Show Your Work: courts start ordering expert witnesses to produce their AI prompts

Together with lawyer misusing generative AI, expert witnesses are now increasingly finding out that the use of generative AI exposes them to more potential criticism than they might have thought in the moment when they sought AI's help. Experts are using tools like ChatGPT and Claude to sort discovery documents, survey literature, and draft sections (or in some cases significant parts) of their reports. Courts in a number of jurisdictions are now exploring whether the record of that process and the prompts themselves form part of the expert's discoverable methodology, or whether it stays behind the curtain as a private drafting aid.

Connecticut sets an early precedent

In Conservation Law Foundation v. Shell Oil Co. (D. Conn., 22 May 2026), an expert in a dispute used OpenAI's GPT-4o, through a private Microsoft Azure server, to filter the defendant's discovery documents into a working subset as part of preparing a report. The Defendant sought to compel production of the prompts used by the expert as part of the usual process whereby an expert must disclose the methodology used to reach their opinion. On 18 May 2026, Magistrate Judge Thomas O. Farrish granted the motion, in what appears to be the first US federal ruling requiring an expert to hand over generative AI prompts. The court rejected each of Plaintiff's three objections which had been raised:

1. that the prompts fell outside the scope of discovery,

2. that they were shielded by the parties' own agreement under Connecticut's Rule 29; and

3. that no separate prompts existed because the expert had only used search terms.

An expert witness's methodology is fair ground for discovery, and under the facts of this case, the process by which Dr Oreskes culled down the defendants' document production into a subset to be worked with is an aspect of that methodology.

Judge Thomas O. Farrish, Magistrate Judge, US District Court for the District of Connecticut

The Plaintiff has since filed an objection, and the District Court has stayed the order for production pending review.

Grinding away at an AI report?

The most recent, and most vivid, illustration of why expert use of AI is incredibly relevant arrived this week. In ongoing litigation over a 2020 explosion at Watson Grinding in Houston, which killed three people and destroyed around 200 homes, the defendant 3M retained an expert who prepared a report. Opposing counsel spotted an AI-generated exhibit comparing the expert's report with another expert's report during discovery and demanded the underlying prompts used in connection with the 3M expert's report.

What followed was 350 pages of ChatGPT conversations in which 3M's expert asked the tool to help him "create an exceptional expert witness report defending the standard of care at 3M" and to "show how 3M is 0% at fault for the explosion."  Astonishingly, it is reported that roughly 90 percent of the report filed with the court tracked ChatGPT's own output and revisions.

Plaintiffs' counsel called the defendant's own expert as his witness once he saw the prompts.

They hired him for the sole purpose of changing the outcome of the case. They hired him and he used ChatGPT to write these reports, so really, ChatGPT was the expert in the case. There's just no question about that.

Will Moye, plaintiffs' counsel, in comments to 404 Media

The transcript of the expert's cross-examination is a sobering reminder of where no expert witness would ever want to be, as the expert was educated that his entire conversation with ChatGPT was public and so was admissible in court:-

The prompts alone in the case are not something any trial lawyer would want to see produced by an expert:-

The expert was engaged at a cost of around US$90,000, and the jury awarded the plaintiffs more than US$61 million, apportioning 30 percent of the liability to 3M.

Expert Witness Duties

Many jurisdictions are publishing guidance for experts, for example in Australia, the Federal Court's Use of Generative Artificial Intelligence Practice Note (GPN-AI), which took effect on 16 April 2026, already applies to witnesses and other third parties including those who are required under subpoena or other orders to produce documents. It requires disclosure wherever Gen AI is used to summarise or analyse information upon which a witness relies to make a statement of fact or express an opinion, and confirms that an expert report should contain that expert's own opinion and process of reasoning, to be read together with the Expert Evidence Practice Note (GPN-EXPT).

The disclosure obligation sits alongside, not instead of, ordinary discovery and subpoena powers, which the ongoing case law development suggests could reach an expert's prompts in an Australian proceeding in much the same way.

In the Cayman Islands, guidance was given in 2025 to parties in Bradley & Chuang v Linda Frye-Chaikin [2025] CIGC (Civ) 5 where Justice Jalil Asif KC said:-

Users of AI tools must take personal responsibility for the accuracy of material produced, and be prepared to face personal consequences, including the possibility of wasted costs orders, if the work product that they put forward to the Court is not accurate.

While this comment was about hallucinations in submissions, the same holds true for an expert report. How AI is used, and the importance of upfront disclosure, is essential for any expert.

What now

Lawyers instructing expert witnesses should be extremely cautious around AI if they want to ensure their experts are not put though the grinder by:-

• ensuring expert instructions are clear as to the use (or non-use) of AI and warn about the loss of privilege which may occur if AI tools are used;

• asking experts directly, before a report is finalised, whether and how generative AI tools were used at any stage, including document review and literature searches,

• reviewing the material, and considering whether the expert's use of AI has waived privilege over materials provided to them;

• retaining and preserving any prompts, chat logs, and AI-generated drafts as part of the expert's working file, on the assumption that they may later need to be disclosed or produced under subpoena;

• ensuring any AI-assisted work is disclosed in the report itself, rather than left for opposing counsel to uncover;

• treating AI output as a starting point for the expert's own reasoning, not a substitute for it; and

• reviewing any discovery agreements to ensure they state explicitly whether AI prompts and queries fall within or outside "expert notes," rather than leaving that question to be litigated after the fact.

  
  

The courts are working out case by case how to deal with AI disclosure, with the above Connecticut ruling still under challenge and no appellate guidance yet in place in the US.

It remains clear, as it always has, that an expert witness who cannot stand by their methodology and expertise in the witness box will not be an asset to the court, but rather a liability.

By Michael Bacina