Michael C. Baldwin, Certified General Real Estate Appraiser • Baldwin Appraisal Services LLC
This August, in a Houston courtroom, a jury put a $61.5 million price on a question an expert witness asked a chatbot.
The case was the Watson Grinding explosion — the January 2020 blast that killed three people and leveled some two hundred homes. 3M was defending claims that it failed to maintain the facility’s gas detection system, and its standard-of-care expert, an engineer with twenty years in the field, submitted a report concluding that 3M bore no responsibility. During discovery, plaintiffs’ lawyers noticed something off about the report’s citations. They asked the judge to order production of the expert’s ChatGPT logs.
They got 350 pages.
The logs showed the expert instructing the AI to “create an exceptional expert witness report defending the standard of care at 3M” and to “show how 3M is 0% at fault for the explosion at Watson Grinding.” They showed him uploading a photograph of the gas detector — the device the entire case turned on — and asking the machine to tell him what he was looking at. They showed him asking the AI, later, to soften the “0 percent responsible” language, because even he could see it wouldn’t survive.
Here is the detail worth sitting with. Plaintiffs’ counsel never moved to disqualify him. They called him, and let him explain his process to the jury. The jury found 3M thirty percent at fault.
The expert was an engineer, not an appraiser. That is a distinction, not a comfort. It means the profession that signs a certification page every time it renders an opinion of value has been given a preview — at someone else’s expense — of a cross-examination that is coming for us.
The Opinion Is Admissible Because It Is Yours
An expert opinion gets into evidence on a theory: that a qualified human applied a reliable method to the facts and reached a conclusion that is genuinely his own. Every rule that governs expert work — Rule 26’s report requirements, the Daubert reliability standard, and for appraisers the whole architecture of USPAP — is a load-bearing wall holding up that theory. Delegate the reasoning to a language model and the wall doesn’t crack; it disappears. The report may read beautifully. It just isn’t an expert opinion anymore, and courts have spent the last three years saying so, in an increasingly impatient line of cases.
A New York Surrogate’s Court got there first, in a case uncomfortably close to valuation practice. In Matter of Weber (2024), a fiduciary-damages expert used Microsoft Copilot to run the calculation at the center of his testimony. The court did something elegant: it ran his own prompts again, from the bench. Three runs, three different answers. The court found the testimony not credible, and held that AI-assisted expert evidence in New York now requires a reliability hearing before it comes in at all.
The federal cases followed. In Kohls v. Ellison (D. Minn. 2025), an expert on — of all things — AI misinformation was excluded after his declaration turned out to contain AI-fabricated citations. The court’s language is the kind that ends careers: “The Court cannot accept false statements — innocent or not — in an expert’s declaration submitted under penalty of perjury.” In In re Celsius Network (Bankr. S.D.N.Y. 2023), a 172-page report assembled with AI in roughly seventy-two hours was found unreliable; the expert had not reviewed his own sources. In Concord Music v. Anthropic (N.D. Cal. 2025), a single hallucinated citation — one — got a paragraph struck and, the court noted, cast a shadow over the credibility of the entire declaration.
Read together, the cases are not anti-technology. They are anti-substitution. The proof is Ferlito v. Harbor Freight (E.D.N.Y. 2025), where an expert used ChatGPT after the fact, to double-check conclusions he had reached through decades of hands-on experience — and the court let his testimony stand. The line the courts are drawing runs in one direction: a machine may check the expert’s work; it may not do the expert’s thinking. The 3M expert’s failure wasn’t that he touched AI. It’s that he ran the inference backwards — he fed the machine his conclusion and asked it to manufacture the analysis.
The Questions You Will Be Asked Under Oath
Every appraiser doing litigation work should assume the following exchange is now a standard module of deposition prep, because it is:
Did you use artificial intelligence in preparing this report? Which portions? What were your prompts? Are the logs preserved? Who selected these comparables — you, or the tool? Walk me through this adjustment without referring to your report. You’ve testified to twenty years of experience — why did you need to ask a chatbot what this was?
Two facts make these questions dangerous. First, your prompts are discoverable — the 3M case established that a judge will order the logs produced, and Weber established that a court may rerun them and watch the answers wobble. Second, “I don’t recall what I asked it” is not a safe answer. From the witness chair, it sounds exactly like I don’t know what’s in my own report.
