Trust but Verify: Use of AI in Expert Testimony

As with all applications of legal AI, expert witnesses’ use of AI presents both opportunities and risks. Recent headlines demonstrate the pitfalls and underscore the need for outside counsel to implement safeguards to ensure that testifying experts use AI appropriately and disclose it where needed.

Recent Cases Demonstrate the Risk of Improper Expert Use of AI

In recent weeks, three cases involving expert AI use have generated substantial coverage in the legal press—and heartburn in litigators.

Most notably, in litigation against 3M involving a deadly explosion at a Houston manufacturing plant, a standard-of-care expert engineer hired by 3M used ChatGPT to help write his report. Plaintiffs counsel recognized evidence that the expert draft was AI-generated and demanded mid-deposition production of the expert’s AI prompt history. 3M’s counsel reviewed the prompts for privilege during a three-hour break in the deposition, and then produced over 300 pages of records.

The details of the ChatGPT prompts were devastating to the expert and his client. The expert explicitly asked the tool to build “an exceptional expert witness report defending the standard of care at 3M” and to show that 3M was “0% at fault.” That apportionment conclusion ultimately did not make it into his report because ChatGPT told the expert that writing “0% responsible” in a report would be an easy target for opposing counsel, and such conclusions “let opposing counsel paint him as an advocate rather than an expert.” The expert also used ChatGPT to identify basic information in the case and was forced to admit that the drafted report (for which he charged more than $90,000) was “90 to 85% ChatGPT.” These disclosures were so unhelpful that the plaintiffs ultimately called this expert in their case.

This followed an earlier decision by a magistrate judge in the District of Connecticut concluding that a plaintiff’s expert had to disclose her AI chat prompts used to identify relevant materials from the defendant’s productions. In that case, the plaintiff’s expert referenced the use of “prompts” in her filed declaration. The defendant requested the expert’s AI prompts and the plaintiff’s counsel objected to the discovery as outside the scope of discovery under Rule 26(a). The magistrate judge disagreed, explaining that “an expert witness’ methodology is fair ground for discovery” and the use of AI to “cull down” what documents the expert would consider is an aspect of the methodology. The magistrate judge’s decision is pending appeal to the district judge.

And most recently, a judge in the Western District of Washington struck an expert report and, in the absence of expert testimony, granted summary judgment in favor of defendants after the plaintiff’s causation expert was found to have submitted AI-hallucinated citations. This, the judge concluded, shattered the expert’s credibility, requiring the entire report to be excluded, even though the hallucinations were limited to specific sections of the report. The judge grounded the exclusion in Federal Rule of Evidence 702, reasoning that the lack of diligence in verifying sources meant the opinion was not based on sufficient facts or reliable methods.

Discoverability and Impact of AI Prompts

Under the Federal Rules of Civil Procedure and equivalent rules of most states (including, as relevant to the 3M case, Texas), expert reports are governed by the rule that while an expert’s methodology and the sources considered in reaching their opinions must be disclosed, draft reports are protected from disclosure. Recent case law indicates that an expert’s use of AI has the potential to elide that distinction and lead to discoverability of the underlying prompts and AI use.

This area of law remains unsettled, but it is developing rapidly as courts confront the intersection of generative AI and long-standing work-product protections. The early decisions on this question demonstrate that experts’ use of AI can go beyond mere drafting aides and actually become a disclosable methodological component or source considered. For example, in both the Texas state court and Connecticut federal court cases described above, the courts required the experts to produce AI prompts and chat history because those materials went beyond mere drafting aides—they were part of the experts’ methodology and materials considered. Thus, the protection for mere drafts was inapplicable where the AI work went to the bases and substance of the expert’s opinions.

As the Texas and Connecticut cases illustrate, there is a danger that an expert’s AI prompts—and the iterative process behind them—are relevant discovery tied to expert’s methodology or bias rather than as draft material protected by discovery rules. This is in contrast to AI prompts from counsel, which courts have protected on the grounds that they reflect the mental impressions of counsel and therefore are protected under opinion work product. For example, in Tremblay v. OpenAI, the Northern District of California found that AI prompts reflecting counsel’s mental impressions were shielded from disclosure as opinion work product, describing them as “virtually undiscoverable.” However, although there are arguments that an expert’s AI prompts may reflect opinion work product, the Texas case shows that counsel should not assume that an expert’s use of AI will be protected to the same degree as traditional draft work, and should monitor this space closely as additional courts weigh in.

