Nov 12, 2026

Can expert witnesses use AI?

Mustafa Awad

Can expert witnesses use AI?

Short answer: Yes. Expert witnesses can use AI to organize, search and check case records, as long as the opinion and the checking of every fact behind it are their own. Courts have excluded an expert declaration with AI-invented citations (Kohls v. Ellison, D. Minn. 2025), but accepted a declaration from an expert who wrote his report from his own experience and used AI only to double-check it (Ferlito v. Harbor Freight Tools, E.D.N.Y. 2025).

What have courts said about expert witnesses using AI?

No court has banned expert witnesses from using AI. Where expert evidence has been excluded or struck, the problem was with AI output that was unchecked, or situations where AI was used but the expert could not explain how it was used. As we are in the early days of AI adoption, two federal trial courts have reached opposite results on whether an expert's AI prompts must be handed over. One of those orders is stayed, and no appellate court has ruled.

Date	Case, court and judge	What happened	What it means for experts
Sep 29, 2026	Villanueva v. Las Vegas Metropolitan Police Dept., No. 2:24-cv-00125 (D. Nev.), District Judge Anne R. Traum	Partly reversing a magistrate judge's order to produce the expert's whole ChatGPT log, the court protected his prompts, his AI-assisted draft material and his deposition outline. It compared prompts to an expert's "online research queries." It ordered production of the ChatGPT case-law summaries he considered, as "facts or data." Using ChatGPT did not waive privilege.	AI output you consult or consider while forming an opinion can be discovered, not only what you rely on
May 18, 2026	Conservation Law Foundation v. Shell Oil Co., No. 3:21-cv-00933 (D. Conn.), Magistrate Judge Thomas O. Farrish	Ordered the plaintiff to produce the generative AI prompts its expert used to cull the defendants' document productions, holding that "an expert witness's methodology is fair ground for discovery." The expert had disclosed her AI use in her report. District Judge Vernon D. Oliver stayed the order on June 3, 2026, pending the plaintiff's objection.	On these facts, AI used to narrow the documents was treated as methodology. The order is stayed
May 23, 2025	Concord Music Group v. Anthropic, No. 24-cv-03811 (N.D. Cal.), Magistrate Judge Susan van Keulen	Counsel used an AI tool to "properly format" citations in a declaration from the defendant's data scientist. It invented the title and authors for one citation, although the underlying article was real. The court called it "a plain and simple AI hallucination," struck part of paragraph 9, and said the error "undermines the overall credibility" of the declaration.	Only part of one paragraph was struck, but one unchecked citation hurt the credibility of the whole declaration
Apr 23, 2025	Ferlito v. Harbor Freight Tools USA, No. 2:20-cv-05615 (E.D.N.Y.), District Judge Gary R. Brown	The expert asked ChatGPT a question after finishing his report, which was based on decades of experience, and it matched his opinion. The court found "no indication" he used ChatGPT "to generate a report with false authority" and denied the motion to exclude.	AI as a checker, and not relied upon in reaching conclusions and opinions on own analysis, held up
Jan 10, 2025	Kohls v. Ellison, No. 24-cv-3754 (D. Minn.), District Judge Laura M. Provinzino	An expert declaration drafted with help from GPT-4o cited two nonexistent articles and got the authors of a third wrong. The court excluded it, denied leave to file an amended version, and said the fake citations "shatter[] his credibility." It also said it "does not fault" the expert "for using AI for research purposes."	Verification of citations is required.
Oct 10, 2024	Matter of Weber, 2024 NY Slip Op 24258 (Sur. Ct., Saratoga County), Surrogate Jonathan G. Schopf	The court found an expert's damages calculations unreliable on grounds unrelated to AI, including a wrong start date and a speculative investment proxy. In a "further and separate discussion," it noted he had used Copilot to cross-check his figures but could not recall his prompts or explain how Copilot reached its outputs. It held that counsel has "an affirmative duty to disclose the use of artificial intelligence" and that AI-generated evidence should face a Frye hearing before admission.	AI output isn't evidence on its own say-so. Your numbers and conclusions need your own method behind them, and in this court AI use must be disclosed.

A trial event, not a ruling: the Watson Grinding explosion litigation (Harris County, Tex., 2026). In the fourth bellwether trial against 3M over a fatal 2020 Houston explosion, a defense expert's ChatGPT history became evidence. He had prompted ChatGPT to "show how 3M is 0% at fault for the explosion at watson grinding." About 350 pages of chat logs were produced after plaintiffs' counsel asked for his prompts, and he was examined on them at trial. The report was characterized at trial as 85 to 90% AI-written; accounts differ on whether the expert agreed to that figure. No court ruled on his AI use, and the jury found 3M 30% responsible (Above the Law, Bracewell, The Texas Lawbook).

