What went wrong in the Watson Grinding case?
In August 2026, a Harris County, Texas jury found 3M 30% liable. It awarded $61.5 million to plaintiffs in one of several lawsuits from a 2020 explosion at Watson Grinding, a Houston manufacturing facility. The blast killed three people and destroyed roughly 200 homes. The verdict followed a failure of expert witness AI governance.
3M retained Josh Autenrieth of KnightHawk Engineering at $475 per hour. His task was an expert report on the company’s standard of care. KnightHawk received about $90,000 for the work. Then, during discovery, plaintiffs’ attorney Will Moye found a five-page document titled “Citation Overlay”. It appeared to be AI-generated. As a result, he requested production of all AI prompts used to prepare the report.
The result was 350 pages of ChatGPT conversation logs. They showed Autenrieth instructing the chatbot to “create an exceptional expert witness report defending the standard of care at 3M”. He also told it to “show how 3M is 0% at fault for the explosion at Watson Grinding.” In addition, he uploaded case documents. He asked ChatGPT to identify the equipment at the centre of the dispute (“what am I looking at?”). He then asked the chatbot to grade its own output. ChatGPT gave the report 97 out of 100.
When asked about his qualifications, Autenrieth even consulted ChatGPT. He asked whether his credentials were sufficient. He also asked how to defend against challenges to his lack of prior expert witness experience.
The report was about 30 pages long. According to the plaintiffs’ attorney, up to 90% of it was AI-generated content. Nearly 2,000 additional plaintiffs remain in the pipeline. The next trial is scheduled for October 2026.
Why does this matter for expert witness firms?
However, the Watson Grinding case is not an anomaly. It is the logical consequence of expert witnesses adopting AI tools without governance, disclosure frameworks, or quality assurance processes.
Three specific failures define the case. Each applies directly to any expert witness practice using AI today.
Discoverable prompt histories replace controlled audit trails. Autenrieth’s ChatGPT logs became the primary exhibit against 3M’s own expert. Every prompt was produced in discovery. So was every instruction to bias the output, and every request for self-grading. Therefore, firms using consumer AI tools are generating records they do not control and cannot redact. In US mass tort proceedings, AI chat histories are now established as discoverable work product.
Self-assessment replaces independent review. Asking ChatGPT to grade its own report is a closed loop with no quality signal. It is the AI equivalent of marking your own homework. The report never passed through an independent review process. That review could have identified bias, unsupported assertions, or structural weaknesses before it reached the courtroom.
Credential verification happens too late. An expert who asks ChatGPT whether his qualifications are sufficient is admitting a doubt. He is doing so in a discoverable record. That question should be resolved at instruction, not during report drafting. And it should not be answered by the same tool writing the report.
What are the regulatory obligations for expert witnesses using AI?
In England and Wales, the regulatory framework is tightening. CPR Part 35.3 requires that expert evidence be the independent product of the expert. It must be uninfluenced by the pressures of litigation. The Ikarian Reefer principles, established in 1993, remain the leading authority. They reinforce that an expert must not adopt the role of advocate.
In January 2026, the Academy of Experts published guidance for expert witnesses on AI use. Lord Neuberger of Abbotsbury endorsed it. The guidance distinguishes high-risk uses from low-risk uses. High-risk uses include generating substantive content or conducting material analysis on which the expert’s opinion will be based. Low-risk uses include administrative tasks such as transcription or summarisation. For high-risk uses, experts should disclose their AI use to instructing lawyers. They should also obtain approval before proceeding.
The Civil Justice Council launched a consultation in February 2026, chaired by Lord Justice Birss. It asked whether Practice Direction 35 should require experts to declare which AI tools they used, and for what purposes, in their statement of truth. The consultation closed in April 2026. A final report is expected later this year. The direction of travel is clear: proportionate transparency for expert evidence is coming.
A High Court judge issued an explicit warning to experts in January 2026 about AI-generated draft reports. And in Ayinde v The London Borough of Haringey [2025] EWHC 1383 (Admin), the President of the King’s Bench Division cautioned that guidance alone is insufficient to address AI misuse.
The Watson Grinding case shows what happens in a jurisdiction without these safeguards. The UK is building them. The question for expert witness firms is whether they will be ready.
CPR Part 35.3
An expert’s overriding duty is to the court. This duty overrides any obligation to the party instructing them or paying them.
How should expert witness firms respond?
The answer is not to stop using AI. AI can accelerate research, improve consistency, and help experts manage larger caseloads. Instead, firms should use AI within a governed framework. That framework must preserve the expert’s independence. It must also create a defensible audit trail. And it must ensure every output passes through qualified human review before it leaves the firm.
Four capabilities distinguish governed AI use from the ungoverned approach that failed in Watson Grinding.
Provenance and chain of custody. Every AI-assisted action should be logged with a timestamp. The log should name the expert who directed it, the model used, and the input provided. This is not surveillance. It is the digital equivalent of a lab notebook: a contemporaneous record of methodology and judgment. If AI prompt logs are going to be discoverable, they should tell a story of professional rigour, not of outsourced thinking.
Independent quality gates. AI output must pass through structured review before it contributes to a final report. A qualified person other than the AI user should conduct that review. The criteria should be defined, and the review itself recorded. A report that has only been reviewed by its author and the AI that drafted it has not been reviewed at all.
Credential and scope matching at engagement. The suitability of an expert for a particular instruction should be verified at the point of engagement. Check qualifications, experience, and prior testimony before the report is written. AI can assist with this matching. However, the decision is a human one, and it must be documented and traceable.
Disclosure-ready output. The Academy of Experts guidance is in force, and PD35 amendments are on the horizon. Firms therefore need a clear, accurate disclosure of AI use for any report. That disclosure should be generated from the audit trail. It should not be reconstructed from memory after the fact.
Ecsper provides this governed AI infrastructure for expert witness firms. AI actions, human decisions, and supporting evidence are captured in a tamper-evident chain of custody. The result is a defensible record of how work was produced, reviewed, and approved. With Ecsper, accountability never ends with “the AI did it.” It ends with a named person.
Check your governance exposure in three minutes.
See the full tracker: over 2,000 documented cases of AI misuse in court proceedings. Ecsper AI Risk Intelligence
