From Work Product to Exhibits: The Emerging Discoverability of AI Prompts

Generative AI has rapidly moved from experimental tool to everyday assistant in law firms and corporate legal departments. But as lawyers use AI to brainstorm, research, and draft, a new question is surfacing in discovery battles: are AI prompts protected work product, or are they discoverable evidence? This emerging issue is reshaping how litigators think about confidentiality, privilege, and ethics in a world where much early case analysis now happens through an AI interface.

Why AI Prompts Are Becoming a Discovery Flashpoint

Litigators have long relied on mental impressions, handwritten notes, and internal memos to analyze cases—materials historically protected as attorney work product. Generative AI introduces a new category of material: the prompts and instructions lawyers feed into AI tools, and the outputs those systems return. As these tools become embedded in research, drafting, and case strategy, courts and counsel are beginning to ask whether this digital dialogue is shielded from discovery, or whether it can be compelled like emails or draft documents.

The answer is rarely binary. Instead, it depends on how AI is used, what information is shared, the nature of the claims and defenses, and the specific discovery requests at issue. In some circumstances, prompts may look like classic attorney work product; in others, they may function as fact statements, admissions, or even standalone exhibits.

What Exactly Is an AI Prompt in the Legal Context?

To evaluate discoverability, it helps to be precise about what counts as an "AI prompt" in legal practice. Generative AI interactions often include several layers of content that may have different legal implications.

Core Components of an AI Interaction

• Initial instruction: The high-level query or command (e.g., "Summarize the key issues in this complaint").
• Context and facts: Case-specific descriptions, client names, dates, or document excerpts pasted into the tool.
• Constraints and preferences: Stylistic or strategic directions ("argue aggressively," "assume we represent the plaintiff").
• Follow-up questions: Iterative clarifications and refinements entered after the initial output.
• AI-generated output: Draft arguments, outlines, summaries, timelines, or checklists produced in response.

Each of these elements can raise distinct privilege and work product questions. A generic instruction detached from facts may be relatively benign, while detailed factual inputs could reveal client confidences or litigation strategy.

Types of AI Tools Used by Legal Teams

The discoverability analysis also varies based on the type of platform and deployment model:

• Public consumer tools: Web-based interfaces available to anyone, potentially with broad data-use rights in their terms of service.
• Enterprise SaaS products: Contracted tools with business-grade security, configurable logging, and data protections.
• Law firm–hosted models: Systems deployed in a firm’s own environment or virtual private cloud with stricter access controls.
• Client-owned or on-prem solutions: AI tools integrated into a client’s infrastructure, subject to its information governance rules.

Even when the underlying technology is similar, courts may view the control over data, logging, and sharing differently depending on this ecosystem.

The Work Product Doctrine Meets Generative AI

The work product doctrine protects materials prepared in anticipation of litigation or for trial by or for a party or its representative. Historically, this has covered attorney notes, research memos, drafts, and strategy documents. AI prompts and outputs can fit comfortably into this category—or fall well outside it—depending on circumstances.

When AI Prompts Resemble Classic Work Product

In many scenarios, prompts look very much like internal attorney thought processes, which are at the core of work product protection.

• Strategy brainstorming: Asking an AI tool to “list potential counterclaims” based on a confidential case summary reflects legal theories rather than underlying facts.
• Issue spotting and outlines: Prompts that request possible arguments, defenses, or cross-examination themes mirror traditional legal research workflows.
• Draft refinement: Using AI to polish language, rephrase sections, or suggest better transitions for a brief that will later be manually reviewed by counsel.

In these use cases, the prompt and resulting output are intertwined with the lawyer’s mental impressions. Many courts are likely to treat them as opinion work product, which receives especially strong protection and is very rarely discoverable.

When Prompts Drift Toward Factual Discovery

Not all AI interactions are primarily strategic. Some prompts are essentially factual questions, and those can bear more directly on discoverable information.

• Recitations of events: A long prompt describing what witnesses said, what documents exist, or what the client reported.
• Embedded documents: Large blocks of contract language, emails, or reports pasted into the AI system for summarization.
• Numeric analyses: Transaction data, sales figures, or other quantitative information provided so the tool can create charts or models.

When prompts contain or restate core facts, opposing parties may argue that they are simply another manifestation of relevant information—akin to a summary chart or an internal investigative note. Courts may need to distinguish between protecting the lawyer’s framing and impressions versus withholding factual content that is otherwise discoverable through other means.

Privilege, Confidentiality, and Ethical Duties

Attorney–client privilege and confidentiality also intersect with AI prompts, often in ways that are distinct from work product analysis. Even if a prompt is unquestionably privileged, that status can be jeopardized by how the tool is configured and who has access to the data.

