By Bill Hartzer · Published · How this site is researched
Experts, their reports and the systems behind them
This section records matters in which an expert witness used an AI system in preparing evidence, and what the tribunal did in response. It also records court rules written specifically for expert reports. Lawyers' own filings with hallucinated case citations are recorded elsewhere on this site. The line matters because an expert witness's report is a different kind of document. It is the expert's statement of opinion, usually made under oath or a declaration of truth, and courts test it under evidence rules such as Federal Rule of Evidence 702 rather than under the rules that govern counsel's briefs.
The matters fall into 4 groups: declarations and reports that carried fabricated citations; reports in which the analysis itself was produced or checked by a system; challenges to experts that did not succeed despite AI use; and rules outside the United States that govern how experts may use these systems.
Citations, calculations and unexplained output
The most common pattern is the fabricated or corrupted citation. In Kohls v. Ellison, an expert declaration drafted with GPT-4o cited 2 articles that do not exist. In Concord Music Group v. Anthropic, the source was real and had been found by a person, but Claude.ai, used by counsel to format the reference, produced a wrong title and wrong authors. In the Valsartan multidistrict litigation, a toxicologist's report cited non-existent sources after he used an AI tool without verifying them. The mechanism differs in each matter. The result in each was a citation that did not match a real source.
The second pattern is delegated or unreproducible analysis. In In re Celsius Network, a 172-page valuation report was written by an AI system at the expert's direction in 72 hours. In Matter of Weber, a damages expert cross-checked his figures with Microsoft Copilot and could not recall his prompts. In the 2026 Watson Grinding explosion trial in Houston, an engineering expert for 3M produced about 350 pages of ChatGPT conversations, and prompts asking the system to show that his client was 0% at fault became the subject of examination at trial.
The third pattern is less often reported: an expert who used a system in a limited, ordered way. In Ferlito v. Harbor Freight Tools, the expert asked ChatGPT a question after his report was complete, and the court found no reason to exclude him.
Matters in this topic
10 matters, most recent first within each group. 9 were checked against the primary document, and the others rest on reporting and say so.
Declarations with fabricated citations
- LeDoux v. Outliers, Inc.Primary document readPlaintiff's counsel used an AI system to generate a citation table and gave it to two experts, who attached it to their reports. The court personally sanctioned counsel $3,000 and later excluded a third expert whose report contained fabricated citations.
- Concord Music Group v. Anthropic (Chen declaration)Primary document readCounsel used Claude.ai to format citations in a sampling expert's declaration, and it produced a wrong title and wrong authors for a real article. The magistrate judge struck paragraph 9 and found the error undermined the declaration's credibility.
- Kohls v. EllisonPrimary document readAn expert defending Minnesota's election deepfake law used GPT-4o to help draft his declaration, which cited 2 articles that do not exist. The court excluded the declaration and denied leave to file a corrected one.
AI-generated analysis
- Watson Grinding explosion trial (3M expert)Reported; document not locatedAn engineering expert retained by 3M in Houston explosion litigation used ChatGPT to draft most of his report and produced about 350 pages of chat logs. He was examined on them at trial. The jury found 3M 30% responsible.
- Matter of WeberPrimary document readA damages expert in a New York trust accounting used Microsoft Copilot to cross-check his calculations and could not recall his prompts. The Surrogate's Court rejected his analysis and announced a disclosure and Frye hearing requirement for AI-generated evidence.
- In re Celsius Network LLC (Faraj report)Primary document readAn objector's valuation expert in the Celsius bankruptcy directed an AI system to write a 172-page report in 72 hours. The bankruptcy court excluded the report as unreliable and admitted his live testimony.
Discovery of the expert's AI use
- Conservation Law Foundation v. Shell (expert AI prompts)Primary document readAn expert's team used a generative AI system to cull the defendants' document production. A magistrate judge ordered the prompts disclosed as part of the expert's methodology. The order is stayed while an objection is decided.
Challenges that did not succeed
- In re Valsartan MDL (Dr. William Sawyer)Primary document readA plaintiff's toxicology expert in the Valsartan MDL cited non-existent sources after using an AI tool without verifying them. The special master denied the motion to exclude and left the issue to costs and cross-examination.
