Are AI Prompts Discoverable?

A Michigan Litigator’s Guide to Gen AI Records Retention

The better question for Michigan litigators in 2026 is not whether generative AI is “special,” but whether the data it creates fits comfortably within the rules that already govern electronically stored information. In most cases, it does. A prompt is text. An output is text, image, code, or another digital artifact. An activity log is metadata about who used a system, when, and how. Once those items are relevant to a claim or defense, within a party’s possession, custody, or control, and not shielded by privilege or another protection, they begin to look very much like ordinary discoverable ESI. Recently 2026 commentary is moving in that direction, and it is doing so by applying familiar discovery principles rather than inventing an entirely new body of AI law. ⁴ ²

Please note this blog post should be used for learning and illustrative purposes. It is not a substitute for consultation with an attorney with expertise in this area. If you have questions about a specific legal issue, we always recommend that you consult an attorney to discuss the particulars of your case.

That shift matters because lawyers and clients have now embedded generative AI into everyday business conduct. Employees use AI systems to summarize contracts, draft emails, analyze datasets, prepare talking points, outline investigations, generate code, compare policy language, and brainstorm litigation themes. Once that happens, the AI interaction is no longer a futuristic novelty. It becomes part of the information ecosystem from which facts are created, refined, communicated, and sometimes memorialized. If a dispute later turns on who knew what, when they knew it, what instructions they gave, how a document was created, or whether a party acted reasonably, AI prompts and outputs may become as important as emails, text messages, spreadsheets, or tracked changes in a Word file. ⁴

Michigan practitioners should resist two equal and opposite mistakes. The first is the casual view that AI chats are too novel, too informal, or too ephemeral to matter in discovery. The second is the alarmist view that every AI interaction must now be preserved forever. Neither approach is defensible. The developing authorities instead point toward a more disciplined middle position: treat AI-generated material like any other ESI category. Preserve what is relevant or reasonably likely to become relevant, negotiate scope early, use proportionality seriously, and build governance that can be explained to a court without embarrassment. ¹ ² ⁴

Michigan’s civil discovery structure already gives litigators the framework they need. The Michigan bench book’s discussion of MCR 2.302 explains that, in cases requiring initial disclosures, parties must provide a copy or description by category and location of all documents, ESI, and tangible things they may use to support their claims or defenses. The same Michigan materials also explain that ESI need not be produced if it is not reasonably accessible because of undue burden or cost, though a court may still order production on a showing of good cause and may impose conditions, including cost allocation and limits on scope. In other words, Michigan already has the doctrinal tools for AI disputes: breadth at the front end, proportionality as a limiting principle, and judicial control over difficult ESI sources. ¹

The production rules are equally important. Michigan’s materials on MCR 2.310 state that when ESI is requested and no form is specified, it must be produced in the form in which it is ordinarily maintained or in a reasonably usable form. That simple rule has major implications for AI evidence. If the relevant record is a native export from an enterprise AI platform with timestamps, user identifiers, thread structure, and linked outputs, a flat screenshot may not always be enough. Conversely, if the dispute only concerns the substance of a particular output, a reasonably usable export may be all that is necessary. The point is not that AI data always requires exotic collection. The point is that the ordinary production analysis still applies, and counsel should think about format early rather than after motion practice begins. ¹

Federal practice tells the same story, which matters because many Michigan litigators split time between state and federal forums, and even state-court arguments often borrow federal proportionality language. Rule 26 requires disclosure of documents, ESI, and tangible things a party may use to support its claims or defenses, and Rule 34 expressly authorizes the production of designated ESI stored in any medium from which information can be obtained in a reasonably usable form. Rule 34 also allows the requesting party to specify the form of production. Nothing in those rules excludes AI artifacts because they originate in a chatbot interface rather than a conventional software application. ²

