AI use under FRE 702 and Daubert: how a physician expert's process is tested
Answer in brief: courts don't test the AI; they test you. Under Rule 702 the party offering you must show your opinion rests on sufficient facts or data and a reliable method reliably applied. AI that organizes the record and that you checked rarely threatens that. AI that produced a fact, a number or a citation you didn't verify gives a Daubert motion its best argument.
Limits: We read every rule, ruling and product page on the date shown and did not test any product with patient records. Court summaries here are not legal advice; how your court treats AI use depends on the jurisdiction and the judge, so ask retaining counsel.
The rule a physician expert is judged by
Federal Rule of Evidence 702 lets a qualified expert give an opinion if the proponent shows the court that it is more likely than not that:
- (a) the expert's specialized knowledge will help the trier of fact;
- (b) the testimony is based on sufficient facts or data;
- (c) the testimony is the product of reliable principles and methods; and
- (d) the opinion reflects a reliable application of those principles and methods to the facts of the case.
The "more likely than not" wording arrived with the amendment effective 2023-12-01. Its committee note says courts that treated the sufficiency of an expert's basis and the application of the method as questions of weight, for the jury, were applying the rule incorrectly (Rule 702 and notes, LII, read 2026-11-12). For you, that means the judge can look at how you built your opinion, including any AI step, before a jury hears it.
Element by element: where AI helps and where it hurts
| Element | What a challenger asks | How AI creates the problem | What protects you |
|---|---|---|---|
| (a) Helpful knowledge | Is this opinion medicine, or something a layperson or a chatbot could produce? | Report language that reads like generic AI text makes your specialty look interchangeable | Reasoning in your own clinical terms, tied to your training and practice |
| (b) Sufficient facts or data | Did you review the whole record? What did you rely on? | A tool that dropped pages, merged patients or missed a late entry leaves a gap you don't know about | A page-cited chronology, a records-reviewed list with page counts, and spot-checks of the lines you rely on |
| (c) Reliable method | What method did you use, and could another physician repeat it? | "I asked the AI" is not a method; outputs can differ on each run, as the Weber court found | Describe the method as your clinical review; describe AI as a sorting and search aid you checked |
| (d) Reliable application | How did you get from these records to this conclusion? | A conclusion drafted by AI leaves an analytical gap you can't fill on the stand | Reasoning written by you, each step pointing to record pages and named sources |
The 5 Daubert factors, applied to an AI-assisted opinion
Daubert v. Merrell Dow (1993) gave trial judges a gatekeeping role and a non-exclusive list of factors. Kumho Tire (1999) held that the gatekeeping applies to all expert testimony, not only testimony based in science, so a physician's experience-based opinion is covered too. The 2000 committee note to Rule 702 lists the 5 factors; here is how each meets AI use.
- Can the method be tested? A chronology can: anyone can open page 412 and see whether the entry is there. A chatbot's reasoning can't be re-run to the same answer.
- Peer review and publication. Your clinical method has it. "AI-assisted record review" for litigation has little published validation. In Weber, the expert told the court AI use was generally accepted in his field and could not name a single publication saying so.
- Known or potential error rate. Don't quote a vendor's accuracy claim at deposition; you can't defend a number you didn't measure. Say what you checked and what you found.
- Standards and controls. Your controls are the ones you can show: a log of AI steps, a sample of lines checked against source pages, and a rule that nothing goes in the report unread.
- General acceptance. Literature search is ordinary clinical practice, which helps Zone 2 work, and a page-cited index is easy to show a court. We found no field that accepts an AI-written causation opinion.
The same 2000 note lists more factors courts use. 2 bite here. First, whether the expert used "the same level of intellectual rigor" as in practice (from Kumho): if you would never trust an unchecked chatbot answer at the bedside, a court will ask why you trusted it here. Second, whether there is "too great an analytical gap between the data and the opinion" (General Electric v. Joiner, 1997). An opinion supported only by the expert's say-so, which Joiner called ipse dixit, fails; an AI paragraph you can't explain is the modern version of that gap.
Frye states and other state tests
Not every case is in federal court, and not every state uses Daubert. New York still applies the Frye general-acceptance test, which is how Matter of Weber framed its AI holding: the court said AI-generated evidence should face a Frye hearing before admission and that counsel has an affirmative duty to disclose AI use first (Matter of Weber, 2024, read 2026-11-12). Under Frye, the question becomes whether the relevant field accepts the method. That favors page-cited record tools and literature search, and works against anything that looks like AI forming the opinion. Ask retaining counsel which test the forum uses before you plan your workflow, because the answer changes how much you document.
Proposed Rule 707: where it stands on 2026-11-12
Rule 707 would make machine output offered without an expert witness meet the Rule 702 test. It targets parties who put AI output in front of a jury with no human expert sponsoring it. It does not govern your testimony, but its path tells you how cautious the rule-makers are.
On 2026-05-07 the Advisory Committee on Evidence Rules voted not to advance the original draft or any of 3 redrafts. Per Redgrave, the leading redraft would usually require an expert witness to explain how an AI system produced the evidence. If that version returns, the message for experts is the same as today's: the human who testifies has to understand the machine's part.
What a Daubert motion aimed at your AI use looks like
From the rulings so far, a motion targeting an expert's AI use makes some mix of these arguments. Pre-empt each in your report and your log.
| Argument | Built on | Your answer, if true |
|---|---|---|
| The opinion isn't the expert's own | Rule 26's "prepared and signed by the witness"; the 3M deposition | "I wrote the opinions and their reasoning. I used software to sort and index the records, and I checked what I relied on." |
| The basis includes invented facts or sources | Kohls v. Ellison | "Every source in my reference list is one I read. Here are the record pages for each fact." |
| The method can't be explained or repeated | Matter of Weber | "My method is clinical review of the record. Here is the list of records reviewed and the chronology with page citations." |
| Undisclosed AI use is a discovery violation | Villanueva, CLF v. Shell | "I told counsel which tools I used before I started, and the log is available." |
| The prompts show a predetermined conclusion | The 3M prompt "Show how 3M is 0% at fault" | Only possible if your prompts were neutral. Write them that way from day 1. |
Next: the deposition questions those arguments turn into, and the disclosure wording that heads them off.