- Ferlito v. Harbor Freight Tools USAPrimary document readA plaintiff's product-design expert asked ChatGPT a question after completing his report, and the answer matched his opinion. The defendant moved to exclude him on that basis. The court denied the motion.
Outside the United States
- NSW Supreme Court Practice Note SC Gen 23Primary document readThe Supreme Court of New South Wales issued a practice note barring the use of generative AI to draft or prepare expert reports without prior leave. Where leave is given, the expert must disclose the use and annex a record of it.
From a struck paragraph to an excluded report
Responses have varied with the facts. The remedies recorded here run from exclusion of the whole declaration (Kohls) to striking a single paragraph (Concord), to excluding a written report while admitting the expert's live testimony (Celsius), to rejecting a damages analysis on grounds that were largely independent of AI (Weber). In the Valsartan matter a special master denied exclusion and said the citation problem might warrant costs and cross-examination instead. In Ferlito the court denied a motion to exclude outright.
Sanctions against expert witnesses are rare in this record. None of the court orders reviewed for these entries imposed a monetary sanction on an expert. Courts have instead acted on the evidence itself, by excluding it, striking part of it, or giving it less weight, and several have addressed counsel. The Kohls order suggested that a reasonable inquiry under Rule 11 may now require attorneys to ask their witnesses whether they used AI and how they verified it. Weber announced that counsel must disclose AI-generated evidence before it is introduced in that court and that such evidence should be subject to a Frye hearing (New York's general-acceptance test for scientific methods).
Discovery is a newer front. In Conservation Law Foundation v. Shell Oil Co., a federal magistrate judge in Connecticut ordered production in May 2026 of the AI prompts an expert's team used to cull a document production, calling an expert's methodology "fair ground for discovery." That order was stayed pending an objection. The Watson Grinding logs came to light through discovery as well.
Sorting the expert matters by clause
Read against the Expert Record Standard, the matters sort by clause more cleanly than by outcome. The citation matters engage Clause 3, independent authority: nothing enters the work on a system's authority, and every citation is confirmed against the source. They also engage Clause 4, because a citation is the most checkable element of a report, and in each matter the check that would have caught the defect was simple.
The analysis matters engage Clause 2, non-delegation, and Clause 5, reproducible analysis. Celsius and the Watson Grinding trial raise the question of whose opinion the report states. Weber shows what happens when a figure depends on a chat session that cannot be re-run. The court put the same question to Copilot on 3 computers and received 3 different answers.
Clause 1, sequence, separates Ferlito from the Watson Grinding trial. In Ferlito the expert formed and wrote his opinion first and consulted the system afterward. In the Watson Grinding trial the reported prompts stated the conclusion before any analysis. Clause 7, contemporaneous record, runs through all of them. Experts who could not say what they had asked a system were at a disadvantage. Where a complete log existed, it became the evidence on which the expert's process was tested.
Use of AI has not disqualified an expert on its own
The record does not show that courts treat AI use by an expert witness as disqualifying in itself. Ferlito admitted testimony from an expert who used ChatGPT, and in the Valsartan matter an expert whose report contained non-existent sources was not excluded. Nor does it show a uniform remedy for fabricated citations. The orders discuss how much of the opinion rested on the defective material and whether the underlying method was otherwise sound.
The record also does not yet contain a reported decision from England and Wales, Canada or Australia in which an expert's own use of AI was ruled upon, as far as the searches for this section found. Activity outside the United States has taken the form of rules. The Supreme Court of New South Wales has required leave since February 2025 before generative AI is used to prepare an expert report. The Civil Justice Council in England and Wales consulted in 2026 on requiring experts to explain their use of AI. Finally, several matters sometimes cited in this area, including the Bard IVC Filters and Social Media Adolescent Addiction multidistrict litigations, were checked and no AI-related expert ruling was found in them. They are not recorded here.
Reading an expert-AI ruling from the document outward
When a new matter is reported, start with the document. Many early accounts blur the difference between an expert witness's error and a lawyer's, as the Concord matter showed when the error turned out to come from counsel's formatting. Establish who used the system, at what stage, and for what task. Then establish what the tribunal actually did: exclusion, striking, reduced weight, costs, or no action. Commentary often describes an expert as "thrown out" when the order struck a paragraph or left the issue to cross-examination.
