17
JulGuidance for our expert witnesses and attorney clients
Artificial intelligence has arrived in expert witness work, whether we invited it or not. Many of the experts we place already use tools like ChatGPT, Claude, or Copilot to organize research, summarize literature, or polish drafts. Used carefully, these tools genuinely improve the work. Used carelessly, they can end your involvement in a case, and judges have now said so, repeatedly and in writing.
This is not a hypothetical risk. A legal database maintained by researcher Damien Charlotin now tracks more than 1,700 court decisions worldwide in which a judge found that someone relied on AI-hallucinated material, most often fake citations, but also fabricated quotes and research that does not say what it was claimed to say. A growing share of those decisions involve expert witnesses specifically. What follows are the rulings every testifying expert should know, the practices that keep your reports safe, and, because we know some of you are wondering, what to do if AI already touched something you filed.
This is the landmark case, and the irony is hard to overstate. In a First Amendment challenge to Minnesota’s deepfake law, the state submitted an expert declaration from a Stanford professor who is a leading scholar on AI and misinformation. He used GPT-4o during drafting and left placeholders in his text, intending to add real citations later. The AI filled those placeholders with citations to two academic articles that do not exist and misattributed the authors of a third. He signed the declaration under penalty of perjury without catching any of it.
Judge Laura Provinzino found his explanation plausible and honest — and excluded his testimony anyway, writing that citing fake, AI-generated sources in a sworn declaration “shatters his credibility with this Court.” She refused to let the state file a corrected version. The expert’s substantive opinions may have been entirely sound. It did not matter. He was out, and the Eighth Circuit later affirmed.
Two lessons stand out. Good faith does not save you: the court believed the errors were innocent and excluded the testimony anyway. And the attorneys were warned too — Rule 11 imposes a “personal, nondelegable responsibility” to verify everything filed, and the judge suggested lawyers may now need to ask their experts directly whether they used AI and what they did to verify the output. Expect that question on your next engagement. Kohls v. Ellison, No. 24-cv-3754, 2025 WL 66514 (D. Minn. Jan. 10, 2025) — read the opinion.
In a New York trust accounting dispute, a fiduciary-services expert admitted he used Microsoft Copilot to cross-check his damages calculations. On the stand, he could not recall what prompt he entered, could not identify what sources the tool relied on, and could not explain how it reached its numbers. He insisted AI use was “generally accepted” in his field but could not name a single publication supporting that claim.
Then Surrogate Jonathan Schopf did something memorable: he ran the expert’s own investment calculation through Copilot on three separate court computers and got three different answers — $949,070.97, $948,209.63, and roughly $951,000 — none matching the expert’s figure. When court staff asked Copilot whether its calculations were reliable enough for court, the chatbot itself answered that its output should be verified by experts first. The court held that AI-assisted evidence in New York must be affirmatively disclosed by counsel and may require a Frye hearing before admission. If any part of your analysis involves AI, you must be able to explain what you asked, what it did, and how you verified the result — under cross-examination. “The computer said so” is not a methodology. Matter of Weber, 220 N.Y.S.3d 620 (N.Y. Sur. Ct. 2024), read the opinion, including its section titled “Use of Artificial Intelligence.”
In this copyright case, an expert declaration cited a real article — but with a fabricated title and a list of co-authors who had never worked together, an error introduced when counsel ran the citation through an AI tool for formatting. The court called it “a plain and simple AI hallucination,” faulted the lawyers for missing it in their manual citation check, struck the affected paragraph, and noted on the record that the episode undermined the expert’s overall credibility. One citation. That is all it took. Concord Music Grp., Inc. v. Anthropic PBC, No. 24-cv-03811-EKL, 2025 WL 1482734 (N.D. Cal. May 23, 2025) — read the opinion.
This is the newest frontier. In May 2026, a federal magistrate judge in Connecticut ordered a party to produce the actual generative AI prompts its expert used to filter documents while preparing her report — apparently the first federal decision treating an expert’s AI prompts as discoverable methodology. The court’s logic: if the prompts shaped which documents the expert ever saw, they shaped the foundation of her opinion, and the other side is entitled to test that. The order is currently stayed pending objection, so the final word is not in, but the direction of travel is unmistakable. Assume that anything you type into an AI tool in connection with a case may one day be an exhibit. Prompt accordingly, and keep records. Conservation Law Found., Inc. v. Shell Oil Co., No. 3:21-cv-00933 (D. Conn. May 18, 2026) — analysis from Mayer Brown.
