AI Prompts in Discovery: What Florida Businesses and Litigants Should Do Now

Written By Patrick Dempsey

THE SHORT ANSWER

AI prompts, outputs, account settings, and related records can become relevant in litigation. That does not make every AI interaction discoverable. The ordinary rules still control: relevance, proportionality, possession or control, preservation, attorney-client privilege, work product, expert discovery, and any governing protective order.

The reported decisions are early and fact-specific. They do not establish a single rule that all AI material is protected or that all of it must be produced. They do show that the result can depend on choices made before a discovery request arrives: who used the tool, at whose direction, for what purpose, what information was entered, what the provider could do with it, whether the material was preserved, and whether anyone later relied on it in a filing or expert opinion.

As of September 11, 2026, our research identified no reported federal circuit or Florida appellate decision squarely deciding the privilege or work-product status of generative-AI prompts and outputs. Businesses should not mistake the absence of appellate authority for the absence of risk.

START WITH FOUR DIFFERENT QUESTIONS

Disputes over AI records often become confused because several legal questions are treated as one. They should be separated.

Is the material within the permissible scope of discovery? A stored prompt, output, chat history, export, or related record may be electronically stored information. A requesting party must still connect it to a claim or defense and satisfy the governing proportionality requirements. The responding party must also determine whether the material is within its possession, custody, or control. The fact that someone used AI does not automatically make every interaction relevant or producible.

Does attorney-client privilege apply? Privilege ordinarily protects confidential communications between lawyer and client for the purpose of requesting or providing legal advice. An AI provider is not the client's lawyer. A separate issue arises when a user puts privileged information into a third-party system. The provider's terms, retention practices, training practices, access rights, and contractual confidentiality obligations can affect whether confidentiality was reasonably preserved. Product labels such as “enterprise” or “closed” are useful starting points, not legal conclusions.

Is the material work product? In federal civil litigation, Rule 26(b)(3) can protect documents and tangible things prepared in anticipation of litigation by or for a party or its representative. Materials revealing counsel's mental impressions receive especially strong protection. Work-product waiver is not identical to attorney-client-privilege waiver; disclosure must ordinarily be assessed by asking whether it was made to an adversary or substantially increased the likelihood that an adversary would receive the material.

Did a testifying expert use the material? Expert discovery has its own rules. Draft reports and many attorney-expert communications receive protection, while facts or data considered by the expert and the basis and methodology of the opinion may be discoverable. An AI prompt used to filter evidence, select documents, or perform analysis can generate a dispute over which side of that line it occupies.

WHAT THE REPORTED DECISIONS ACTUALLY SHOW

A represented client using a public chatbot independently

In United States v. Heppner, No. 25 Cr. 503 (JSR), 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026), a represented criminal defendant independently used the consumer version of Claude to analyze the investigation and possible defenses. Counsel did not direct the searches. The court held that the resulting documents were protected by neither attorney-client privilege nor work product.

The court reasoned that Claude was not an attorney, the interactions were not confidential attorney-client communications, and giving the resulting documents to counsel later did not create privilege retroactively. The work-product claim also failed because the defendant acted independently rather than at counsel's direction and the materials did not reflect counsel's strategy. The court additionally considered the consumer platform's data-use and disclosure terms.

Heppner should not be read as deciding the status of lawyer-directed work performed through a system with materially different confidentiality protections. The opinion expressly arose from a public consumer tool and a client acting apart from counsel.

Pro se litigants preparing their own cases

Two federal civil decisions reached more protective results for self-represented litigants.

In Warner v. Gilbarco, Inc., No. 2:24-cv-12333-GAD-APP, 2026 WL 373043 (E.D. Mich. Feb. 10, 2026), the defendants sought broad discovery concerning a pro se plaintiff's use of third-party AI tools. The court denied the request as untimely and also found relevance and proportionality problems. Alternatively, it held that litigation-preparation materials could receive work-product protection and that use of ChatGPT did not automatically waive that protection because work-product waiver generally requires disclosure to an adversary or conduct likely to place the material in an adversary's hands.

In Morgan v. V2X, Inc., No. 1:25-cv-01991-SKC-MDB (D. Colo. Mar. 30, 2026), the court concluded that Rule 26(b)(3) can protect a pro se party's AI-assisted litigation preparation. But the result was not complete protection. The plaintiff had to identify the AI tool used with confidential discovery because he failed to show that the tool's identity revealed protected strategy. The court also amended the protective order to restrict the use of confidential material in AI systems unless specified contractual security, non-training, non-disclosure, and deletion protections were present.