In federal court, one rule sits behind these rulings. Federal Rule of Evidence 702 requires the party offering an expert to show the court "that it is more likely than not" that each of its requirements is met, including that "the expert's opinion reflects a reliable application of the principles and methods to the facts of the case." New York state courts, as in Matter of Weber, apply the Frye test instead: whether a method is generally accepted in its field. A separate proposed Rule 707 on machine-generated evidence was published for comment in 2025. In May 2026 the federal advisory committee said it "does not recommend action on the proposed Rule 707 at this time" and would study a revised version further (committee report).

"Did you use AI?" How to answer under oath

Be ready to describe your work with AI before anyone asks: what the tool did, what you did, and how you checked it. Then answer truthfully. A flag is thrown when the honest answer is "it wrote my conclusions" or "I don't know how it got there."

Based on the current trend, opposing counsel asks about AI to suggest the opinion isn't yours. In our interviews, experts raised this question without being asked. The experts who were comfortable with it could describe their process in a sentence.

A truthful answer, if it matches what you actually did, sounds like this:

"Yes. I used software to organize and search the records, and to build a timeline where every entry links to its source page. I read the records myself, checked the entries I relied on against the original pages, and the opinions in my report are my own."

Expect follow-up questions, and have an answer ready for each:

Which tool, and what did it do? Name it and describe its role in one sentence.
Did it write any part of your report? Say which sections, if any, and that you reviewed and approved each one.
Did it form or suggest your opinions? The answer must be no.
How did you check its output? Explain how you traced facts back to source pages.
Do you have your prompts or chat history? Know what exists before your deposition (see the discovery section below).

This is not legal advice. Talk to the retaining attorney before your deposition about how they want AI use described and disclosed.

Can you say you reviewed every page if AI helped?

Yes, if you actually read the records and used AI to find, sort and cite pages rather than to replace your reading. The statement "I reviewed every page" is about what you did, so it is only true if you did it.

Where AI helps here is in making your review more complete. Record sets in our interviews ran from 300 to 35,000 pages per case, and the fear experts named most was missing something, not the hours. A tool that links every fact to its page, flags pages it could not read, and shows conflicting entries side by side helps you prove you saw everything. A tool that hands you a summary with no page references does the opposite: you can't show what it skipped.

On this point, it’s worth highlighting what using general chatbots and generic AI summaries entails. The risk isn't using AI technology. It's using a tool that wasn't built for record review, then testifying that you reviewed every page. General chatbots and one-click summaries compress hundreds of pages into a few paragraphs, rarely show which pages they read or skipped, and can cite facts and references that no source page supports. If you relied on output like that, you can't honestly say you reviewed every page, and opposing counsel will find the gap.

Do expert witnesses have to disclose AI use?

There is no uniform rule yet, but disclosure is increasingly expected, and hiding AI use is the bigger risk. In CLF v. Shell, the expert had disclosed her AI use in her report, and a magistrate judge still treated her prompts as discoverable methodology.

Three places decide what you disclose:

The court. Some judges have standing orders on AI; one in the Northern District of California requires lead trial counsel to certify they have personally verified the accuracy of AI-assisted content. In New York, Matter of Weber held that counsel has an affirmative duty to disclose AI use. Federal Rule 26(a)(2)(B) already requires a report to state the facts or data the expert considered.
Your engagement letter. Law firm guidance now recommends AI terms in expert retention agreements, such as disclosing the tools used, preserving prompts and outputs, and independent human verification. If your letter has AI terms, read them before you start.
The retaining attorney. Ask early how they want AI use described in the report and at deposition.

The Mayer Brown guidance for litigators recommends addressing AI use, preservation and disclosure in engagement letters and naming AI expressly in expert discovery stipulations.

Are an expert witness's AI chats and prompts discoverable?

They can be. As of November 2026, two federal trial courts have reached opposite results on prompts, one of those orders is stayed, and no appellate court has ruled as of yet.

Prompts: CLF v. Shell (D. Conn., May 2026) ordered an expert's prompts produced as methodology; the order is stayed pending an objection. Villanueva (D. Nev., Sep 2026) protected them, comparing them to an expert's online research queries. Mayer Brown notes the two may be partly reconciled on their facts: in CLF the AI was used to cull documents, in Villanueva the prompts were research questions.
AI output you considered: Villanueva ordered ChatGPT case-law summaries produced because the expert considered them. It applied the Ninth Circuit rule that disclosure reaches facts or data an expert "considered," "not only those relied upon."
AI-assisted drafts: Villanueva protected draft material under Rule 26(b)(4)(B), which covers drafts "regardless of the form in which the draft is recorded."
Whole chat histories: In the Watson Grinding explosion trial, about 350 pages of a defense expert's ChatGPT history were produced.

Which AI uses are safe for expert witnesses?

The line courts draw is between AI that helps you handle the record and AI that does your thinking. Uses on the left keep the opinion yours and every fact traceable. Uses on the right are where experts have been excluded or challenged.

