The use of artificial intelligence in expert evidence has moved from a theoretical concern to a courtroom reality in an interesting Texas explosion case in which an engineer retained by 3M acknowledged that ChatGPT generated between 85 and 90 per cent of his expert witness report.
The episode has raised uncomfortable questions about independence, methodology, disclosure and the line between legitimate technological assistance and the delegation of expert opinion to an artificial intelligence system.
The litigation arose from the January 2020 explosion at the Watson Grinding and Manufacturing facility in Houston. The blast killed three people, injured others and caused extensive damage to surrounding property. Plaintiffs alleged, among other matters, that 3M had failed properly to service the facility’s gas detection system. The resulting litigation ultimately produced a jury award of more than $61 million, with 30 per cent of responsibility apportioned to 3M.
The Engineering Firm was Paid Approximately $90,000, with Autenrieth Billing at $475 per Hour.
To defend its position, 3M retained Josh Autenrieth of Knighthawk Engineering as an expert witness. Court-related records indicate that the engineering firm was paid approximately $90,000, with Autenrieth billing at $475 per hour. The report became an inherently significant issue when plaintiffs’ counsel Will Moye discovered a five-page document entitled “Citation Overlay” that appeared to have been generated using ChatGPT.
What followed was potentially more damaging to the defence than the original discovery of AI assistance.
Moye sought production of the prompts and conversations Autenrieth had entered into ChatGPT. The production reportedly ran to approximately 350 pages. The material provided an unusually detailed record of how the expert’s report had been developed. (Incident Database)
Among the ai prompts was a request to “create an exceptional expert witness report defending the standard of care at 3M” and another asking the chatbot to “show how 3M is 0% at fault for the explosion at Watson Grinding.”
For an expert witness, the wording is inherently significant An expert’s overriding duty is not to act as an advocate for the party retaining him, but to provide objective, reliable and properly reasoned opinion evidence within his or her area of expertise. A prompt instructing an AI system to establish a predetermined allocation of liability creates an obvious forensic issue: was the expert independently forming an opinion from the evidence, or was he using AI to construct arguments supporting a conclusion he had already reached? The ChatGPT record apparently provided plaintiffs’ counsel with an opportunity to explore precisely that question.
The chatbot initially produced a report containing the assertion that, from a technical and standard-of-care perspective, 3M was “0% responsible” for the explosion. That wording was subsequently removed after Autenrieth asked the system to review the report from the perspective of opposing counsel. ChatGPT reportedly identified the “0% responsible” language as an easy target because it could make the expert appear to be acting as an advocate rather than an independent witness.
The irony was considerable. The AI system had effectively identified a potential weakness in the expert’s methodology before opposing counsel had to do so. At trial, Autenrieth defended his professional qualifications and experience. He said: “I’ve got a body of work and 20-plus years of experience in the industry.”
He characterised ChatGPT as a drafting aid rather than the source of his substantive opinions, explaining that “my opinions were put in there, and AI helped me to draft a straw man to build off of.”
He further maintained that he had supplied information and opinions before the AI generated material and that, where the output did not accord with his views, he altered it.
At Odds With Recorded AI Involvement
That explanation, however, collided with the extent of the recorded AI involvement.
During examination, plaintiffs’ counsel and Autenrieth agreed that approximately “90 to 85 percent” of the submitted report was ChatGPT-generated. The dispute became even more heated because the prompts were not limited to grammatical editing or formatting. Autenrieth had asked ChatGPT to analyse technical material, consider photographs of gas-detection equipment, identify potential weaknesses in his report and assess his own suitability to act as an expert witness.
At one point, he asked the system to identify the principal weaknesses that opposing counsel might attack and how he could defend them. He also asked ChatGPT to grade his report. It reportedly gave the document a score of 97 out of 100.
The once inherently professional relationship between the expert and the retaining party subsequently deteriorated. According to reporting on the proceedings, 3M sought to have Autenrieth disqualified after the extent of his AI use became apparent. Plaintiffs’ counsel took the unusual step of calling Autenrieth as a witness rather than simply seeking to exclude his evidence.
For expert witnesses, that development may be the most instructive feature of the case.
An expert who loses the confidence of the party that retained him can become an awkward witness. An expert whose own working papers contain evidence apparently inconsistent with his sworn testimony faces an even greater forensic challenge. Here, the ChatGPT conversations supplied opposing counsel with a chronological record against which Autenrieth’s methodology and credibility could be tested. The case also illustrates the emerging discovery implications of generative AI.
It would be premature to describe the litigation as having established a universal rule that every AI conversation is automatically discoverable. Discovery is governed by the applicable procedural rules, privilege doctrines and the particular circumstances of each case. Nevertheless, the episode demonstrates that AI prompts and chat histories may become discoverable where they bear directly upon the formation of an expert’s opinions, the preparation of a report or the basis for testimony.
For lawyers instructing experts, the practical lesson is straightforward. AI cannot be treated as an invisible drafting assistant if it is materially involved in forming the substance of an expert opinion. Counsel should establish at the outset what AI tools may be used, what information may be uploaded, whether prompts and outputs must be preserved, and how any AI-generated material will be independently verified.
For experts, the fundamental obligation remains unchanged. The report must be their opinion, supported by their expertise, analysis and examination of the evidence. The Watson Grinding litigation does not establish that an expert can never use ChatGPT. It demonstrates something more consequential: the case demonstrates how an expert’s AI interactions can become relevant to discovery and cross-examination.
In a profession built upon independence and credibility, that trail can become almost as important as the report itself.
- Dangers of AI Misinformation – Maryland State Bar
- Join Witness Directory as a member