That is why the most useful framing for clients is not “Are prompts discoverable?” in the abstract. The better framing is “When do prompts, outputs, and logs become relevant evidence?” A prompt may matter because it shows the factual assumptions a user fed into the model. An output may matter because it became the first draft of a final business communication or legal analysis. A log may matter because it shows when a party began using an AI tool, what subject was being explored, whether multiple iterations were run, or whether a key document was created through AI-assisted drafting rather than independent human work. In a fraud case, that might bear on scienter or intent. In a trade secret case, it might bear on what confidential information was entered into the system. In a contract dispute, it might bear on interpretation, negotiation history, or internal understanding. The discoverability question, then, is almost always really a relevance-and-proportional question. ² ⁴

Recent commentary has sharpened that point with concrete examples. A February 2026 K&L Gates analysis identified In re OpenAI as the defining early ruling on GenAI data discoverability and explained that the court compelled production of millions of GenAI logs, including user prompts and model responses, subject to anonymization. Just as important, that commentary emphasized the companion principle many litigators prefer to forget: AI data is not automatically discoverable merely because it exists. In the same litigation, a separate request for internal AI-tool content was denied as irrelevant and disproportionate. That pairing is the real lesson. Courts are not carving out a categorical exemption for GenAI data, but they are also not authorizing unlimited fishing expeditions into every AI interaction an organization has ever generated. ⁴

For Michigan litigators, that means records retention and litigation holds must become more precise, not more panicked. A client should not respond to this emerging issue by saving every prompt forever, any more than a sophisticated organization would keep every temporary browser artifact or every auto-saved document fragment indefinitely. Over-retention drives cost, inflates review populations, and creates new risks. But under-retention is just as dangerous when the organization has failed to identify where relevant AI data lives, how long the platform keeps it, whether users can delete it, whether exports preserve context, and whether logs are retained separately by a vendor. The goal is not maximal retention. It is defensible retention. ⁴ ¹

That is where prompt governance becomes a business-litigation issue rather than a purely technical one. Prompt governance, properly understood, is the old discipline of information governance translated into an AI environment. It asks who may use which tools, for what purposes, using what categories of information, under what retention settings, with what access controls, and with what procedures for hold implementation when a dispute arises. Once a client uses generative AI in contract administration, HR workflows, product design, incident response, marketing review, or legal operations, that client has created a new source of ESI. The responsible answer is to inventory that source, map its settings, and decide in advance how litigation preservation will work. ³ ⁴

A sensible Michigan hold notice in 2026 therefore should say more than “preserve emails and texts.” If the matter potentially touches AI-assisted work, the hold should prompt custodians to identify enterprise AI platforms, browser-based consumer tools, plug-ins embedded in ordinary office software, and any saved exports or linked files. It should instruct them not to delete relevant prompt threads or outputs, not to “clean up” drafts in a way that strips context, and not to move only cherry-picked outputs into a folder while allowing the surrounding conversation to disappear. K&L Gates specifically notes that preservation steps may include disabling auto-delete settings, exporting chat histories, saving key exchanges in document repositories, and coordinating with IT regarding log and metadata retention. That is not exotic advice; it is simply the AI version of a modern legal hold. ⁴

The same principle should shape early case assessment. At the start of a matter, counsel should ask whether any custodian used AI to draft or analyze content tied to the dispute, whether the platform preserves prompts and outputs, whether relevant activity logs exist separately, and whether the tool was personal, enterprise, or embedded in another product. That inquiry belongs besides the usual questions about email, shared drives, phones, collaboration tools, and backup systems. If it is omitted, counsel risks learning too late that a central document was AI-generated, that the underlying prompt chain was auto-deleted, or that a vendor not the client controls the relevant logs. ⁴

None of this means every AI artifact is worth the cost of collecting. Michigan law already accommodates that reality. The Michigan bench book explains that a party needs not provide ESI that is not reasonably accessible because of undue burden or cost, and the court may limit the frequency or extent of ESI discovery even when the information is accessible. That language should encourage practical, targeted decision-making. If a case turns on a single generated report, preserving the final output and a narrow set of surrounding prompts may be enough. If a case turns on how an AI system was repeatedly used across a large business process, logs and metadata may matter more. The key is being able to explain why a preservation line was drawn where it was. ¹