Generally safe	Risky or unsafe
Organizing, sorting and deduplicating large record sets	Asking a chatbot to write your analysis or conclusions
Searching every mention of a term, drug, date or person	Asking AI to argue for one side ("show the defendant was 0% at fault")
Building a timeline where each entry links to its source page	Citing studies, cases or records you haven't opened yourself
Flagging unreadable pages and conflicting entries for you to check	Accepting a summary with no page references
Checking your finished draft against the record	Using AI before forming your own view, then adopting its answer
Drafting repetitive sections (case background, records reviewed) in your template, which you review and approve	Uploading patient records to a consumer chatbot
Grammar, spelling and formatting	Using a tool you can't describe or whose process you can't explain

Can expert witnesses put medical records into ChatGPT?

Not into a consumer chatbot plan. As a best practice, treat protected health information as if HIPAA applied: use only tools whose vendor will sign a Business Associate Agreement (BAA), and check your protective order. Consumer chatbot plans don't come with a BAA.

HIPAA itself binds covered entities, such as providers and insurers, and their business associates. An expert retained by counsel is generally neither. Records that reach an expert through litigation are usually governed instead by a protective order, which may be a HIPAA "qualified protective order" under 45 C.F.R. § 164.512(e), and by state privacy law. Check what yours allows before any record goes into an AI tool.

The major AI vendors generally offer BAAs only on enterprise, API or business-workspace agreements, not on free or individual plans.

Before uploading any case records, check that the tool:

does not train its models on your data;
has zero data retention agreements with the AI model providers it uses;
stores data in the US, if your engagement requires it;
has an independent security audit such as SOC 2 Type II.

What expert witnesses told us about AI

Between August and November 2026, Newcase conducted interviews with expert witnesses and expert firms across medicine, engineering, finance, EMS, premises liability and liquor liability. Most already use AI. What worries them is defending it.

70% already use general AI tools (ChatGPT, Claude, Gemini, Perplexity, NotebookLM) in their expert work.
60% raised the risk of being asked about AI use under oath without being prompted.
60% named HIPAA or data privacy as a condition for using any AI tool.
87% named manual record review as their main bottleneck.
Record sets ranged from 300 to 35,000 pages per case.

Where Newcase fits

Newcase is AI-powered software for expert witnesses that helps you work through large case records while you stay in control of the review. It is built around the safe uses above.

Answers link to their source pages, so you can open the original and check any fact you rely on.
Unclear facts are flagged for you to check, including unreadable pages and entries that conflict, rather than smoothed over.
Repetitive report sections are drafted in your own template, and you review and approve each one.
It never forms or writes opinions. The analysis and conclusions stay yours.
It works for any kind of expert, medical or not.
Security: SOC 2 Type II, HIPAA with a BAA, US-only hosting, zero data retention with model providers and no training on your data.

That means you can answer "did you use AI?" in one honest sentence.

Book a demo

Frequently asked questions

Can an expert witness use AI to write a report? An expert can use AI to organize records, search them and draft repetitive sections they review and approve. The analysis and opinion must be written by the expert. In a 2026 Texas explosion trial, a defense expert who had asked ChatGPT to show his client was "0% at fault" was examined in front of the jury on about 350 pages of chat logs, and his report was characterized at trial as 85 to 90% AI-written.

Will I be asked if I used AI? Increasingly, yes, at deposition or trial. Answer truthfully and describe what the tool did and what you did, in one sentence.

Do I have to disclose AI use as an expert witness? There is no uniform rule. Courts, engagement letters and retaining attorneys increasingly expect disclosure. A federal magistrate judge ordered an expert's AI prompts produced in CLF v. Shell (D. Conn., May 2026; the order is stayed), and a New York court held in Matter of Weber (2024) that counsel must disclose AI use.

Are my AI prompts discoverable? They can be. Two federal trial courts have reached opposite results: CLF v. Shell (D. Conn., May 2026) ordered prompts produced, though that order is stayed, while Villanueva v. Las Vegas Metropolitan Police Dept. (D. Nev., Sep 2026) protected them but ordered production of AI summaries the expert considered. No appellate court has ruled.

Has a court excluded an expert for using AI? Yes, for unchecked AI errors. In Kohls v. Ellison (2025), a court excluded an expert declaration that cited two nonexistent articles, and in Concord Music v. Anthropic (2025), a court struck part of a declaration paragraph over a citation whose title and authors AI had invented. In Ferlito v. Harbor Freight (2025), an expert who used AI only to double-check his own conclusions was not excluded.

Is ChatGPT HIPAA compliant for medical records? Consumer chatbot plans don't come with a BAA. HIPAA may not bind a retained expert directly, but best practice, and often your protective order, points to a tool that signs a BAA, doesn't train on your data and has zero data retention.

Can I say I reviewed every page if I used AI? Only if you did. AI can help you find and cite pages and flag what needs checking, but it can't do the reading for you.

Sources

Court documents and rules

Commentary and reporting

This article is for general informational purposes only and does not constitute legal advice. Talk to the retaining attorney before your deposition about how they want AI use described and disclosed.