Be careful: chatting with AI about your case is discoverable

rogerallen

In United States v. Heppner, Judge Rakoff of the Southern District of New York ruled that written exchanges between a criminal defendant and generative AI platform Claude were not protected by attorney-client privilege or the work product doctrine.

pseingatl

That's one judge. An audio tape made by a criminal defendant is intended for review by his counsel is a non-discoverable privileged communication. The tape retains this character if reviewed by an attorney-authorized paralegal. What difference exists where the attorney has the tape summarized by AI. I respectfully submit that Hizzoner is incorrect. We might also ask if the best venue to decide national AI regulation is a single judge sitting in a criminal case involving a fraudster. If Judge Rakoff is correct, then a trade secret shared with AI is no longer a trade secret. This affects not just a single NY criminal defendant, but anyone that runs a company and wants to keep business practices secret. I would submit that this is no way to regulate a field such as AI.

Terr_

It seems the key here isn't—or shouldn't be—what kind of service the defendant used, but whether something special happens when a service is involved in preparing a message to his lawyer. IMO if the "for my lawyer" purpose/intent is not in dispute, then it shouldn't matter whether the service is a search-engine, an LLM, a browser-based word processor, or the drafts/sent folders of a webmail client. The reverse direction is much clearer: Imagine a client receives an obviously-privileged email from their lawyer, and uses a cloud text-to-speech service to listen to it. Should that audio/text be admissible as evidence? Hell no.

anon373839

This is a really interesting and well written case update/critique. I agree with the author's that the judge's reliance on Anthropic's fine-print privacy policy does not satisfy the actual legal standard governing privilege. Or if it did, it would raise extremely thorny issues around all of the cloud-based technology products that lawyers and clients use every day. That said, I note that the court's opinion specifically calls out Anthropic's practice of *training models on user data* as a reason why the defendant could not have expected confidentiality. I do not use these cloud models for anything important precisely because they are operated by companies, like Anthropic, that are completely untrustworthy.