Generative AI and the Damages Expert: Lessons from Matter of Weber and Kohls v. Ellison

The arrival of generative AI in litigation has, until lately, been told mainly as a story about lawyers: briefs filed with confident citations to cases that turned out not to exist, and the sanctions that followed. The more consequential development for the valuation and damages community is that the same scrutiny has now reached the expert witness. Two decisions from 2024 and 2025 — Matter of Weber in New York and Kohls v. Ellison in Minnesota — are the cases practitioners are citing, and together they sketch the emerging rules: disclose the tool, verify its output, and remember that the opinion on the cover page must remain the expert’s own.

Since 2023, a lengthening line of mostly American decisions has dealt with filings that cited authorities which, on inspection, did not exist; the consequences have escalated from modest fines toward, in at least one matter, the removal of counsel for the remainder of a case. What distinguishes Weber and Kohls is that the problem has migrated from the advocate’s brief to the expert’s report, the part of the record a tribunal leans on precisely because it is meant to carry independent, specialist authority.

I.An unexplained “cross-check” sinks a damages opinion

Matter of Weber, 85 Misc. 3d 727 (N.Y. Sur. Ct. 2024),1 arose from an unremarkable trust dispute. A beneficiary objected to the trustee’s interim account, alleging a breach of fiduciary duty in the retention of trust property rather than its sale and reinvestment. To quantify the loss, the objectant’s expert produced a supplemental damages report modelling what the trust would have earned had the property been sold earlier and the proceeds placed in a balanced index fund.

The difficulty surfaced on cross-examination. By the court’s account, the expert had “relied on Microsoft Copilot, a large language model generative artificial intelligence chatbot, in cross-checking his calculations,” yet “could not recall what input or prompt he used,” and “could not state what sources Copilot relied upon and could not explain any details about how Copilot works or how it arrives at a given output.” There was, the court noted, “no testimony on whether these Copilot calculations considered any fund fees or tax implications.” Surrogate Jonathan G. Schopf was unpersuaded, observing that nothing in the record established the reliability of the tool, whether in general or as applied. The court could not, it held, “blindly accept as accurate” figures generated by a system whose workings no one had explained.

The court then did something instructive: it ran the exercise itself. Entering a comparable prompt into Copilot on three separate court computers produced three different answers, none of which matched the expert’s. Asked whether it was accurate and reliable, the tool itself answered, in substance, that its outputs should always be verified by professionals before being relied on in court. The Surrogate’s summary — “garbage in, garbage out” — captured the point.

Having found the opinion unreliable on the merits in any event — the wrong measurement period, omitted variables, speculative assumptions, no relevant expertise — the court went further and addressed the AI question as one of first impression. Given the technology’s rapid evolution and unsettled reliability, it held that counsel bear an “affirmative duty to disclose the use of artificial intelligence,” and that AI-assisted or AI-generated evidence is subject to a Frye hearing (the test for novel scientific or technical evidence) before it may be admitted, with the court setting the hearing’s scope. The holding speaks in terms of Surrogate’s Court practice, but its logic is not so confined.

II.Hallucinated authorities, and a declaration struck

If Weber concerns an opinion that AI quietly checked, Kohls v. Ellison, 2025 WL 66514 (D. Minn. Jan. 10, 2025),2 concerns one that AI helped write, with sharper consequences. The underlying suit was a First Amendment challenge to a Minnesota statute criminalising the dissemination of election-related “deepfakes.” Resisting a preliminary injunction, the Attorney General filed a declaration from a Stanford professor — a specialist in misinformation and the dangers of AI, no less — on the harms such material poses to democratic debate.

