was masked, which is easier to document as part of the firm's reasonable efforts under Rule 1.6(c). Neither approach takes the remaining text outside Rule 1.6; both reduce what reaches the endpoint. The legal determination is the firm's or its counsel's.

    What about discovery and litigation holds?

    If client information has been pasted into a third-party AI tool, that tool's records may now be in scope for discovery and for any litigation hold the firm must implement. The firm's policy should include a procedure for preserving and producing AI-related records when applicable, and for confirming with the AI vendor what records exist and for how long. We look at why the duty to preserve reaches employees' AI conversations, and why personal accounts make it hard to meet, in Your AI chats are evidence.

    • Adopt an AI policy. Specify which products are approved, which are blocked, and the consent requirements for each matter type.
    • Choose enterprise-tier products with DPAs and the shortest retention the plan allows; block consumer endpoints.
    • Run local-first redaction on devices used for client work, so that less reaches the AI endpoint when policy slips.
    • Update engagement letters and the firm's information-security policy.
    • Train every lawyer and paralegal on Rule 1.1, Rule 1.6, and the firm's specific policy.
    • Verify every AI-generated citation and every legal proposition before relying on it.
    • Audit annually and after any new product or feature is rolled out.

    The bottom line

    For lawyers, AI is a competence question, a confidentiality question, and a workflow question all at once. The rules in 2026 are clear enough to act on: choose tools whose configuration you can demonstrate as "reasonable efforts," reduce what leaves the device by removing client identifiers before the prompt is sent, document what you do, and verify what comes back. Done well, AI saves real time on the work clients pay for. Done casually, it is a discipline case waiting to happen.

    For the specific research tools — Harvey, Westlaw AI, Lexis+ AI, and CoCounsel — and the hallucination risks to know about, see AI Tools for Legal Research in 2026.

    Sources and primary references

    Reviewed November 2, 2026. Regulations are cited to their official text. Vendor statements are cited to the vendor's own documentation as it read on that date; vendors revise these pages, so confirm before relying on them.

    Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works

    This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.

    Reference pages for this post