It is not all cautionary tales. In this product liability case, Judge Gary Brown refused to exclude a design expert who had used ChatGPT — because of how he used it. The expert formed his opinions first, grounded in decades of experience joining dissimilar materials, wrote his report himself, and only afterward ran a query to see whether the AI agreed. He testified he was “quite amazed” that it confirmed what he had already concluded. The court found no reliability problem, and it quoted Kohls for the governing principle: experts may use AI for research; trouble arises only when they abdicate independent judgment in favor of ready-made AI answers.
That is the model. Your judgment first, AI as a verification and efficiency tool second — never the other way around. Ferlito v. Harbor Freight Tools USA, Inc., No. 2:20-cv-05615, 2025 WL 1181699 (E.D.N.Y. Apr. 23, 2025) — read the order.
The courts are not waiting for this to sort itself out. The U.S. Judicial Conference’s Advisory Committee on Evidence Rules has proposed a new Federal Rule of Evidence 707, which would subject machine-generated evidence to the same reliability standards as expert testimony, alongside a proposed amendment to Rule 901 addressing AI-fabricated evidence; the public comment period closed in February 2026. Judges around the country have issued standing orders requiring disclosure or certification of AI use in filings. And after Kohls, more retaining attorneys will ask you directly — possibly in writing — whether AI touched your report and how you verified it. Answer honestly. Across all of these cases, courts have been consistently harsher on concealment than on disclosed, corrected mistakes.
1. Write your opinions yourself. Your report must reflect your independent professional judgment. AI can help you organize, summarize, and check — it cannot form your opinion for you. One federal commentator put it bluntly: experts who recite AI output reduce themselves to parrots.
2. Verify every citation by pulling the actual source. Not by asking the AI whether it is real. Open the journal article, the case, the guideline. Confirm it exists, confirm the authors, confirm it says what you are citing it for. Hallucinated citations come in three flavors: sources that do not exist, wrong citations to real sources, and real sources that do not support the point.
3. Double-check, then triple-check. Have a second set of eyes — a colleague, an assistant, or retaining counsel — independently verify citations and factual claims before anything is signed. You are signing under penalty of perjury.
4. Never put confidential case materials into public AI tools. Medical records, discovery documents, and privileged communications do not belong in consumer chatbots. If your workflow requires AI review of case documents, raise it with retaining counsel first and use secure, appropriate tools.
5. Keep records of your AI use. What tool, what version, what you asked, and what you did to verify the output. After the Shell ruling, assume prompts may be discoverable — and a documented verification process is a strength on cross, not a weakness.
6. Disclose when asked, and consider disclosing when not asked. If counsel asks whether AI was involved, the answer must be complete and accurate. In New York, after Weber, disclosure may be required outright. Coordinated candor between expert and attorney is what separates a correctable issue from a credibility crisis.
7. Be ready to explain your methodology on the stand. If you cannot explain how a number, conclusion, or source made it into your report, it should not be in your report.
If you are reading this and realizing AI played a role in something you have already submitted, take a breath. Using AI is not misconduct, and courts have said so explicitly. Every exclusion described above came from unverified AI output, not from AI use itself.
Review the work now. Go back through the report and independently verify every citation, quote, statistic, and factual claim against the actual sources. If everything checks out, you are in good shape — and you have just built the verification record you should keep going forward.
If you find an error, tell your retaining attorney immediately. Do not wait for opposing counsel or the court to find it. The case law is consistent: judges have repeatedly shown more leniency to experts and attorneys who came forward promptly, took responsibility, and corrected the record than to those who denied or minimized. Prompt, honest correction is damage control; discovered concealment is a career problem. Your attorney will know the right procedural vehicle for a correction in your jurisdiction.
Don’t quietly fix it yourself. A corrected report or amended declaration is a legal filing decision that belongs to counsel. Your job is complete candor with the attorney who retained you.
Going forward, apply the seven practices above on every engagement. One clean verification process, applied consistently, ends this risk entirely.
Courts are not anti-AI. Judge after judge has acknowledged these tools’ legitimate potential — the Kohls court itself said AI may transform legal practice for the better. What no court will tolerate is an expert whose sworn opinion contains things the expert never checked. The technology did not change the standard; it created a new way to fail it. Your expertise, your judgment, and your verification are what attorneys pay for and what courts rely on. AI can make you faster. Only you can make the work true.
If you have questions about AI use on a current or upcoming engagement, reach out to us. We would much rather have that conversation before a report is served than after.
— The Homestead Medical Experts Team