Expert report review
The opposing expert used AI to prepare the report: what does a reviewer check, and when do courts exclude the opinion?
In the decisions discussed below, no court barred experts from using generative AI. The courts excluded or refused to credit work the expert did not verify or could not explain: a declaration under penalty of perjury citing articles that do not exist (Kohls v. Ellison, D. Minn. 2025) and a damages figure cross-checked in a chatbot whose prompt the expert could not recall (Matter of Weber, N.Y. Sur. Ct. 2024). A reviewer tests both points.

May an expert use AI to prepare a report?
No court in the decisions below barred it. Two federal judges wrote that they did not fault the use of AI as such. They asked whether the expert verified what the tool produced and could explain each result.
In Kohls v. Ellison, No. 24-cv-3754 (D. Minn. Jan. 10, 2025), Judge Laura M. Provinzino excluded an expert declaration that contained fabricated citations and wrote in the same order: "To be clear, the Court does not fault Professor Hancock for using AI for research purposes."
In Ferlito v. Harbor Freight Tools USA, Inc., No. 2:20-cv-05615 (E.D.N.Y. Apr. 23, 2025), Judge Gary R. Brown denied a motion to exclude an expert who had run a ChatGPT query after writing his report. The court placed the problem elsewhere: it arises "when attorneys and experts abdicate their independent judgment and critical thinking skills."
For the party that receives a report, two facts matter: what the tool produced, and what the expert did with the output before signing.
What happened to the declaration with AI-generated citations in Kohls v. Ellison?
The court refused to consider it. The declaration of Professor Jeff Hancock cited two academic articles that did not exist and gave the wrong authors for a third, and the court denied leave to file a corrected version.
Christopher Kohls and Mary Franson challenged Minn. Stat. § 609.771, a Minnesota statute on election deepfakes, and moved for a preliminary injunction. Two expert declarations were filed in opposition. One of them, by Professor Jeff Hancock, contained "citations to two non-existent academic articles" and named the wrong authors for a third. Hancock had used GPT-4o in drafting. In the court's words, "GPT-4o provided Professor Hancock with fake citations to academic articles."
The plaintiffs moved to exclude both declarations. Attorney General Ellison asked for leave to file an amended Hancock declaration. The court excluded the Hancock declaration, kept the declaration of Professor Jevin West, and denied the request to amend as moot. The citation to fake sources, the court wrote, "shatters his credibility with this Court."
The declaration was made under penalty of perjury. The court wrote that it "does not believe that Professor Hancock intentionally cited to fake sources," and excluded the declaration all the same: he "submitted a declaration made under penalty of perjury with fake citations."
The order also addressed counsel. Rule 11 imposes a "personal, nondelegable responsibility" to validate what is filed, and attorneys may now need to "ask their witnesses whether they have used AI in drafting their declarations" and what the witnesses did to verify the output.
What did the New York court decide about Copilot in Matter of Weber?
It found the expert's damages calculation unreliable and held that counsel must disclose the use of AI before such evidence is offered, with a Frye hearing on the evidence prior to its admission.
Matter of Weber, 2024 NY Slip Op 24258 (Sur. Ct., Saratoga County, Oct. 10, 2024), decided by Surrogate Jonathan G. Schopf, was a trust accounting. The beneficiary objected that the trustee had kept and used a property on Cat Island in the Bahamas, which was sold in 2022 for $485,000. His damages expert was Charles Ranson. The hearing took place on May 8, May 22 and June 7, 2024, and the testimony revealed that Ranson "relied on Microsoft Copilot, a large language model generative artificial intelligence chatbot, in cross-checking his calculations."
The court recorded what followed: "Mr. Ranson could not recall what input or prompt he used to assist him with the Supplemental Damages Report." He could not explain how Copilot works. The court added that "the record is devoid of any evidence as to the reliability of Microsoft Copilot in general."
The court then ran its own test. On a court-issued computer it asked Copilot to calculate the value of $250,000 invested in the Vanguard Balanced Index Fund between December 31, 2004 and January 31, 2021. The answer was $949,070.97. Two more court-issued computers returned $948,209.63 and "a little more than $951,000.00." The court wrote that "the fact there are variations at all calls into question the reliability and accuracy of Copilot."