Privilege Concerns in AI Workflows

Privilege generally protects confidential communications between lawyer and client made for the purpose of seeking or providing legal advice. AI tools complicate this in at least three ways:

• Third-party involvement: AI vendors may be seen as third parties. Contracts, data handling practices, and the functional role of the tool can influence whether the vendor is treated like a translator/consultant or a waiver-inducing outsider.
• Data storage and reuse: If prompts are stored, logged, or used to train models accessible to others, opposing counsel may argue that confidentiality has been compromised.
• Shared environments: When tools are integrated across multiple clients, business units, or firms, access controls become critical to maintaining privilege boundaries.

Well-structured vendor agreements, clear data segregation, and explicit confidentiality commitments can help align AI use with established doctrine around interpreters, e-discovery providers, and other legal service vendors.

Professional Responsibility and Client Communication

Beyond evidentiary rules, lawyers must consider duties of competence, supervision, and communication. Ethical implications include:

• Explaining AI use: Clients may require disclosure about when and how their data will be input into AI tools, particularly in sensitive matters.
• Accuracy and verification: Reliance on AI outputs without adequate review can lead to misstatements in filings or discovery responses.
• Supervision of nonlawyers and technology: AI vendors and platform administrators may be treated similarly to nonlawyer staff for supervision duties.

These ethical dimensions indirectly shape discoverability. For instance, if a court finds that AI was used in a way that compromises accuracy or confidentiality, it may be more inclined to allow probing into how and why prompts were used.

How AI Prompts Can Become Exhibits

Although prompts may begin life as internal tools, they can end up as exhibits when they become independently probative or when parties try to rely on them to prove or defend claims. Several scenarios could bring prompts into the evidentiary spotlight.

Reliance on AI in Key Decisions

Where a party bases a material decision on AI analysis—such as choosing to terminate an employee, decline a contract, or adopt a safety measure—prompts can become relevant to show what the decision-maker knew and considered at the time.

• Employment disputes: An AI-generated risk assessment used in a termination decision might be subject to discovery, including prompts describing the employee’s performance.
• Compliance investigations: If an internal probe leans heavily on AI summaries, opposing parties may seek the underlying prompts to test completeness and bias.
• Product liability: Safety recommendations generated through AI, particularly when partially followed or ignored, could be central exhibits.

In these contexts, the AI interaction functions more like an expert memo or consultant report than a private brainstorm. Courts may treat the prompt–output record as part of the decision-making file that discovery is designed to illuminate.

AI as a Witness of Process

In disputes over whether a party conducted a reasonable search, responded appropriately to a demand, or made a diligent inquiry, prompts may be used to show the process that was followed.

1. Opposing counsel questions whether a party adequately searched for certain categories of documents.
2. The party explains that it used an AI tool to identify relevant custodians and repositories.
3. The court or adversary requests details: what was asked, which data collections were included, and what limits were imposed.
4. Prompts and log files become a key record of the process, functioning like a search protocol in traditional e-discovery.

Here, prompts act less like privileged musings and more like procedural steps that can be evaluated for reasonableness and completeness.

E-Discovery Implications: Collection, Review, and Production

AI prompts introduce new practical questions for those managing discovery, especially in organizations where AI tools are integrated across multiple practice areas or business units.

Where Are Prompts Stored?

Before you can decide whether prompts are discoverable, you need to know where they live and how they can be accessed:

• Application logs: Many tools keep detailed logs by user, date, and session.
• Exported transcripts: Lawyers may copy prompt–response transcripts into notes, emails, or document management systems.
• Screen captures: Screenshots of particularly useful outputs sometimes end up in slide decks or summaries.
• Structured databases: Certain enterprise tools store queries and outputs in searchable repositories for reuse.

Each of these locations may fall within the scope of a reasonable e-discovery collection, depending on the case’s issues and proportionality considerations.

Searching and Redacting AI Content

Even when prompts are discoverable in principle, they may contain intertwined privileged and non-privileged content. That creates a familiar—but technically challenging—task:

• Separating core facts from attorney mental impressions embedded in the same prompt.
• Applying redactions to retain responsive factual material while shielding strategy.
• Designing search terms and filters that can locate relevant AI interactions across large log repositories.

Legal teams will likely need to collaborate closely with information governance, IT, and AI vendors to develop export formats and redaction workflows that respect both discovery obligations and privilege boundaries.

Comparing Approaches: Conservative vs. Integrated AI Use

Organizations are not uniform in how heavily they rely on AI for legal work. Their posture toward prompts and discoverability tends to fall along a spectrum.

