Expert Witness Asked ChatGPT To Show 0% Fault—The Wrong Way For Experts To Use AI

What The 3M Expert Witness Asked ChatGPT To Do

He told ChatGPT he was “being retained as a professional expert witness by 3M.” He uploaded his résumé and hundreds of case files. Then he gave it the outcome: “I need to show that 3M met the standard of care and why.”

Later that day, he got more specific. He asked it to “create an exceptional expert witness report defending the standard of care of 3M.” He asked it to defeat what the other side’s expert had said. And he asked it to “show how 3M is 0% at fault for the explosion at watson grinding and how my background and experience is well suited to render this professional opinion.” Then he typed: “Create the expert witness report draft now.”

The chat record shows a 16-minute, 57-second exchange that produced a draft report with 14 sections and his name at the top. The draft stated: “From a technical and standard-of-care standpoint, 3M is 0% responsible for the January 24, 2020 explosion.” That sentence was cut later. The reason came from the same model.

Other prompts went to the basics of the job. He uploaded a photo of a gas detector of the type at issue in the case and asked, “what am i looking at in terms of gas detection system.” He asked whether the “prosecution,” his word for the other side in a civil case, would attack his resume because he had never been an expert witness before. He had the model review his drafts. The last score it gave him was 97 out of 100. The night before his deposition, he asked ChatGPT, “What am I basing my standard of care on as an expert witness?”

ChatGPT Pushed Back Only When Told To Play Opposing Counsel

He asked the model to review the draft “as the opposing council on the side of trying to assign blame to the calibration technician and their company.” The answer found the problem. It said “0% responsible” was “an easy target.” It said an expert “should not sound like an advocate assigning legal fault.” It said zero percent “is a jury-allocation conclusion, not an engineering conclusion.” He cut the line from his report. The prompt that produced it stayed in the record.

That is the whole lesson, in two prompts. Same model. Same evidence. Asked to support a conclusion, it did. Asked to attack the conclusion, it found a weak spot.

At the trial in July, the plaintiffs called 3M’s own expert to the stand. Plaintiffs’ counsel Will Moye walked him through the record. The transcript shows what happened. At his deposition, he had said AI did not write his report. At trial, he said it “helped me to draft a straw man to build off of.” By the time of the deposition, he had billed 174 hours at $475 an hour. Moye said 85 to 90 percent of the 30-page report came from the model. Autenrieth would not agree to a number. Then he said, “If you say so.”

Moye asked why an unbiased witness would tell a chatbot to show zero fault. He said he had never been an expert witness before. Moye asked if he had used biased prompts to get a particular answer. “The prompts were biased because I had opinions,” he said. So if the prompts were his opinions, and the prompts were biased, were his opinions biased? “That means my opinions are already set.”

Why AI Can Tell Experts What They Want To Hear

AI chatbots such as ChatGPT run on models that are trained partly on human feedback. Researchers at Anthropic found that this feedback can reward answers that agree with the user more than answers that correct the user. Researchers call the result sycophancy, which means telling people what they want to hear. OpenAI rolled back an update to ChatGPT after it said the model had become “overly flattering or agreeable.” The training had leaned too hard on quick thumbs-up ratings.

A study in the journal Science tested 11 AI models. Across the models, they affirmed users’ actions 49 percent more often than people did, on average. People who got the flattering answers became more sure they were right. They also trusted the flattering model more.

A model does not always tell users what they want to hear. An expert who types the answer into the prompt makes it more likely. Give a model the documents plus the answer, and it can arrange the documents around the answer. None of this requires bad intent. An expert who asks a model to confirm a finding can get a confirmation, because that is what was asked for. The bias can slip in through the question, whether the expert meant it or not.

The Right Way For Expert Witnesses To Use AI

AI will have a place in expert work. It can summarize thousands of pages of records. It can check a calculation. It can help an expert who is better at the work than at the writing put the findings into plain language, so the facts do not get muddied by accident. The expert and counsel must decide where it fits under the engagement terms, court orders, and applicable rules. The opinion itself is never handed off. The order matters, and it runs like this.

• Frame the job as a question, never as the client’s answer.
• Form the opinion first, from the documents, the data, the site, the device, and the standards that govern the work, before the model sees a word of it.
• Then give the model the opposite job from the one it got in Houston: assume I am wrong; find the weakest link in my reasoning; tell me what an opposing expert with these same records would say. Prompts that attack push against the model’s pull toward agreement. Prompts that ask for approval ride along with it.
• Treat what comes back as a list of things to check against the evidence, never as the answer. Every fact, quote, calculation, and citation the model returns gets checked against the original record.
• Keep technical opinions separate from legal conclusions. An expert should distinguish technical causation and standard-of-care opinions from legal allocation of fault, which is generally resolved by the factfinder under the court’s instructions.
• Do not place confidential case files in an outside AI service without authorization. Have counsel assess preservation, disclosure, and discovery obligations under the applicable rules, court orders, protective orders, and engagement terms.
• Work as though prompts, uploads, outputs, and related metadata may be sought in discovery or examined at deposition. In this case, the prompt record was used to question the expert before the jury.

The danger goes past simple mistakes. An expert’s own work habits can become evidence that the analysis started with the answer.

According to 404 Media, plaintiffs’ counsel recognized what looked like ChatGPT output in the expert’s paperwork and sought further materials. Many will not be noticed. Nothing stops a report built the same way from sitting in a case file right now, on either side of a lawsuit, fluent and formatted and signed.