The 3M ChatGPT Deposition That Should Change How Your Company Stores AI Conversations
A single expert witness's prompt history produced 350 pages of previously unseen material mid-deposition. For every organisation using AI in serious work, that is a records problem, not just a legal one.
Key points
- An engineering expert retained by 3M used ChatGPT while building his analysis in the Watson Grinding explosion lawsuit, and a prompt telling the system to "show how 3M is 0% at fault" later surfaced in litigation.
- After plaintiffs' attorney Will Moye questioned the expert about AI-generated material, the deposition went off the record and more than 350 pages of previously unproduced ChatGPT conversation history were handed over roughly three hours later.
- The American Bar Association has identified AI chat histories as an emerging source of discovery material, meaning courts can compel companies to produce them.
- Most organisations have policies about what employees type into AI tools, but few have decided how long those conversation records should be kept, who owns them, or what triggers a legal hold.
- The right level of record-keeping depends on how consequential the AI-assisted work is: helping rewrite an email requires different treatment from supporting a safety analysis or an employment decision.
A deposition in an explosion lawsuit produced one of the clearest illustrations of what AI governance actually means in practice, and it had nothing to do with a data breach or a hacked system.
The case involves Watson Grinding, a Texas industrial facility that exploded in 2020. An engineering expert retained by 3M used ChatGPT (OpenAI's AI assistant, which can answer questions and draft analysis) while developing his work. Among the conversations that later came to light was a prompt asking the system to "show how 3M is 0% at fault." No public finding establishes that 3M told the expert to write that prompt. But the prompt existed, and once plaintiffs' attorney Will Moye spotted material that appeared AI-generated, he demanded the underlying conversation history during the expert's deposition.
The deposition paused. About three hours later, more than 350 pages of ChatGPT material that had never been handed over appeared.
Why does a deposition story matter to ordinary businesses?
It matters because the same thing can happen in any workplace where AI helps produce a consequential decision.
Consider a procurement analyst who uses an AI assistant to build the strongest-sounding case for a supplier management already prefers. A manager who asks AI to help document an employment decision after it's already made. An auditor who keeps refining prompts until a control weakness (a gap in a company's safety checks) sounds less serious on paper. In each case the AI performs exactly as designed, but the conversation history records the assumptions, the rejected phrasings and the preferred outcomes that never make it into the final document.
Courtrooms can demand that history. Regulators can too. So, in some circumstances, can an internal investigation. We covered the parallel risk in Slack messages on 18 August in "Your Team's Slack Messages During a Breach Could Cost More Than the Breach Itself": the principle is the same. The channel changes; the discovery exposure doesn't.
| Aspect of the AI interaction | Who typically governs it today |
|---|---|
| What employees type in (the "prompt") | IT acceptable-use policy |
| The AI's output | The employee who uses it |
| The full conversation history | Usually nobody, formally |
| Retention period | Often the AI platform's default |
| Legal-hold obligation | Rarely defined before a dispute |
So what should companies actually do?
Saving everything isn't the answer. It creates its own privacy and security risks, and mountains of stored chat logs don't automatically mean better oversight.
The practical approach is to match record-keeping to the seriousness of the work. An employee asking AI to tidy up an email needs no special process. An engineer using AI to support a safety analysis, or an executive relying on AI output to inform a major financial decision, is in different territory. For those higher-stakes uses, organisations need enough of the process on record to reconstruct what happened: which AI system was used, the key prompts and outputs, and evidence that a qualified human actually reviewed the result.
As first reported by CSO Online, the governance gap here is rarely about technology. It's about ownership. AI teams decide what the tools do. Legal teams learn how a platform retains data only when a lawsuit arrives. Records-management teams are often not in the conversation at all. Someone with defined authority needs to decide in advance what is kept, what expires, what triggers a legal hold (a formal instruction to preserve records relevant to a dispute), and who is accountable when a regulator or court asks.
The Watson Grinding case is a useful reminder that the prompt itself can become part of the official record of a decision. Most companies have spent years telling staff not to paste sensitive data into AI tools. Fewer have asked what happens to the conversation after the answer appears on screen. That second question is now overdue.
Common questions
Do ordinary employees need to worry about their AI chat logs being read in court?
For most routine work, no. But if you use AI to help produce analysis or decisions that could later face legal or regulatory scrutiny, the conversation history may be discoverable, meaning a court or regulator could require your employer to hand it over.
Does this mean companies should ban AI from serious work?
Not necessarily. It means they need a clear policy: document which AI tool was used, retain the key prompts and outputs for high-stakes tasks, and make sure a qualified person reviews and signs off on the final result rather than simply accepting the AI's draft.
