What Lawyers Paste into ChatGPT: Privilege, Confidentiality, and AI

Why lawyers are heavy AI users, quietly

Legal work is text work: reading, summarising, comparing, drafting. LLMs do all four startlingly well, and associates under billable-hour pressure noticed early. Surveys throughout 2024–2025 consistently showed legal among the highest-adoption professions for generative AI, and among the least likely to have usage formally sanctioned. The result is classic shadow AI: heavy, valuable, invisible use.

What does that use look like in practice? These are the prompt patterns that appear over and over:

• “Summarise this agreement”, followed by an entire executed contract: party names, pricing, indemnities, change-of-control terms.
• “Draft a settlement counter-offer based on this”, pasting the existing offer with client and opposing-party identities and amounts.
• “Rewrite this email to the client”, including the client's name, the matter, and the advice itself.
• “Find the inconsistencies in these two depositions”, pasting discovery material, sometimes under protective order.
• “Check this due-diligence list”, including target-company financials and employee data in an M&A context.

The privilege problem

Attorney-client privilege survives only while confidentiality is maintained. Voluntary disclosure to a third party outside the privilege relationship is the classic route to waiver. A consumer-tier AI chatbot (whose terms of service permit input retention, human review, and model training) is a third party with none of the safeguards of, say, a vendor under a confidentiality agreement.

Has a court squarely held that a ChatGPT paste waived privilege? The case law is still developing. But the asymmetry is brutal: the upside of the paste was twenty saved minutes; the downside is arguing novel waiver questions in front of a judge while opposing counsel subpoenas AI usage records. No litigation department wants to be the test case.

The professional-conduct problem

The ABA addressed generative AI directly in Formal Opinion 512 (July 2024). Its core holdings map onto exactly the behaviour above:

Lawyers must evaluate their duty of confidentiality under Model Rule 1.6 before inputting information relating to the representation into a generative AI tool, and where the tool is self-learning, informed client consent may be required.

Competence (Rule 1.1) requires understanding how a tool handles inputs; supervision (Rules 5.1 and 5.3) extends those duties to the conduct of associates and staff. A firm that cannot say which AI tools its people use cannot honestly claim to be supervising that use. And for EU client data, GDPR runs in parallel: an unauthorised disclosure of client personal data is a reportable breach analysis on top of the ethics issue.

Why policy memos haven't worked

Most firms responded with a memo: “do not put client confidential information into AI tools.” The memo fails for the same reason it fails everywhere: it asks a busy professional to weigh an abstract future risk against a concrete present deadline, with no friction at the moment of action. The associate pasting a contract at 11pm is not consulting the memo.

Effective control has to live where the paste happens: the browser.

A workable AI programme for a firm

1. Sanction real tools. Provide enterprise AI (with no-training terms and DPAs) good enough that lawyers do not need workarounds. Legal-specific assistants with confidentiality commitments count.
2. Define matter-data rules. Names, identifying facts, financial terms, and privileged analysis stay out of unsanctioned tools, stated in one page, not thirty.
3. Block at the point of paste. Browser-level detection that recognises client PII, financial data, and credentials and cancels the paste into unsanctioned AI tools, explaining why in the moment. The associate learns; the client data never leaves.
4. Keep the supervision log. A timestamped record of detections and blocks is what Rules 5.1/5.3 supervision looks like in evidence, and it answers clients (and increasingly, cyber insurers and outside-counsel guidelines) who now ask firms to attest to their AI controls.

The client question is coming either way

Corporate clients have started adding AI-usage clauses to outside-counsel guidelines: disclose which tools touch our matters, or certify that none do. Firms that can answer with an audit log turn a risk conversation into a differentiator. Firms that cannot are one associate's paste away from a very difficult phone call.