Beyond the threshold question of discoverability, the way an expert uses AI can itself become a liability. While litigators cannot control the process or outcome of experts, unethical or sloppy use of AI will certainly impact the credibility of the party and its counsel. If opposing counsel can show that the expert’s methodology was flawed because, for example, use of AI introduced unverified data or fabricated citations, or introduced analysis the expert cannot independently explain and defend, the expert’s credibility may be seriously called into question. As seen in the Western District of Washington case described above, such a showing can support a motion to exclude the expert’s testimony or to strike the report altogether.

Ultimately, whether an AI chat is a drafting tool or a source is going to be a question of degrees. At one end of the continuum, using an AI model to check a report for typos and consistent terminology is little more than an advanced spell check, and is likely to be considered shielded as part of the drafting process. On the other end, using large language models and machine learning to scrape, process, and analyze large volumes of data is a methodological design that almost certainly has to be (and in most cases, already is) disclosed. But there is a lot of gray area between those two poles, and very little guidance to date from courts.

Key Considerations and Best Practices

Given the emergent risks, litigation counsel should be prepared to speak with expert witnesses as part of the engagement process about their use of AI and what safeguards they have in place to ensure that the process is reliable, preserved and disclosed when necessary.

The safest approach, of course, is to instruct expert witnesses not to use AI at all. That, however, may not prove feasible, particularly in data-heavy reports and in fields where standard practice now encourages or even requires the use of AI. If an expert is to use AI, then the baseline rule should be the same rule always given (and frequently ignored) for electronic communications: the expert should use AI with the assumption that every prompt she writes and response she receives will be read in open court.

Beyond that, the following practical safeguard should limit the risk that expert use or misuse of AI will undermine the client’s case:

• Establish clear ground rules: For every engagement, counsel should speak with the expert at the outset about permitted uses of AI, and consider including in the engagement letter that AI should not be used absent agreement between counsel and the expert.
• Carefully draft prompts: Experts should frame prompts in a way that they would feel comfortable being read in open court.
• Consider the appropriate use case: Experts and counsel should carefully consider the appropriate use cases for AI. For example, machine-learning AI can be a useful tool for analyzing large volumes of data, generating samples, or identifying trends. But using AI to draft sections of reports opens the expert to credibility attacks, even if the draft ultimately reflects the expert’s opinions based on a sound methodology.
• Avoid being an AI rookie in court: If AI is not something an expert uses ordinarily in the course of his or her work, or is not employed by people in the field, the expert and counsel should think very carefully before deploying it for the first time in the litigation context. This not only minimizes the risk of sloppy mistakes, but is consistent with the requirement that an expert’s opinions be a “product of reliable principles and methods.”
• Filter evaluative AI analyses through other members of the case team: When AI is used to review, pressure test, or red team an expert report, that should be performed by counsel or by consultants retained to support the expert work stream. AI will often identify analytical gaps or additional sources, and counsel can communicate the feedback back to the expert. Filtering those recommendations through lawyers or consultants will limit the risk that the entire AI chat will have to be disclosed.
• Disclose all sources considered: Where an AI tool directs the expert to an article, document, or other source, the underlying source should be disclosed as part of the expert’s materials considered.
• Cite check, cite check, cite check: One of the easiest protections to implement is a thorough check of all sources and citations within an expert report. The report should rely on real authorities, and both counsel and experts alike should ensure that nothing is an AI hallucination.
• Check local rules: Most importantly, counsel should be aware of the rules in the jurisdictions where they practice, and keep abreast of new developments in this space.

The lessons of AI are converging fast. For experts and the lawyers who retain them, the message is blunt: AI can assist, but the opinion has to be the expert’s, and everything the expert types into the machine may one day be read aloud to a jury.

Reprinted with permission from the September 28, 2026 edition of “The Legal Intelligencer” © 2026 ALM Global Properties, LLC, trading as Centellic. All rights reserved. Further duplication without permission is prohibited, contact 877-256-2472 or asset-and-logo-licensing@alm.com.