The prompt is the evidence: a 3M expert's ChatGPT log became 350 pages of discovery

An expert witness retained by 3M in litigation over the January 2020 explosion at Watson Grinding in Houston used ChatGPT to write significant portions of his expert report, and instructed the model to produce something that would "show how 3M is 0% at fault for the explosion at Watson Grinding," according to 404 Media ^[1][2]. The report is not the interesting artefact here; the prompt log is, because it became discoverable ^[3].

The underlying case is not marginal. The explosion killed three people and destroyed roughly 200 homes ^[4], and the U.S. Chemical Safety and Hazard Investigation Board attributed it to a "degraded and poorly crimped rubber welding hose" that leaked a flammable gas ^[5]. Dozens of homeowners sued 3M and Watson Grinding, alleging among other things that 3M did not properly service the facility's gas detection system ^[6]. 404 Media reports that hundreds of millions of dollars in total liability are at stake ^[7].

The discovery chain is the part operators should read twice. 3M hired Josh Autenrieth of Knighthawk Engineering to prepare an expert report ^[8]. During discovery, plaintiffs' attorney Will Moye found a five-page document titled "Citation Overlay" that appeared to have been generated by AI, recognised it as ChatGPT output, and demanded every prompt from 3M's lawyers ^[9]. The deposition was paused for three hours while they were collected, and Moye received 350 pages of ChatGPT conversations, including public links to Autenrieth's full conversations ^[10][11]. One anomalous five-page attachment unspooled into the entire drafting history.

That history includes Autenrieth telling ChatGPT he was "being retained as a professional expert witness by 3M" for the Watson Grinding proceedings, that he needed a report to defend 3M's standard of care, and that it had to counter an opposing witness's "outlandish and false claims" ^[12][13]. He attached hundreds of court records and asked the model to find standards violations ^[14]. ChatGPT produced a roughly 30-page report containing the sentence "From a technical and standard-of-care standpoint, 3M is 0% responsible for the January 24, 2020 explosion" ^[15]. The line did not reach the filed report, because Autenrieth asked ChatGPT to "review this as the opposing council," and the model warned that "0% responsible" is "an easy target" and among several "phrases [that] let opposing counsel paint you as an advocate rather than an expert" ^[16][17]. The sanitised output survived. The instruction that produced it also survived, in the log.

There is a competence layer too. In a separate chat, Autenrieth uploaded an image of a gas detector and asked ChatGPT "what am I looking at?" and "what are model names of industrial gas detectors" ^[18]. Court transcripts suggest 3M paid Knighthawk roughly $90,000, at a stated rate of $475 per hour ^[19][20], which works out to about 189 billable hours ^[21].

What to watch: whether your own AI-assisted deliverables have a retention policy at all. Reports, risk assessments, incident narratives and post-mortems drafted with a chatbot leave a timestamped record of what the author asked for before they got it, and that record is a separate document from the deliverable. Treat prompt histories as you would draft memoranda: assume they are producible, assume the framing you used will be read aloud, and decide now who is allowed to type "show how we are 0% at fault" into anything that logs.