Discovery of ESI in the Age of Artificial Intelligence  

PDF

Years ago, the creation of online search engines, instant messaging, and social media revolutionized discovery. Now, the prevalence of Artificial Intelligence (“AI”) has caused a similar technology-fueled upheaval. AI tools are being automatically integrated onto devices, even where users never explicitly enable them. Oftentimes, such AI functionality includes the ability to access everything shared on that device – from chats, to screenshots, to voice recordings or browser content. Conversational AI assistants even have the ability to listen into conversations and provide prompts based on discussion topics. Practitioners and stakeholders should be aware of practical tips for the discovery of AI and potential pitfalls.   

Discovery Requests 

Step 1: Ask the right questions 

Consider the following document request: 

REQUEST NO. 1: Produce any and all correspondence, search history, hand-written notes or other documentation relating and/or referring to the diagnosis of COPD.  

In the age of AI, the terminology used might not capture all of the information the requestor would ideally like to discover. For example, imagine the responding party produces a document in response to this request that was created by Microsoft CoPilot. The party asserts that the document lists all the times they were hospitalized because of their COPD in the last two years.  

But what prompts were used to create the document? Perhaps CoPilot was prompted to exclude visits to specific hospitals, or just focus on selective medical history. Was an AI agent used to generate the document? If so, was the agent reactive or model-based? This information reveals numerous factors bearing on the reliability of the document, and the responding party’s thought process. While it is possible to ask these questions during a deposition, the responses may be unhelpful due to the technical nature of the subject matter or for other reasons. Thus, practitioners should consider methods to obtain reliable answers to these questions in written discovery. 

Then there is the question of how the responding party searches for materials responsive to this request. The request above fails to specify a date range, or which platforms the requesting party intends to be searched for responsive documents. Do they simply scroll through their AI platforms history looking for the word ‘COPD’? Additionally, it may be useful for the requesting party to seek information regarding the AI source used, as some open source AI do not require an account and thus do not allow a user to obtain historical data.  

Rather than negotiating search methodology later down the line, practitioners may consider crafting requests affirmatively specifying search terms, date ranges, and examples of platforms. This avoids any ambiguity as the scope of the request, both with opposing counsel and the court. It also guides the responding party in searching for responsive material. As always, such requests must be relevant and proportional to the needs of the case. See Fed. R. Civ. P. 26(b)(1).    

Step 2: Consider who you are seeking discovery from  

While sometimes cumbersome, it is best practice to seek the ESI discovery from the opposing party in the case before subpoenaing third parties (this can generate objections relating to the burden and necessity of the requests). One problematic scenario this causes is when the responding party does not know how to obtain their search history on various AI platforms. Consider providing instructions with your discovery requests on how to obtain the relevant information. For example, a plaintiff’s Gemini activity can be downloaded using the following method: 

• Navigate to takeout.google.com. 

• Click "Deselect all" (so you don't end up downloading gigabytes of unnecessary data from your Google account). 

• Scroll down, look for the "My Activity" option, and check the box. 

• Click the button that says "All activity data included". 

• In the pop-up window, click "Deselect all" and then check only the "Gemini Apps" option. Click OK. 

• Scroll down, click “Next step” and then “Create export”.  

• Wait for Google to email you saying your file is ready, then download it. 

• Look for the main file named “My Activity.html”. Double click it, and you should have a giant, single webpage containing your entire chat history in plain text. Use Ctrl+F to search for the specific conversations or keywords you need.  

Format of Production 

As with traditional ESI, “[i]f the responding party ordinarily maintains the information it is producing in a way that makes it searchable by electronic means, the information should not be produced in a form that removes or significantly degrades this feature.” See Fed. R. Civ. P. 34(a). Additionally, under the Federal Rules of Civil Procedure, the format of production should not make it more difficult or burdensome for the requesting party to use the information efficiently in litigation. Often, this means the responding party should not be producing screenshots of their search history or documents generated using AI. Production in PDF format can be acceptable, if searchable and organized correctly. Any search history using AI tools should also have dates and timestamps included.  

Asserting Privilege 

Several courts have recently discussed how various privileges apply to AI searches. In United States v. Heppner, 820 F. Supp. 3d 292 (S.D.N.Y. 2026), the court considered whether written exchanges that the defendant had with a generative AI platform were protected by attorney-client privilege or the work product doctrine. Id. at 296. The defendant argued that he had input information he had learned from his counsel into Claude and had shared the AI-generated documents with his counsel to obtain legal advice. Id. He admitted, however, that counsel had not directed him to run the searches. Id. The court ruled that the materials were not privileged, and that inputting the information into Claude was akin to sharing it with any other third party. Id.   

In contrast, the court in Warner v. Gilbarco, Inc., 820 F. Supp. 3d 629 (E.D. Mich. 2026) found that “ChatGPT (and other generative AI programs) are tools, not persons, even if they may have administrators somewhere in the background”. Id. at 636. Accordingly, AI materials prepared by the pro se plaintiff in preparation of litigation were subject to work-product protection. Id. at 637. 

Takeaway: Discovery of information related to generative AI can provide a host of highly relevant information. However, like with the discovery of ESI more generally, it has its constraints. As such, knowing what questions to ask, and the proper scope of such discovery is vital.