Privilege adds another layer, and here Michigan litigators should be especially careful not to assume too much. The State Bar of Michigan’s 2025 report emphasizes both technological competence and the duty of confidentiality. It stresses that lawyers must understand the benefits and risks of AI tools, remain responsible for what those tools produce, and protect client information when using externally hosted systems. That is already enough to require caution with prompts that contain strategy, sensitive facts, privileged communications, or trade secrets. ³

A February 2026 O’Melveny analysis of United States v. Heppner shows why. There, Judge Rakoff held that written exchanges a criminal defendant had with the consumer version of Claude were not protected by attorney-client privilege or the work product doctrine. According to the summary, the court reasoned that the communications were not confidential and were not made at counsel’s direction. The article also notes that the court left open the possibility that attorney-directed use under a different confidentiality structure could present a different analysis. The practical takeaway is not that AI use automatically destroys privilege. It is that privilege analysis remains technology-neutral and intensely fact specific, with confidentiality and attorney direction still doing the heavy doctrinal work. ⁵

That distinction is vital for business clients and litigators alike. A public or consumer tool with training-on inputs, broad disclosure language, or unclear enterprise controls raises a very different risk profile from a vetted enterprise environment with contractual confidentiality protections, segregated data handling, and documented legal workflows. Michigan lawyers advising clients on AI use should therefore stop treating vendor terms as procurement trivia. In an AI discovery fight, the platform’s privacy terms, retention settings, and administrative controls may become central to both preservation and privilege arguments. ³ ⁵

This is also why lawyers should be cautious about conflating internal business use with legal use. If an employee asks a public AI system to “summarize our likely defenses” or “draft talking points for the dispute,” the result may feel work-product adjacent, but feeling adjacent is not the same as being protected. The safer course is to separate legal workflows from ad hoc consumer-tool use, direct legal experimentation through approved systems, document the purpose and supervision of the work, and avoid feeding sensitive matter-specific content into tools whose confidentiality terms are too weak to defend later. That is not paranoia. It is simply applying ordinary privilege hygiene to a new interface. ³ ⁵

For litigators on the requesting side, AI discovery should also be handled with discipline. Discovery requests that ask for “all prompts, outputs, logs, and communications related to any AI system” are almost inviting a proportional objection. Better requests tie the AI material to a disputed issue, date range, custodian, workflow, or specific final document. If the theory is that a party used AI to generate misleading advertising, then requests should target the prompts and outputs used to create that campaign. If the theory is that a party used AI to analyze copyrighted works, then logs and outputs tied to that analysis may be central. Precision not only improves the chance of obtaining useful material; it also makes the request sound like a serious discovery tool rather than a press release. ² ⁴

The same goes for meet-and-confer practice. AI discovery should now be on the agenda whenever the facts suggest it may matter. Counsel should be prepared to discuss what tools were used, whether they were enterprise or consumer tools, what data is retained, how long it is kept, whether logs exist separately, what export options are available, and what level of burden would be involved in collection and review. K&L Gates emphasizes negotiating scope early and using protective orders or anonymization to manage confidentiality concerns. That advice fits comfortably with the Michigan and federal emphasis on proportionality, specificity, and practical ESI management. ¹ ² ⁴

In-house legal departments and outside counsel should also rethink retention schedules through a litigation lens. The right policy will vary by industry, system, and risk profile, but the policy should at minimum distinguish between incidental experimentation and business-substantive AI use. A routine, short-lived brainstorming exchange may warrant brief ordinary-course retention. By contrast, AI outputs used in decision-making, incorporated into final work product, tied to regulated activity, or relevant to customer, employee, or litigation events may justify longer retention or separate storage. The important thing is that the organization can articulate why the policy exists, apply it consistently, and suspend ordinary deletion when a preservation duty arises. ⁴ ¹