The court held that "counsel has an affirmative duty to disclose the use of artificial intelligence" and that evidence generated by it "should properly be subject to a Frye hearing prior to its admission." This is a decision of one New York Surrogate's Court. The calculation failed for other reasons as well: the court found that the expert used an incorrect time period, relied on a proxy index fund and left out taxes, fees and the actual 2022 sale. Its conclusion: "Mr. Ranson's damages calculations cannot be credited as they are unreliable."
When did a court find no problem with an expert's use of ChatGPT?
In Ferlito v. Harbor Freight, where the expert formed his opinion on a safer design from his own experience, wrote the report, and only afterwards ran a ChatGPT query that agreed with what he had written.
Joseph Ferlito bought a splitting maul in 2017. The head came off the handle and injured him, and he sued Harbor Freight on a design defect theory. His expert, Mark Lehnert, has no engineering degree; he reported more than a dozen patents and decades in management of engineering departments. He proposed an alternative design, a metal pin securing the head, which other mauls on the market already had.
After writing the report Lehnert asked ChatGPT about the best way to secure a hammer head to a handle. He testified that he did not rely on the answer in reaching his opinions. The court found that "there is little risk that Lehnert's use of ChatGPT impaired his judgment" and that "There is no indication that Lehnert used ChatGPT to generate a report with false authority." It concluded: "Accordingly, the Court finds no issue with Lehnert's use of ChatGPT in this instance."
The court cited Kohls for the contrast. The sequence in Ferlito was opinion first and tool second, and the expert could say under oath which was which. The objections to his lack of degrees and to the absence of peer-reviewed sources were left for cross-examination.
Which rules apply to an expert report prepared with AI?
Rule 702 on reliability and Rule 26(a)(2)(B) on the content of the report. A proposed Rule 707 on machine-generated evidence, published for comment in August 2025, addresses machine output offered without an expert witness.
Rule 702 of the Federal Rules of Evidence puts the burden on the party offering the expert. That party must show the court that it is more likely than not that the testimony is "based on sufficient facts or data" and is "the product of reliable principles and methods," applied reliably to the facts of the case. The 2023 committee note adds that "each expert opinion must stay within the bounds of what can be concluded from a reliable application" of the expert's basis and methodology. A number that comes out of a tool the expert cannot explain is measured against that sentence.
Rule 26(a)(2)(B) of the Federal Rules of Civil Procedure requires a report "prepared and signed by the witness" that lists "the facts or data considered by the witness in forming" the opinions. The 1993 committee note says the rule does not bar counsel's "assistance to experts in preparing the reports." Assistance with drafting is permitted. The opinions and the signature are the witness's.
The Preliminary Draft of Proposed Amendments released on August 15, 2025 contains a new Rule 707, "Machine-Generated Evidence." Its text: "When machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, the court may admit the evidence only if it satisfies the requirements of Rule 702(a)-(d)." The rule would not apply "to the output of simple scientific instruments." Comments were due no later than February 16, 2026.
The draft committee note also speaks to the expert who uses such a tool: "Where a testifying expert relies on such a method," it "will be scrutinized under Rule 702." The note lists the risks: "analytical error or incompleteness," "inaccuracy or bias built into the underlying data or formulas," and "lack of interpretability of the machine's process." Rule 707 was published as a proposal for public comment; the rule applied in Ferlito was the existing Rule 702.
What does a reviewer check when the opposing report may rest on AI output?
Whether every cited source exists and says what the report claims, whether each calculation can be reproduced independently, and whether the report discloses the tool, the inputs and the expert's own verification.
The three decisions supply the checklist. Each item is a fact that can be established from the report, its exhibits and the deposition transcript:
- Does every cited article, standard and decision exist, and does it say what the report attributes to it? Kohls turned on two citations to articles that did not exist.
- Can each number be reproduced from the stated inputs in a spreadsheet? In Weber one prompt returned three different totals on three computers.
- Does the report name the tool, the prompts and the output among the facts or data considered under Rule 26(a)(2)(B)(ii)?
- Can the expert explain each step of the method without the tool, as Rule 702 requires of the principles and methods behind an opinion?
- Did the opinion exist before the tool was used, as in Ferlito, or did the tool supply the opinion?
- Does the text describe the objects, measurements and documents of this case, or general propositions that would fit any case?
