Using AI on an expert report without a Daubert problem: what's safe, what's discoverable

An expert's AI prompts have already been ordered produced in federal discovery, and a proposed rule would put AI output under the same scrutiny as expert testimony. The good news: the line that keeps AI use defensible is the same one that already governs your opinions.

By CitePage · Updated June 2026 · ~6 min read

Experts are using AI to get through record bundles — that ship has sailed. The open question is narrower and more practical: can you use it without handing opposing counsel a new line of attack? The honest answer is yes, but only if you understand two things that have changed in the last year — what's now discoverable, and where the rules are heading — and then keep your AI use on the right side of one bright line.

What a real court has already done

In Conservation Law Foundation v. Shell Oil (D. Conn.), a magistrate judge on May 18, 2026 ordered the plaintiff to produce the generative-AI prompts its expert and her research assistant used to filter a large document production down to a working subset. The court's reasoning is the part to internalize: "An expert witness's methodology is fair ground for discovery," and the process of culling the records "is an aspect of that methodology." A stipulation shielding an expert's "notes, drafts, or communications" was held not clear enough to cover the prompts.

Read it precisely. The order was stayed pending the plaintiff's objection, so it is not the final word, and one magistrate's discovery ruling does not bind other courts. But the direction is unmistakable: when AI shapes what an expert looked at or how, courts are willing to treat the prompts and inputs as discoverable methodology — not protected work product by default.

The bright line: organizing facts vs. forming opinions

The reason an AI prompt becomes a problem in discovery is that it can blur into methodology — into the reasoning behind the opinion. So the safest architecture keeps AI strictly on the facts-organizing side of the line and keeps you, the expert, as the sole source of every opinion and inference. The same line that already separates a record fact from an expert determination is the one that keeps AI use defensible.

✓ Generally safe (organizing facts)

Reading every page so nothing is skimmed under deadline
Extracting dates, providers, and findings into a structured timeline
Attaching a source-page citation to each entry
De-duplicating an event recorded across documents
Surfacing contradictions and gaps for your review
Producing an exportable record of what was done

✗ Keep it yours (forming opinions)

Deciding whether a finding is traumatic or degenerative
Apportioning causation between accident and pre-existing disease
Resolving a record-versus-deposition contradiction
Drafting the reasoning that carries your name
Writing any sentence of your actual opinion
Standing in for your verification of the source page

When AI only organizes facts and you author every conclusion, a discovery demand for your "AI methodology" turns out to be far less threatening — because the AI never was your methodology for forming opinions. It assembled a cited, verifiable chronology; you reasoned from it. That distinction is the whole game.

What makes it discoverable-ready instead of discovery-exposed

The expert who pasted records into a public chatbot has the worst of both worlds: no organized, exportable record of what they did, and a potential privilege problem on top. If asked to produce their "AI methodology," they're reconstructing it after the fact. The defensible posture is the opposite — assume it may be discoverable and build the record on purpose:

Keep a clean audit trail. What was processed, what was extracted, what each entry cites, and every place you edited or overrode the draft — exportable, so disclosure is a download, not a scramble.
Cite every fact to a page. A page-cited chronology is the visible path from record to reasoning that a Daubert reliability inquiry asks for. When each fact traces to a source page, your "basis and reasons" are auditable end to end.
Use private, no-train tooling. Privileged material does not belong in a public model that may retain or train on it. An isolated tenant that never trains on your data is safer for privilege than a consumer chatbot.
Sign every word of the opinion. The output that goes to the court is yours, written by you, defensible as your work — with AI's role confined to the facts underneath.

An honest line on where the rules are heading. Proposed Federal Rule of Evidence 707 (projected default effective Dec 1, 2027) would extend Daubert-style reliability scrutiny to machine-generated evidence offered without a sponsoring expert, and the UK's proposed PD35 §3.3 AI-use declaration is a consultation-stage amendment — neither is binding law today. They tell you the direction of travel; they are not current mandates. Disclose under the rules as they actually stand in your jurisdiction, and treat CLF v. Shell as a persuasive signal, not settled nationwide law.

A short defensibility checklist

AI organized facts; you formed every opinion and inference.
Every fact in the chronology cites a source page you verified.
Contradictions and gaps were surfaced for your judgment, not silently resolved.
There is an exportable audit trail of prompts, sources, and your edits.
Privileged records stayed in a private, no-train environment.
The report is written and signed by you.

Get AI leverage with a discovery-ready record.

CitePage returns a page-cited chronology — every fact stamped to its source, conflicts flagged for your judgment — with an exportable audit trail built for privileged material and never used to train AI. You author every opinion. Your first chronology is free.

Try your next case free →