Federal judges and major law firms are sounding a very clear alarm: anything you tell a consumer AI chatbot can be dragged into court. The practical rule lawyers are pushing on clients is blunt and not especially high tech: if it matters to your case, do not feed it into a public chatbot unless counsel is directing and supervising how it is used.
The wake up call came after a Manhattan federal judge ordered a defendant to hand over 31 documents created with Anthropic’s Claude in a criminal case and followed up with a written opinion on February 17, 2026. That ruling, and orders that followed, sent law firms scrambling to revise engagement letters and circulate fresh AI warnings to clients, according to Reuters.
What The Judge Actually Said
In United States v. Heppner, U.S. District Judge Jed S. Rakoff held that the AI generated materials were not protected by privilege. As summarized by Debevoise & Plimpton, the court found that the Claude exchanges failed basic privilege requirements: they were not communications with a lawyer, the user could not reasonably expect confidentiality in light of the platform’s terms, and the work was not performed at the direction of counsel.
The opinion underscored that the fine print matters. Platform privacy, retention and model training clauses can be outcome determinative when a judge is deciding whether AI related material is confidential or fair game in discovery.
Federal Courts Are Split
Not every court is taking Rakoff’s hard line, especially outside the criminal context. In Warner v. Gilbarco, Magistrate Judge Anthony P. Patti of the Eastern District of Michigan treated a pro se litigant’s ChatGPT assisted filings as protected work product. He viewed the chatbot as a drafting tool instead of a third party that automatically destroys protection just by being in the loop, according to coverage at JD Supra.
That split result, with a tough criminal ruling in Manhattan and more forgiving language in some civil orders elsewhere, leaves lawyers juggling inconsistent risk profiles depending on jurisdiction, case type and how clients are using AI.
Firms Are Rewriting Client Playbooks
In response, more than a dozen large firms have pushed out internal guidance and revised client contracts to spell out what AI use is acceptable. The message is less "embrace the future" and more "do not blow privilege." According to Reuters, firms are warning that dropping privileged advice or sensitive facts into a public chatbot can waive protection.
Common recommendations include: avoid pasting privileged narratives into consumer tools, document when AI work is done at a lawyer’s direction, and favor closed or enterprise AI systems with contractual confidentiality protections for anything substantive.
Practical Steps Lawyers Recommend
Litigation teams are being urged to build a clearly defined, legal grade lane for AI use. That typically means securing enterprise AI licenses that promise not to train on client data, updating ESI maps and legal hold protocols so prompts and outputs are treated like other documents, and preserving negative testing logs that might later be requested in discovery. Guidance from Gibson Dunn notes that litigators should assume prompts are documents the other side will eventually try to see.
Lawyers are also telling teams to memorialize legal direction inside the prompts themselves. A simple clarification like “I am conducting this research at the direction of counsel for [X] litigation” can help later if a judge needs to decide whether the work was created as part of a legal strategy.
On the law, the near term implication is straightforward even if the technology is not. Voluntarily sending attorney generated or attorney informed material to a third party consumer platform can amount to a waiver of privilege when the platform’s terms allow retention or disclosure to others. Analysis by Debevoise & Plimpton suggests courts are likely to focus on two variables in future disputes: whether a lawyer actually directed the AI work and whether a contract guarantees meaningful privacy.
For now, in house counsel and outside firms are tightening vendor agreements and internal rules with a conservative baseline: treat prompts and outputs as discoverable unless you are inside a documented, counsel directed enterprise workflow that preserves confidentiality. The practical takeaway, echoed in litigation playbooks from Gibson Dunn, is simple enough to print on a conference room poster. Treat AI artifacts like any other piece of potential evidence and govern them with the same discipline.