Designing AI Policies with Discoverability in Mind

To manage emerging risks, both law firms and corporate legal departments are drafting AI usage policies. A thoughtful policy can greatly influence how courts view the role of AI and the reasonableness of your discovery practices.

Core Elements of a Litigation-Aware AI Policy

• Approved tools list: Identify which AI platforms may be used for client matters and under what conditions.
• Confidentiality rules: Define what types of client information may be entered, and when anonymization or obfuscation is required.
• Logging and retention: Specify whether prompts are logged, for how long, and how they are segregated by matter or client.
• Review expectations: Require human verification of outputs before they influence filings, advice, or major decisions.
• Discovery readiness: Explain how AI records will be handled in preservation, collection, and privilege review.

Training Lawyers and Staff

Written policies are only effective if they are internalized in daily work. Training should cover:

• Concrete examples of appropriate vs. risky prompts.
• How to recognize when an AI interaction is functionally creating a record that might later become an exhibit.
• Steps to preserve, tag, or document important AI-derived conclusions in a way that supports future explanations to a court.

Embedding these concepts into onboarding and periodic refreshers can materially reduce the likelihood of unexpected discovery disputes.

Seven Practical Safeguards for Legal Teams Using AI

Legal professionals do not have to wait for comprehensive case law to adopt prudent practices. The following safeguards can help teams benefit from AI while preparing for potential discoverability challenges.

1. Separate Factual Uploads from Strategy Prompts

Where feasible, use different sessions—or even different tools—for factual summarization versus high-level legal brainstorming. This can make it easier to argue that certain logs are largely factual and others are steeped in opinion work product.

2. Minimize Identifiers When Possible

For early-stage idea generation, consider anonymizing parties and locations, especially in tools where data segregation is less robust. Reducing specific identifiers can mitigate exposure if prompts are later scrutinized.

3. Control Access to Prompt Logs

Restrict who can see historic prompts for a given matter, and align access controls with your broader privilege protocols. Logs should not be a general firm-wide resource when they embed confidential strategy.

4. Align Vendor Contracts with Privilege Needs

Ensure AI vendors commit to:

• Confidential handling of client data.
• Clear restrictions on training models with your prompts, unless expressly agreed.
• Detailed descriptions of where prompts are stored and for how long.

These terms support arguments that vendor involvement does not waive privilege or undermine confidentiality.

5. Document Critical AI-Assisted Decisions

When an AI output has a material influence on a legal or business decision, consider documenting:

• The nature of the tool used.
• Key assumptions or data provided.
• The human review and judgment applied before acting.

Such documentation can be valuable if courts later examine whether your reliance on AI was reasonable and appropriately supervised.

6. Embed AI Repositories in Litigation Holds

Update your litigation hold templates to explicitly reference AI tools and prompt logs. This helps ensure that potentially relevant AI records are preserved alongside email, chat, and document repositories.

7. Coordinate Early with Opposing Counsel

In complex matters where both sides rely on AI, consider addressing AI use in early meet-and-confers. Setting expectations about scope, formats, and privilege-handling for AI-derived materials may prevent later disputes and motion practice.

Looking Ahead: How Courts May Shape the Doctrine

The law governing AI prompts is still nascent. Over time, a few themes are likely to emerge as courts confront concrete disputes:

• Function over form: Courts will probably focus less on whether something is labeled a "prompt" and more on what role it played—fact statement, strategic memo, or procedural record.
• Analogy to existing categories: Prompts may be analogized to search queries, drafts, consultant reports, or work product, depending on context.
• Proportionality and burden: The sheer volume of logged interactions may push courts to limit AI discovery to clearly relevant subsets.
• Standards for reasonable AI use: As norms solidify, courts may express expectations for how counsel should supervise AI-assisted tasks, including discovery.

For now, practitioners must operate in a gray space, applying familiar principles of work product, privilege, and proportionality to a new class of digital artifacts.

Final Thoughts

AI prompts sit at the intersection of innovation and legal tradition. They can function as protected windows into attorney thought processes, as factual recitations, or as documentary evidence underpinning critical decisions. Their ultimate treatment in discovery will depend heavily on how they are used, where they are stored, and whether legal teams proactively manage them as part of their broader information governance strategy.

By clarifying internal policies, tightening vendor arrangements, training lawyers on safe prompting, and planning for how AI artifacts fit into litigation holds and discovery workflows, organizations can harness the power of generative AI while staying prepared for the day those prompts might be requested, scrutinized, or even shown to a jury.

Editorial note: This article provides a general discussion of emerging issues around the discoverability of AI prompts in litigation and is not legal advice. For further context and commentary, see the original insight at Baker Botts Our Take.