Most attorneys are already using ChatGPT, Claude, or Gemini for drafting, research, and summarization. And most of them are sending prompts that contain client names, case numbers, financial figures, and other information they are ethically obligated to protect.
This is not a theoretical risk. ABA Formal Opinion 512, issued in July 2024, makes clear that Rule 1.6's confidentiality obligations apply directly to AI use — and a 2026 federal court ruling has added a second layer of exposure that many attorneys haven't absorbed yet.
This article explains what the rule requires, what the current legal landscape looks like, and what "reasonable precautions" actually means in practice in 2026.
Model Rule 1.6(a) prohibits attorneys from revealing "information relating to the representation of a client" without informed consent or a recognized exception. Rule 1.6(c) adds an affirmative obligation: lawyers must "make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client."
Two points matter here. First, the rule covers all information relating to the representation — not just attorney-client privileged communications, not just documents marked confidential. A client's name combined with any contextual detail about their legal matter is covered. Second, the "reasonable efforts" standard is not absolute, but it does require attorneys to actively take precautions.
On July 29, 2024, the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512 — its first formal guidance specifically addressing generative AI. The opinion identifies four ethical obligations that apply when lawyers use AI tools:
The opinion explicitly flags consumer AI tools — where prompts may be retained, reviewed by vendor employees, or used to train future models. Consumer ChatGPT (Free, Plus, and Pro tiers) trains on conversations by default unless users opt out. OpenAI is also subject to a court order in the New York Times v. OpenAI litigation requiring it to preserve chat logs, including deleted conversations, indefinitely.
Key takeaway from Opinion 512: A lawyer who sends confidential client information to a consumer AI tool is disclosing that information to a third-party vendor — which is a prima facie violation of Rule 1.6 unless the attorney has taken reasonable precautions or obtained client consent.
In United States v. Heppner (S.D.N.Y. 2026), Judge Jed Rakoff held that documents generated using consumer AI tools are not protected by attorney-client privilege or the work-product doctrine. The court's reasoning: an AI tool is not an attorney, owes no duty of confidentiality, and cannot form an attorney-client relationship — so communications routed through a third-party AI system do not carry privilege.
The practical consequence: if a lawyer uses consumer ChatGPT to develop legal strategy, draft client communications, or analyze facts, those AI conversations may be discoverable. The privilege protections attorneys rely on to shield their work do not extend to AI-assisted work when the AI tool is a non-confidential third party.
Attorneys who want to use AI tools have three practical options:
ChatGPT Enterprise and the OpenAI API with zero-data-retention enabled contractually prohibit training on your conversations or retaining data beyond the session. These options exist but require IT involvement, firm-wide licensing decisions, and significant cost. Data still traverses OpenAI's infrastructure.
The attorney manually replaces client names, case numbers, and identifying details before submitting a prompt. Free but time-consuming, error-prone, and adds friction to every interaction. Studies of manual de-identification in healthcare settings find error rates of 10–15% even by trained professionals.
A purpose-built tool intercepts prompts in the browser, automatically detects and replaces sensitive data with placeholders before the prompt is submitted, and then restores the original values in the AI's response. Values that are replaced never reach the AI server. Because detection can miss details, the attorney still reviews the prompt before sending; anything not replaced is sent as written.
naosuu is a Chrome extension built for this problem. When an attorney types a prompt into ChatGPT, Claude, or Gemini, naosuu intercepts it before submission. On-device detection rules (running locally, no server calls) look for names, organizations, case numbers, financial figures, SSNs, and dates. A pre-flight review panel shows the attorney exactly which values will be redacted. The attorney reviews and approves, and the prompt is sent with those values replaced. When the AI responds, naosuu replaces every placeholder with the original real value.
This satisfies Rule 1.6(c)'s "reasonable efforts" standard in the most direct way available: the AI platform receives no confidential information to disclose.
naosuu costs $9.99/month with a 30-day free trial. It requires no IT configuration and no data processing agreement.
naosuu redacts before the prompt is sent, on your device. Review each prompt before you send it. $9.99/mo — 30-day free trial.
Try naosuu Free