The ABA’s Formal Opinion 512 warns that lawyers using generative AI must consider competence, confidentiality, communication with clients, supervision, candor, and fees. The California State Bar’s 2026 guidance goes further, emphasizing that lawyers must understand AI risks, protect client confidences, supervise AI use, verify AI outputs, and consider client communication where AI materially affects the representation.
The problem is not AI. The problem is uncontrolled disclosure.
Lawyers already know the old rule: privileged information generally must remain confidential. The problem is that many clients do not think of an AI platform as a third party. They experience the tool like a private conversation, a search engine, or a smarter version of Google.
Courts are beginning to test that assumption.
In United States v. Heppner, commentary on the decision describes the court as ruling that written exchanges between a criminal defendant and Claude were not protected by attorney-client privilege or work product. The analysis focused heavily on the fact that the AI interaction was not with counsel and was not done at counsel’s direction.
That does not mean every AI-assisted document is automatically discoverable. But it does mean lawyers should stop assuming that “AI use” will be treated like a private legal notebook.
The client may have created evidence.
In Tremblay v. OpenAI, a court compelled production of OpenAI account information, prompts, outputs, and testing-process documents related to plaintiffs’ pre-suit ChatGPT testing. The court rejected arguments that prompts and outputs not supporting the plaintiffs’ claims were irrelevant or protected work product.
That case is not a typical commercial dispute or personal injury case. But the lesson travels: when a party uses AI to test facts, claims, allegations, examples, damages, or theories, the prompt history may become part of the factual record.
Think about how this plays out in ordinary litigation:
A terminated employee asks AI to draft a narrative of why the firing was discriminatory.
A business owner asks AI to summarize a contract dispute and “identify the weakest facts.”
A medical-malpractice plaintiff uploads medical records and asks AI whether the doctor breached the standard of care.
A corporate executive pastes a whistleblower complaint into AI and asks whether the company should settle.
Each one may feel harmless. Each one may create a record the lawyer later has to preserve, analyze, and possibly fight over.
The intake form needs a new question.
Litigation intake used to ask about documents, emails, text messages, witnesses, insurance, social media posts, and prior lawyers.
Now it should ask about AI.
A simple version:
Have you used any AI tool, chatbot, legal app, writing assistant, search assistant, summarizer, or automated drafting tool to discuss, summarize, analyze, research, translate, or draft anything related to this dispute?
Follow-up questions should include:
| Intake Question | Why It Matters |
|---|---|
| Which AI tool did you use? | Different platforms have different retention, training, and privacy terms. |
| Was it a personal, free, paid, business, or enterprise account? | “Paid” does not automatically mean confidential. |
| What did you enter, upload, or paste? | Prompts and uploads may contain privileged or confidential information. |
| What did the AI produce? | Outputs may become relevant to claims, defenses, damages, or impeachment. |
| Did you save, copy, forward, or rely on the output? | Reliance may affect discoverability and preservation. |
| Did anyone else see the prompt or output? | Sharing may affect privilege and confidentiality. |
Clients need a plain-English warning.
Law firms should not wait until the first discovery request to educate clients. The warning should appear in the engagement letter, intake packet, and first litigation hold notice.
Suggested language:
Do not enter, upload, paste, summarize, or discuss anything about your legal matter in ChatGPT, Claude, Gemini, Copilot, Perplexity, or any other AI tool unless our firm has approved the tool and the specific use. This includes letters, contracts, emails, texts, pleadings, medical records, employment records, financial records, discovery materials, settlement discussions, legal strategy, or your personal notes about the case. AI prompts, outputs, uploads, account data, and logs may be discoverable and may create privilege, confidentiality, privacy, or protective-order issues.
The takeaway
AI is now part of the factual environment of litigation. Clients may use it before counsel is retained. Employees may use it before a lawsuit is filed. Witnesses may use it before a statement is taken.
The safest firms will treat AI the same way they treat social media, email, texts, and cloud storage: as something that must be addressed early, preserved correctly, and controlled before it causes damage.
The new litigation intake question is simple:
“Have you used AI for anything related to this matter?”
That question may protect privilege, preserve evidence, and prevent a discovery problem before it starts.