These decisions do not establish that attorney supervision is unnecessary in a represented party's case. They concern the distinct position of a pro se litigant, who is both the party and the advocate. They also show that an underlying analysis may receive protection even when the tool's identity, security practices, or handling of confidential discovery remains discoverable.

AI work performed by a nonlawyer third party

In Shealy v. Seaside Investments, LLC, No. 2684CV00799-BLS2 (Mass. Super. Ct. June 16, 2026), a represented party sent dispute-related documents to his romantic partner, who put them into ChatGPT and returned generated drafts. Counsel did not direct that work. The Massachusetts trial court concluded that neither the partner's AI queries nor the resulting output was protected work product because the partner was not the party's representative within the meaning of the applicable rule and did not act at counsel's direction.

The important fact was not simply that ChatGPT was used. The person conducting the work was neither counsel nor a qualifying representative, while the party was represented by lawyers who had no role in directing the exercise.

Lawyer-created prompts used to investigate a claim

In Tremblay v. OpenAI, Inc., No. 23-cv-03223-AMO, 2024 WL 3748003 (N.D. Cal. Aug. 8, 2024), plaintiffs' counsel tested ChatGPT while investigating copyright claims. The plaintiffs relied on favorable prompt-output pairs in their complaint but withheld other testing that did not support their allegations.

The district court treated the undisclosed, attorney-crafted prompts as opinion work product because they reflected counsel's mental impressions about how to test the system. It declined to compel the withheld negative testing. Materials affirmatively used in the complaint were a different matter. The case supports a careful distinction between private attorney-directed investigation and material selected for affirmative public reliance. It does not support a rule that every disclosure of AI output waives all related work product.

AI prompts used in a testifying expert's document review

In Conservation Law Foundation, Inc. v. Shell Oil Co., No. 3:21-cv-00933 (VDO), ECF No. 970 (D. Conn. May 18, 2026), a magistrate judge ordered revised discovery responses concerning prompts or queries used by a testifying expert and her team to narrow a document production. The magistrate judge treated that process as part of the expert's methodology and concluded that the parties' agreement protecting expert notes, drafts, and communications did not clearly protect the prompts.

That is not the end of the procedural history. On June 3, 2026, the district court stayed the production order pending resolution of the plaintiff's Rule 72(a) objection. The order therefore remains a significant warning about expert workflows, but it should not be presented as settled or controlling law. It is a stayed magistrate-judge discovery order under review.

FLORIDA'S CURRENT RULES

Florida's state-court requirements changed during 2026. Any discussion that stops with the early circuit administrative orders is now incomplete.

Florida state courts

In May 2026, the Florida Supreme Court amended Florida Rule of General Practice and Judicial Administration 2.515(d)(2). Effective June 15, 2026, a signer of a document filed in a Florida court represents, among other things, that the legal authorities identified in the filing exist and are accurately cited. The rule authorizes sanctions after notice and an opportunity to be heard, including striking the document, costs, attorneys' fees, contempt, or dismissal.

The statewide rule does not require disclosure that generative AI was used. In Administrative Order AOSC26-12, the Florida Supreme Court explained that Rule 2.515 replaced the varied circuit AI-disclosure and certification requirements. Effective June 15, state courts may not impose those separate requirements through local administrative orders, court policies, or judicial practices.

The earlier Miami-Dade and Broward circuit orders are therefore important history, but they do not describe the present statewide filing requirement.

Individual federal judges

Florida's federal courts are separate systems. Some individual federal judges have adopted their own AI-related filing requirements. For example, Judge Wendy Berger's April 2, 2026 standing order requires parties appearing before her to certify whether generative AI was used in preparing a filing and, if it was, that a human personally reviewed the language for accuracy and verified the citations.

That is an individual judge's order, not a district-wide Middle District of Florida rule. Counsel should check the assigned judge's current orders and practices in every case rather than assume one judge's requirement applies throughout the district.

Florida lawyers' ethical duties

Florida Bar Ethics Opinion 24-1 is an advisory, nonbinding opinion, but it provides important guidance. A Florida lawyer using generative AI must protect client confidentiality, understand relevant data-retention and data-sharing practices, supervise the work, verify its accuracy, charge reasonable fees, and comply with advertising rules.

