AI in expert reports: what the 2025–2026 rulings mean for your methodology
A plain-English guide for forensic experts and the lawyers who retain them — what has sunk experts who used AI, what has been fine, and how to stay on the right side of it.
For most of the last two years, the headlines about AI in litigation were about lawyers, with briefs citing cases that never existed. The story has now moved to the witness stand. In a string of 2025–2026 decisions, courts have struck expert declarations built on AI-fabricated sources, rejected AI-derived analysis an expert couldn't explain, and, in one closely watched order, directed an expert to hand opposing counsel the AI prompts behind her work as discoverable methodology.
None of this means an expert can't use AI. It means the way you use it, and your ability to account for it, is now part of your credibility. Here is what the cases actually say, and a practical way to stay on the right side of them.
What has sunk experts who used AI
The losing pattern is consistent. An AI tool introduced something the expert didn't supply and didn't catch, and the unverified output, not the use of AI as such, is what did the damage.
In Kohls v. Ellison (D. Minn., Jan. 10, 2025), the Minnesota Attorney General's office offered a declaration from a Stanford misinformation scholar in a challenge to the state's political-deepfake law. The expert had used GPT-4o to expand his notes and let it fill in citations; it produced two articles that do not exist and misattributed a third. The court struck the entire declaration, writing that the fabricated citations “shatter[] his credibility with this Court.” It also noted the irony of a misinformation expert failing to verify AI output.
A few months later, in Concord Music Group v. Anthropic (N.D. Cal., May 2025), a data-scientist's declaration cited a real journal (correct volume, page, year, and link) but with a title and authors the AI had invented. The court struck the paragraph and drew the line plainly. There is “a world of difference between a missed citation and a hallucination generated by AI,” and even a single one “undermines the overall credibility” of the testimony.
And in Matter of Weber (N.Y. Surrogate's Court, 2024), an expert used an AI chatbot to “check” his valuation math but could not recall the prompts or explain the tool's methodology. The court found the valuation unreliable and signaled that AI-derived evidence may require disclosure and a reliability hearing before it can come in at all.
What has been fine
The contrast is just as instructive. In Ferlito v. Harbor Freight Tools (E.D.N.Y., Apr. 2025), a long-experienced expert drafted his report independently and used an AI tool only to confirm conclusions he had already reached on his own. The court allowed the testimony. The difference wasn't whether AI touched the work. It was that the expert's own judgment drove the opinion, the AI didn't originate anything, and the expert could stand behind every word.
That is the safe-harbor pattern the cases keep pointing to: AI as an assistant the expert supervises and verifies, never a substitute for the expert's analysis.
The new frontier: your AI use may be discoverable
Striking a bad citation is one thing. The more structural shift is in discovery. In Conservation Law Foundation v. Shell Oil (D. Conn.), a magistrate judge ordered an expert to produce the AI prompts she had used to narrow a large document production. The judge held that the process was “part of that methodology and therefore discoverable” under Rule 26, and rejected the argument that the prompts were protected “notes.”
Two cautions on that ruling. First, it is a magistrate decision that, as of this writing, has been objected to under Rule 72(a) and stayed pending the district judge's review. So it is a signal of where things are heading, not settled law. Second, the principle behind it is not new. An expert's methodology has always been fair game on cross. What is new is the recognition that how you prompted an AI tool can be part of that methodology. If that holds, an expert who can't reconstruct what the tool was given and what it produced is exposed.
Why this fits Rule 26 and Rule 702
None of this is a special “AI rule.” It is the existing framework applied to a new tool. Rule 26(a)(2)(B) already requires a report to disclose the basis and reasons for each opinion and the facts or data the expert considered. Amended Rule 702 (effective December 2023) requires that the opinion rest on sufficient facts or data and reflect a reliable application of reliable methods, and it puts the burden on the party offering it. Outsource part of the analysis to a tool you can't explain or verify, and you put both of those at risk. Keep the opinion your own, with the tool in a supporting role you can account for, and you don't.
A practical checklist for using AI in an expert report
- Stay the author. The opinions, findings, and conclusions are yours. A tool can organize and format what you supplied; it should never originate a fact, a number, an opinion, or a citation.
- Verify every citation and figure against the source, by hand. The fastest way to lose a declaration is to let a tool insert a reference you didn't check.
- Don't let a tool reach past your evidence. If a sentence can't be tied to something you actually provided, treat it as a gap to resolve, not text to keep.
- Keep a methodology record. Note which tool and version you used, on which sections, and what you gave it, at the time and not reconstructed after a motion to compel.
- Be ready to produce it. Assume your AI interactions could be discoverable as methodology, and keep them in a form you'd be comfortable handing to opposing counsel.
- Tell retaining counsel. Disclosure questions are easier to answer before a deposition than during one.
Where this is heading
As of mid-2026 there is still no uniform, expert-specific rule mandating AI disclosure; the obligation is being shaped case by case. But the direction is hard to miss. The experts who get hurt are the ones who let a tool do the thinking and couldn't account for it. The experts who are fine kept their own judgment in the chair and could show their work. The safe posture is the same one good experts have always had. Own your methodology, and extend that habit to a new tool you should expect to explain.
How we think about this
This is the exact problem Disclosed. is built around. It structures your own findings into a Rule 26(a)(2)(B) report, refuses to add a fact or citation that isn't already in your evidence, flags anything ungrounded for your review, and keeps an automatic, tamper-evident AI-disclosure record of how the tool was used. That is the kind of methodology record these cases are starting to ask for. You remain the preparing and signing expert; admissibility is, as always, the court's call.
Sources
- Kohls v. Ellison, No. 24-cv-3754, 2025 WL 66514 (D. Minn. Jan. 10, 2025).
- Concord Music Group, Inc. v. Anthropic PBC, No. 24-cv-03811, 2025 WL 1482734 (N.D. Cal. May 2025).
- Matter of Weber, 85 Misc. 3d 727 (N.Y. Sur. Ct. 2024).
- Ferlito v. Harbor Freight Tools USA, Inc., No. 20-CV-5615, 2025 WL 1181699 (E.D.N.Y. Apr. 23, 2025).
- Conservation Law Foundation v. Shell Oil Co., No. 3:21-cv-00933 (D. Conn.) (magistrate order on AI prompts, under Rule 72(a) review).
- Fed. R. Civ. P. 26(a)(2)(B); Fed. R. Evid. 702 (as amended Dec. 1, 2023).
General information, not legal advice. Disclosed. is a software company, not a law firm. Case descriptions are summaries; Conservation Law Foundation v. Shell in particular is a non-final order under review and may change. Consult counsel about your own matter, and verify any authority against the official reporter before relying on it.