What can be asked about AI use at the expert's deposition?
After the report is served, the expert can be deposed on what he considered and how he reached each opinion. Drafts of the report and most attorney communications are protected; the facts or data considered are not.
Under Rule 26(b)(4)(A), "A party may depose any person who has been identified as an expert whose opinions may be presented at trial," and where a report is required, only after the report is provided. Rule 26(b)(4)(B) protects "drafts of any report or disclosure required under Rule 26(a)(2), regardless of the form in which the draft is recorded." Rule 26(b)(4)(C) protects communications between the attorney and the expert, with three exceptions: communications that relate to compensation, that identify facts or data the attorney provided and the expert considered, or that identify assumptions the attorney provided and the expert relied on.
The questions follow from Kohls and Weber. Which tool was used? What was typed into it, and what came back? Which passages of the report came from the output? Who checked each citation against the source itself? Was the calculation repeated by another method? An expert who cannot answer the second question is in the position of the witness in Weber.
A rebuttal report is an expert report under Rule 26(a)(2)(B) as well, with the same signature and the same list of facts or data considered. Under Rule 26(a)(2)(D)(ii), absent a stipulation or a court order, it is due "within 30 days after the other party's disclosure."
What does IAE prepare when the opposing report shows signs of AI use?
A review of the opposing expert's report on six questions and, on request, questions for the deposition. The format, cost, deadline and specialist are confirmed within one business day; the signed PDF follows in up to 5 business days.
IAE reviews a court-appointed or opposing expert's report in 16 fields. The review answers six questions: whether the method fits the object and the questions and was followed at every stage; whether all objects, materials and questions were examined, with no substituted questions or gaps; whether the conclusions follow from the examination, with no contradictions or inadmissible assumptions; whether the expert's education, specialty and experience match the subject; whether the report meets the formal requirements and the examination procedure of the case's jurisdiction; and whether the report can be relied on.
The client sends the report with its exhibits and describes the task: a review of the opposing report, deposition questions, or both. The secretariat confirms receipt the same day. Within one business day IAE confirms the format, the cost, the deadline and the specialist from its registry, and the client approves the candidate and the CV before work starts. A 48-hour rush is available for cases with tight deadlines.
The work is done by members of the public IAE registry, who can be verified by Member ID or QR code. The specialist confirms in the document that he or she has no interest in the outcome of the case. Materials are handled confidentially and are not passed to third parties without the client's consent. IAE is not a law firm and does not advise on case strategy. A review is evidence that the court weighs together with the other materials, and neither IAE nor the reviewer guarantees the outcome.
Sources
- Kohls v. Ellison, No. 24-cv-3754 (LMP/DLM), Document 46 (D. Minn. Jan. 10, 2025) (Justia)
- Matter of Weber, 2024 NY Slip Op 24258 (Sur. Ct., Saratoga County, Oct. 10, 2024) (Justia)
- Ferlito v. Harbor Freight Tools USA, Inc., No. 2:20-cv-05615 (GRB) (SIL), Memorandum of Decision and Order (E.D.N.Y. Apr. 23, 2025) (GovInfo)
- Federal Rules of Evidence, Rule 702 with committee notes (LII, Cornell Law School)
- Federal Rules of Civil Procedure, Rule 26 with committee notes (LII, Cornell Law School)
- Preliminary Draft of Proposed Amendments to the Federal Rules, released August 15, 2025 (United States Courts)
- Preliminary Draft of Proposed Amendments, August 2025, text of proposed Evidence Rule 707 and committee note (United States Courts, PDF)
Read also
All articles
Must a rebuttal expert offer an alternative opinion, or is a critique of the other report enough?
A rebuttal expert does not have to build a competing model. In Medidata v. Veeva (S.D.N.Y. 2022) the court repeated that "there is no requirement
Can the opposing expert change deposition testimony on an errata sheet?
A deponent who asks for review before the deposition ends has 30 days to sign a statement listing changes "in form or substance" with the reasons
Tribunal-appointed expert in arbitration: how does a party answer the report?
An arbitral tribunal may appoint its own expert unless the parties have agreed otherwise. Under Article 26 of the UNCITRAL Model Law and ArticleExpert report review
Ready to discuss your case?
Send the materials: the secretariat will confirm the format, cost, deadline and the specialist within one business day.