The opinion recommends obtaining informed client consent before using a third-party generative-AI system when the use would disclose confidential information. The precise duty depends on Rule 4-1.6, any applicable exception, the nature of the system, and the information involved. An in-house or otherwise isolated system may mitigate some confidentiality concerns, but the lawyer remains responsible for understanding how the system handles client information.

A PRACTICAL CONTROL PLAN

The emerging cases do not justify preserving or producing every AI interaction in every dispute. They do justify adding AI to the existing process for identifying, preserving, protecting, and reviewing potentially relevant information.

1. Map actual use before litigation

Identify which employees, lawyers, consultants, and experts use AI; which products and account types they use; what kinds of information they enter; whether chat histories or logs are retained; and whether administrators or providers can retrieve them. Do not assume that a product's marketing label answers questions about training, human review, retention, deletion, or disclosure.

2. Put legal work under a defined workflow

When AI will be used for litigation strategy or legal analysis, counsel should define the purpose, authorized users, approved systems, permitted information, and review requirements. Direction by counsel can strengthen a work-product argument, but it does not guarantee protection. The content, purpose, confidentiality controls, and later use still matter.

3. Preserve relevant material proportionately

When litigation is reasonably anticipated, determine whether relevant AI prompts, outputs, logs, exports, or settings exist and fall within the organization's possession, custody, or control. A legal hold should address AI expressly when the facts make AI activity a likely source of relevant evidence. The preservation instruction should be tailored to the claims and custodians rather than collecting unrelated AI histories indiscriminately.

4. Address confidential discovery in protective orders

Protective orders and ESI protocols should state whether confidential discovery may be put into an AI system and, if so, under what safeguards. Relevant terms may include isolation of customer data, no training on submitted material, limits on provider and subcontractor access, encryption, deletion rights, breach notice, audit documentation, and restrictions on onward disclosure.

5. Establish expert rules before substantive work begins

An expert engagement should address whether AI may be used, what may be entered, what must be preserved, how outputs will be validated, and how the workflow will be described if challenged. Counsel should consider the expert-discovery rules and the pending posture of Conservation Law Foundation rather than promise that all prompts will be protected or assume that all must be produced.

6. Make discovery requests specific

A request for every AI interaction may be irrelevant, disproportionate, and vulnerable to a work-product objection. When AI use is genuinely connected to a claim, defense, filing, investigation, or expert opinion, requests should identify the relevant custodians, subjects, time period, systems, and categories of records. The same discipline should govern objections and privilege logs.

7. Treat affirmative reliance as a separate decision

Before quoting or relying on AI output in a pleading, report, declaration, or presentation, consider what related prompts, settings, and testing may become discoverable. Selective public reliance can create disclosure and fairness arguments that would not exist if the material remained part of a private litigation-preparation process.

OUR TAKE

AI does not require courts to abandon ordinary discovery doctrine. It creates new records, new custodians, and new third-party systems to which familiar doctrine must be applied. The reported cases disagree in part because their facts are different: a represented client acting independently, a pro se litigant preparing a civil case, a romantic partner generating drafts, a lawyer testing a claim, and an expert filtering evidence are not equivalent situations.

The sound response is not to assume that all AI use is discoverable or that an enterprise account makes it privileged. It is to create a defensible record of purpose, direction, confidentiality, preservation, and human review. Organizations that can explain those choices will be better positioned to protect legitimate work product, comply with discovery obligations, and challenge requests that go too far.

WHEN TO CALL US

Counsel should be involved before confidential business information, client information, or protected discovery is introduced into a new AI workflow. Legal review is also appropriate when a litigation hold may need to reach AI systems, an expert proposes to use AI, a protective order is silent about AI, or an opposing party serves AI-specific document requests or deposition topics.

Hirzel Dreyfuss & Dempsey represents clients in commercial litigation and discovery disputes in Florida state and federal courts. More information is available on our commercial litigation page.

SOURCES

DISCLAIMER

This post is provided for general informational purposes only and does not constitute legal advice or create an attorney-client relationship. The decisions discussed above are trial-level rulings and do not bind Florida state courts, the Eleventh Circuit, or other trial courts. The Conservation Law Foundation production order was stayed pending district-court review. Whether particular AI records are relevant, preserved, discoverable, privileged, or protected work product depends on the governing law and the specific facts, including who created the material, for what purpose, under whose direction, in which system, and how it was later used.

Patrick Dempsey
Previous

A $250,000 Deposit, a $2.5 Million Escrow, and Why a Sale Order Could Not Settle the Difference
Next

Two Franchisees, Two Ways to Close a Restaurant, and the Decisions That Separate Them