Training is part of that governance story. Michigan’s AI report stresses competence, continuous monitoring, and the need to understand technology well enough to use it responsibly. In the discovery setting, that means custodians should know that a prompt is not “just a question” if it embeds key facts, strategy, or confidential content. They should know whether their AI tool saves history by default, whether they may delete threads, whether exports preserve metadata, and when they must notify legal that AI-assisted materials exist. Lawyers, meanwhile, should know enough about the client’s tools to ask the right preservation questions before the first scheduling conference. ³

The broader business lesson is straightforward. AI records retention is no longer only a compliance or IT topic. It is a litigation-readiness topic. The companies that handle it best will not be the ones that ban all AI or preserve all AI. They will be the ones that know where their AI data resides, have approved tools and use policies, align retention with business value and legal risk, and can implement targeted holds without improvisation. Those are the same companies that usually perform better in ordinary ESI disputes, because AI has not changed the habits of good information governance; it has simply exposed where those habits were weak. ³ ⁴

So, are AI prompts discoverable? Often yes, but not because the law has suddenly become fascinated with prompts as a category unto themselves. They are discoverable when they function as relevant, nonprivileged, proportional ESI. Sometimes the prompt will matter more than the output because it shows assumptions, instructions, or intent. Sometimes the output will matter more because it became the operative business record. Sometimes logs will matter because they reveal timing, frequency, or scope of use. And sometimes none of it will be worth pursuing because the burden outweighs the probative value. That is not uncertainty so much as ordinary discovery analysis applied to new data. ¹ ² ⁴

For Michigan litigators, the actionable conclusion is clear. Do not wait for a published Michigan appellate decision that says “AI prompts are discoverable” in those exact words. The existing rules are already broad enough, the recent commentary is already pointed out enough, and the early decisions are already instructive enough to justify action now. Update holds notices. Inventory AI tools. Separate consumer use from approved enterprise use. Revisit confidentiality assumptions. And when a case arises, ask at the outset whether AI-generated materials played a meaningful role in the events at issue. The lawyers who do that will be better positioned to preserve the right material, resist the wrong requests, and explain their choices persuasively when the court asks the inevitable question: what did you do about the AI data? ¹ ³ ⁴ ⁵

Contact Tishkoff

Tishkoff PLC specializes in business law and litigation. For inquiries, contact us at www.tish.law/contact/. & check out Tishkoff PLC’s Website (www.Tish.Law/), eBooks (www.Tish.Law/e-books), Blogs (www.Tish.Law/blog) and References (www.Tish.Law/resources).

Sources:

1. Michigan Courts, Michigan Judicial Institute Civil Benchbook, Chapter 5: Discovery, including sections on MCR 2.302 and MCR 2.310. https://www.courts.michigan.gov/4aeeef/siteassets/publications/benchbooks/civil/civilresponsivehtml5.zip/Civil/Ch_5_Discovery/Disclosure.htm
2. Federal Rules of Civil Procedure, Rule 26 and Rule 34, Legal Information Institute / Cornell Law School.  https://www.law.cornell.edu/rules/frcp/rule_26  https://www.law.cornell.edu/rules/frcp/rule_34
3. State Bar of Michigan, Transforming the Legal Profession in the Age of AI, June 2025.  https://www.michbar.org/AI
4. Julie Anne Halter and Alexa Stemmler, K&L Gates, Litigation Minute: Is AI-Generated Content Discoverable? What Companies Need to Know in 2026, February 12, 2026.  https://www.klgates.com/Litigation-Minute-Is-AI-Generated-Content-Discoverable-What-Companies-Need-to-Know-in-2026-2-12-2026
5. O’Melveny & Myers LLP, S.D.N.Y. First-of-its-Kind Ruling: AI-Generated Documents Are Not Privileged, February 24, 2026. https://www.omm.com/insights/alerts-publications/sdny-first-of-its-kind-ruling-ai-generated-documents-are-not-privileged/

This publication is for general informational purposes and does not constitute legal advice. Reading it does not create an attorney-client relationship. You should consult counsel for advice on your specific circumstances.