There is also no longer any professional ambiguity about the standard. In April 2026 the Appraisal Standards Board adopted Advisory Opinion 41, guidance aimed squarely at technology in appraisal practice, and its premise is the same one the courts reached: tools cannot comply with USPAP — only appraisers can. AO-41 requires that an appraiser understand a tool well enough to know when its output is unreliable, disclose AI use whenever it materially contributes to an analysis, and — this is the part most of the profession has not absorbed yet — document in the workfile the tool’s output, the data it was given, and the prompts, along with the appraiser’s own analysis connecting that output to the conclusion. The workfile question and the deposition question have merged. They are now the same question.
What Responsible Use Actually Looks Like
None of this argues for abstinence. A Connecticut law firm recently ran an experiment feeding identical valuation prompts to several AI models and got divergent answers and invented “phantom sales” — which tells you what AI is bad at: being the source of record. What it is good at is speed — assembling, formatting, cross-checking, surfacing the thing you’d have found anyway an hour later. The discipline is keeping it on the right side of four lines:
The direction of inference. Evidence runs to opinion, never prompt to opinion. The analysis exists before the machine sees it; the machine’s job is to test it, organize it, and attack it — the Ferlito posture, not the 3M posture.
The source of record. Every fact in the report traces to a document, a record, or a verification a human performed. Nothing enters the analysis because a model asserted it.
Reproducibility without the tool. If the power went out, the opinion survives. Every adjustment, every comp selection, every capitalization decision can be defended at a whiteboard by the person who signed the certification.
Disclosure and documentation. AI work product is workfile content — preserved, producible, and disclosed where AO-41 requires it. If a prompt would embarrass you in a deposition, the time to discover that is before you write it, not after it’s produced.
How We Built for This
Our firm uses AI daily, and I will say so under oath without discomfort — because of what sits underneath it.
Baldwin Appraisal Services runs on a proprietary knowledge system we call the Baldwin Vault: a structured database of every comparable we have verified, every market we have surveyed, every methodology decision we have made, built up assignment by assignment across a multi-state practice. It is not a subscription and not a model. It is the firm’s record since inception — verified, market-specific evidence under codified methodology — and it compounds: every job makes the next one better documented. That is what makes it proprietary in the real sense: nobody can license it, because nobody else did the work.
A few of its rules will sound familiar after everything above. Every note in the system cites the workfile that holds its verification — no fact floats free of its source. Cap rates are typed — actual, modeled, or pro-forma — and every rent carries its basis, because mixing those classes is the most expensive category of error in valuation, and a database that lets you mix them silently is a liability generator. Every revision to a report is logged, permanently. And methodology changes have a single human owner: anyone may propose, one credentialed appraiser approves.
Notice what that structure produces as a byproduct: AO-41 compliance. When AI helps assemble one of our analyses, the output, the underlying data, and the reasoning connecting them are already in the workfile, because the system will not hold a conclusion any other way. AI makes us faster at assembling. The Vault is what makes the assembly true — and what makes the deposition questions above easy to answer.
For the Attorneys
The AI question is now a standard cross-examination topic, which means it should be a standard vetting topic. Before you designate an expert, ask what opposing counsel will:
- Do you use AI in your work? (Beware both wrong answers — “never,” from someone who plainly does, is as bad as “extensively,” from someone who can’t say where.)
- Where does the line sit between what the tools do and what you do?
- If your prompts and logs were produced tomorrow, what would they show?
- Can you re-derive your key conclusions without any of it?
- What system — not intention, system — keeps machine output out of your opinion unless a human verified it?
An expert with real answers to those five questions is not a liability you are managing. He is an asset the other side has to deal with.
The One Question
Twenty years of expertise, and the record now permanently contains the moment he uploaded a photograph of the most important piece of equipment in the case and asked a chatbot: what am I looking at?
That is the question an expert must never need to ask. The tools were never the expertise. The expertise is knowing what you’re looking at — and being able to prove, line by line, workfile by workfile, that you knew before the machine did.
Michael C. Baldwin, Certified General Real Estate Appraiser (CT RCG.0001733), is the principal of Baldwin Appraisal Services LLC, Waterbury, Connecticut, practicing commercial valuation, litigation support, and expert witness work in Connecticut and across a multi-state footprint.