Opposing counsel noticed that the declaration cited publications that could not be found, and suspected AI. They were right. The Attorney General’s office conceded that the professor had used GPT-4o to help prepare the declaration, and that it contained citations to two non-existent articles and a misattributed author. The proponent sought leave to file a corrected version, arguing that the substance held up regardless. Judge Laura M. Provinzino declined. A declaration made under penalty of perjury, the court reasoned, is “not a mere formality”; fabricated citations — innocent or not — broke the trust such a document is meant to carry and “shatter[ed]” the expert’s credibility. The court excluded the declaration in its entirety and refused to allow a substitute. The irony of an authority on AI-driven misinformation being undone by AI-generated misinformation was not lost on the court, which marked the moment in two words: “The irony.”

The court was careful, though, about what it was not saying. It did not fault the use of AI as such, acknowledging the technology’s genuine promise; it faulted the abdication of the human duty to check. Echoing the wave of attorney-sanction cases, it grounded the result in Rule 11’s “personal, nondelegable” obligation to verify what one files, and added its voice to a “growing chorus of courts” insisting that AI-generated content in legal submissions be verified before it is put forward.

III.Where the line falls: the expert must remain the author

Read together, the two decisions are not about banning the technology. Both ask a narrower question: when an opinion has passed through a model, who answers for what comes out? The duty of verification stays with the person who signs the report and reaches every conclusion within it; it does not pass to the model that produced the draft. The familiar defence that experts have always relied on juniors does not rescue unexamined AI use: a junior’s work can be questioned, corrected and made the expert’s own, and a junior can be held to account; a model can do none of these things, because it cannot explain its reasoning, exercise judgment under instruction, or stake a reputation on the result.

A model cannot be cross-examined; the expert who signs the report can.

That the line turns on conduct rather than on tools is confirmed by the cases on the other side of it. In Ferlito v. Harbor Freight Tools USA, Inc., 2025 WL 1181699 (E.D.N.Y. Apr. 23, 2025),3 a court declined to exclude an expert who had used a language model, because he formed his opinion first, on the strength of decades of experience, and used the tool only to confirm a conclusion he had already reached and could defend. The distinction tracks one drawn in the wider professional literature between augmentation, where the expert remains demonstrably the author, and automation, where the machine supplies the substance and the expert adds little beyond a signature. Courts are likely to tolerate, and in time to expect, the first; they are unlikely to accept the second.

This has a practical edge. Current models are strong within a band of competence and unreliable just beyond it. The edge of that “jagged frontier” is hard to see from the inside, which is exactly why authoritative-sounding errors slip through. A balanced-fund valuation and a journal citation are both the kind of discrete, checkable output a model will render fluently and may quietly get wrong. The lesson is not to avoid the tools but to check their output, and to use them only where they are reliable.

IV.The regulatory horizon

These decisions arrive as institutions move to formalise expectations. In the United States, the Judicial Conference’s evidence-rules advisory committee has floated amendments that would subject machine-generated evidence to reliability scrutiny akin to that applied to expert testimony, and would let a party challenge evidence as AI-fabricated; neither has yet been adopted. Arbitral bodies have moved faster: the Silicon Valley Arbitration & Mediation Center issued guidelines on AI in arbitration in 2024, and the Chartered Institute of Arbitrators followed in 2025, both contemplating disclosure of AI use balanced against duties of confidentiality. In England and Wales, where experts already owe an overriding duty of independence to the court, the Civil Justice Council has consulted on requiring experts to identify the tools they used and to explain the use made of them. Across these instruments the common thread is disclosure: of the tool, the manner of its use, and how heavily its output was relied on.4

For Singapore-seated work, the principles are readily portable. An expert’s paramount duty is to the court or tribunal, and the opinion must be independent and reliable; those obligations are indifferent to whether part of the analysis passed through a model. The same verification-and-disclosure discipline should be assumed to apply, and the prudent course is to adopt it now rather than wait for a local decision to compel it.

None of this counsels avoidance. Used within its competence and under genuine human supervision, AI can sharpen and speed the technical work behind a valuation or damages opinion. What Weber and Kohls establish is that the analysis remains the responsibility of the professional who signs the report and answers for it under challenge. That responsibility, the courts are now making plain, does not travel with the tool.