Here’s a question I wouldn’t want to answer in a deposition.
“Counsel, when you prepared the summary judgment brief in this matter, did you use an AI system? What did you ask it? What documents did you give it? What did it tell you before you wrote what you wrote? Do those records still exist?”
At most firms using these tools, the answer to every one of those questions is sitting on a disk somewhere. Hardly anybody has decided what that answer should be, where the records live, how long they last, or what happens when a litigation hold lands.
This piece is about that trail. It’s got more questions than answers, because the law here isn’t settled. Pretending it is would be exactly the kind of confident wrongness this section keeps warning you about.
What actually gets created
Start by making a list. Most lawyers underestimate the trail by an order of magnitude.
One AI-assisted drafting session leaves behind:
- The prompt. What you asked, in your words, including how you framed it — and the framing usually carries your read on the case.
- The context you supplied. Documents, excerpts, facts, and often your take on them.
- Retrieved material. If a retrieval system fed the model documents, there’s a record of which ones it picked and why.
- The output, including the drafts you threw out.
- Your back-and-forth. “That’s too aggressive.” “Take out the part about the estoppel argument.” “Assume we lose the motion to dismiss.”
- The transcript, timestamped, which shows when you worked and on what.
- Tool call logs, if you’ve built the kind of connector I’ve described elsewhere in this section.
Now read that list the way the other side would. The back-and-forth is a near-perfect record of how counsel’s view of the case changed over time: which arguments you tried and dropped, which weak spots you flagged, what you thought the exposure was before you cleaned it up into a filing.
Depending on questions nobody’s answered for sure, that’s either the most protected material in the whole lawsuit or one heck of a discovery problem.
Where the doctrine starts
In Minnesota, work product lives at Rule 26.02(d), “Trial Preparation: Materials.” Documents prepared in anticipation of litigation are discoverable
“only upon a showing that the party seeking discovery has substantial need of the materials in the preparation of the party’s case and that the party is unable without undue hardship to obtain the substantial equivalent of the materials by other means.”
And here’s the part that matters most. The rule tells courts they “shall protect against disclosure of the mental impressions, conclusions, opinions, or legal theories of an attorney or other representative of a party concerning the litigation.” That extra protection for opinion work product is where most of an AI transcript belongs.
My read, and I’m fairly sure of it but not certain: a lawyer’s prompts and back-and-forth on a matter are textbook opinion work product. They’re the lawyer’s own words, written in anticipation of litigation, laying out mental impressions and legal theories in so many words. A prompt that says “draft an argument that the notice was untimely, but be careful because the estoppel certificate cuts against us” is about as pure a statement of counsel’s mental impressions as you’ll ever find.
The doctrine was built for a lawyer’s notes, and these are a lawyer’s notes. Typing them into a chat box instead of onto a legal pad shouldn’t change a thing, any more than dictation did.
At least one federal court agrees. In 2024 a district judge in California held that the ChatGPT prompts the authors ran before suing OpenAI were “queries crafted by counsel” that “contain counsel’s mental impressions and opinions about how to interrogate ChatGPT”. She refused to make them hand over the tests that didn’t pan out. The ones they’d put in the complaint, they had to produce. That’s one trial-court discovery order under the federal rule, not an appellate decision and not Minnesota law.
That’s the easy part. The rest isn’t.
Five questions nobody’s settled
1. Does the model’s output get the same protection?
The prompt is your words. Is the output your mental impressions, or is it something a machine spit out that just passed through your hands?
The obvious comparison is a draft an associate or a paralegal writes at your direction. That’s protected. Where the comparison breaks down: the associate is a person inside the privileged circle, and the model is a service. If your judgment did the picking — you asked for it, you sized it up, you kept some and tossed the rest — the mental-impressions argument looks strong. If a workflow ran on its own overnight and no lawyer touched it until morning, it looks a lot weaker.
My guess is the answer turns on how involved a human was. So how you build the workflow may decide what protection its output gets. Not enough people think about that when they set one up.
2. Does using an outside service waive anything?
Sharing with agents who help counsel generally doesn’t waive privilege. That’s the same reasoning that protects what you share with experts, vendors, and litigation-support outfits. A model provider whose contract bars training on your inputs and limits retention looks a lot like a litigation-support vendor.
It looks a lot less like one if the terms let the provider use your inputs to improve its models, or if there’s no confidentiality promise at all, or if you used a consumer product whose terms you never read. The vendor contract may end up doing more of the work in the privilege fight than any doctrine does.
That’s one more reason, on top of the usual confidentiality ones, to know exactly what your provider’s terms say about inputs and retention.