Check whether the ruling rested on AI at all. In Weber the expert's analysis failed on several grounds that had nothing to do with Copilot. In the Valsartan bellwether the special master's ruling became moot when the case ended on a different expert's exclusion. Note whether the matter is news only. The Watson Grinding account, for example, rests on reporting and not yet on a court filing. Each entry in this section states whether the primary document was located and opened, and lists what could not be confirmed.
Primary documents behind this topic
The orders, opinions and published texts that entries in this topic were checked against. Read them before relying on any summary, including the ones here.
- LeDoux v. Outliers, Inc.
Order on Sanctions, LeDoux v. Outliers, Inc., No. 3:24-cv-05808-TMC (W.D. Wash. July 24, 2026), Dkt. 265 - Conservation Law Foundation v. Shell (expert AI prompts)
Conservation Law Foundation, Inc. v. Shell Oil Co., No. 3:21-cv-00933 (D. Conn. May 18, 2026), ECF 970 (order granting motion to compel) - In re Valsartan MDL (Dr. William Sawyer)
Special Master Order, In re Valsartan, Losartan, and Irbesartan Products Liability Litigation, No. 1:20-cv-00946, MDL No. 2875, Dkt. 3167 - Concord Music Group v. Anthropic (Chen declaration)
Order on Joint Discovery Submissions, Concord Music Group, Inc. v. Anthropic PBC, No. 24-cv-03811-EKL (SVK), Dkt. 377 - Ferlito v. Harbor Freight Tools USA
Memorandum of Decision and Order, Ferlito v. Harbor Freight Tools USA, Inc., No. 20-5615 (GRB) (SIL) - NSW Supreme Court Practice Note SC Gen 23
Supreme Court Practice Note SC Gen 23: Use of Generative Artificial Intelligence (Gen AI) - Kohls v. Ellison
Order Granting in Part and Denying in Part Plaintiffs' Motion to Exclude Expert Testimony and Denying Defendant's Motion for Leave to File an Amended Expert Declaration (ECF 46) - Matter of Weber
Matter of Weber, 2024 NY Slip Op 24258, 85 Misc. 3d 727 (Sur. Ct., Saratoga County, File No. 1845-4/B) - In re Celsius Network LLC (Faraj report)
Memorandum Opinion Approving the CEL Token Settlement and Resolving Issue of Collateral Ownership, In re Celsius Network LLC, No. 22-10964 (MG), 655 B.R. 301
Questions about this topic
Can an expert witness use AI in preparing a report?
Yes, nothing in the U.S. decisions recorded here prohibits it, and Ferlito v. Harbor Freight Tools admitted testimony from an expert who used ChatGPT. Some courts outside the United States restrict it: the Supreme Court of New South Wales requires prior leave before generative AI is used to prepare an expert report. Use that fabricates sources or replaces the expert's own analysis has led to exclusion.
Does a fabricated citation always get an expert excluded?
No, the remedy has varied. Kohls v. Ellison excluded the whole declaration, Concord Music Group v. Anthropic struck 1 paragraph, and a special master in the Valsartan litigation denied exclusion and left the issue to cross-examination and possible costs. In the orders reviewed, the courts looked at how much of the opinion depended on the defective material.
Must an expert disclose use of AI?
It depends on the forum and any orders in the case. Matter of Weber stated that counsel in that New York court must disclose AI-generated evidence before introducing it, and the New South Wales practice note requires disclosure in the report where leave is granted. Where no rule applies, disclosure may still be required by the engagement or drawn out in discovery.
Can opposing counsel obtain an expert's AI prompts?
Yes, it has happened. In Conservation Law Foundation v. Shell Oil Co. a magistrate judge ordered production of prompts an expert's team used to filter documents, treating them as part of the expert's methodology, though that order was stayed pending objection. In the Watson Grinding trial, about 350 pages of an expert's ChatGPT conversations were produced and used in examination.
What should an expert keep to answer questions about AI use?
A contemporaneous log of each use: the date, the tool and version, the task, the input, and what was done with the output. The expert in Matter of Weber could not recall his prompts, which left the court without a basis to assess them. A log also lets an expert show that the opinion came first and the system was consulted afterward.