3. Does a litigation hold reach it?
Almost certainly. I think this is the one that’ll catch firms most off guard.
A hold reaches relevant electronically stored information in your possession, custody, or control. An AI transcript of your work on a matter is ESI about that matter, plain and simple. If your system rotates transcripts out after 30 days and a hold attaches, you’ve got a spoliation problem created by a retention setting nobody actually picked. Somebody clicked through a default.
Flip it around and, across the whole profession, it’s worse: plenty of firms keep everything, forever, by default. Every session on every matter, timestamped, in a folder nobody’s ever gone through. That’s a discovery obligation quietly piling up on a disk.
4. Whose records are they?
If the connector runs on your machine, they’re yours. If it runs on a vendor’s servers, the vendor has them, and the vendor can be subpoenaed. If the assistant is a consumer product that keeps your history on its servers, your client’s confidences and your mental impressions are sitting in somebody else’s system on somebody else’s schedule.
How you set things up decides the answer. You settled this question the day you picked your tools, whether you knew it or not.
5. What about the retrieval log?
This one’s the sneakiest. A retrieval system records which documents came up for a query and which didn’t. In a discovery fight over whether a search was good enough, that log is evidence about how you work. It might help. It might not.
And it rubs up against something I’ve argued elsewhere in this section. I’ve said a system should log what it retrieved, so you can answer for what it did. That’s right for competence and for supervision. It also leaves a record somebody could discover. I don’t think that changes my advice. The alternative is not being able to explain your own process, and that’s worse. But keeping the log should be a choice you made, not an accident.
What to do while the law catches up
None of this is a reason to freeze up. It’s a reason to make five decisions on purpose.
1. Pick your retention period, in writing. Not the default. Pick a period, write down why, apply it the same way every time, and make sure your hold process can override it. A policy you can defend, applied consistently, is worth far more than any particular number of days.
2. Keep each matter in its own sessions. One sprawling conversation that covers six clients is a discovery nightmare. You can’t produce what’s about one matter without exposing the other five, and you can’t cleanly claim protection for any of it. Work you keep separate by matter is work you can separate in discovery. That’s one more argument for the board-and-cards setup, and I didn’t think of it when I built it.
3. Read your provider’s terms on inputs, retention, and training. Then write down what they say and when you checked. When the terms change — and they will — you’ll want to know what you were relying on and when.
4. Assume it’s all discoverable and write that way. It’s the oldest advice in litigation, and it applies here as is. Don’t type anything into a prompt you wouldn’t put in an email about the case. The chat window feels casual and here-and-gone, which is exactly why people write things there they’d never put in a memo. It isn’t here-and-gone. It’s a file.
5. Find out what you’ve already got. Most firms a year into these tools are sitting on a pile of saved sessions nobody’s looked at. Find out where it is, how much there is, and what’s in it — before somebody asks you in a deposition.
My prediction
I’ve said elsewhere in this section that I expect real appellate authority on this within a few years, and that firms will be badly surprised by how much they kept without ever deciding to. I’ll add one more here.
The first real fight won’t be about privilege. It’ll be about spoliation. The privilege questions are interesting and will take years to sort out. The spoliation question is simple and mechanical. It shows up the first time a party asks for AI transcripts in a case where the other side’s retention setting quietly deleted them.
That’s no doctrinal puzzle. It’s a records-management screwup, and judges have a well-worn and unsympathetic vocabulary for those.
So the most valuable thing you can do about all this is also the least interesting: know where the records are and how long they last. Everything else can wait for the case law. That can’t.
Sources
- Minn. R. Civ. P. 26.02(d) (Trial Preparation: Materials)
- Minn. R. Prof. Conduct 1.6
- ABA Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512 (July 29, 2024)
- Tremblay v. OpenAI, Inc., No. 23-cv-03223-AMO, 2024 WL 3748003 (N.D. Cal. Aug. 8, 2024)
- Warner v. Gilbarco, Inc., No. 2:24-cv-12333, ECF No. 94 (E.D. Mich. Feb. 10, 2026)
- United States v. Heppner, No. 25 Cr. 503 (JSR), ECF No. 27 (S.D.N.Y. Feb. 17, 2026)
- Morgan v. V2X, Inc., No. 1:25-cv-01991-SKC-MDB, ECF No. 65 (D. Colo. Mar. 30, 2026)
Rules quoted were verified against the Revisor of Statutes. This is commentary on unsettled questions and is expressly not legal advice; the views above are the author’s reading and there is limited authority on most of them. Consult counsel in your own jurisdiction before relying on any of it.