The Laundered Opinion

Every AI-in-the-courtroom scandal for the past three years has followed the same shape. A lawyer, or a paralegal, or a guy representing himself, asks a chatbot for case law. The chatbot — being a plausibility engine that has never once in its life known the difference between a real citation and a well-formed one — invents Varghese v. China Southern Airlines and six of its closest friends. A judge notices. Sanctions follow. Everyone writes the same column about how you have to check the model's work, and we all move on until the next one.

That story is boring now. It's a competence failure with a known fix.

This is a different story, and it is considerably worse.

In a Harris County courtroom this month, plaintiffs' attorneys in a lawsuit over the 2020 Watson Grinding explosion in Houston — three people dead, roughly 450 structures damaged — produced roughly 350 pages of ChatGPT prompt logs belonging to an engineer 3M had hired as its expert witness. The engineer, Josh Autenrieth of KnightHawk Engineering, had asked the model to "create an exceptional expert witness report defending the standard of care at 3M."

And then, in case that left any ambiguity: "show how 3M is 0% at fault for the explosion at Watson Grinding."

Trial testimony put the resulting report at somewhere between 85 and 90 percent machine-generated. According to the plaintiffs' attorney, Autenrieth acknowledged under oath that the prompts he entered were biased toward 3M, to help 3M win the case.

Nothing hallucinated. No fake citations, no invented precedent, no confabulated standard. The model performed flawlessly. It was handed a conclusion and asked to build a road to it, and it built a very clean road — the kind that comes with subheadings and a methodology section and the reassuring cadence of a professional who has considered the evidence.

The corruption did not come from the machine. The corruption arrived at the machine already fully formed, typed into a text box by a human being, and the machine's entire contribution was to make it look like engineering.

i · the prompt is the opinion

The whole architecture of expert testimony rests on a single load-bearing assumption: that the expert forms an opinion, and then writes it down.

That's not a formality. It's the reason the testimony is admissible at all. An expert is permitted to tell a jury things no other witness may — to draw inferences, to characterize causation, to say this hose failed because — precisely because the law presumes the conclusion is downstream of a method. Strip that ordering and you haven't got a weakened expert opinion. You've got advocacy in a lab coat, which is the one thing expert testimony rules exist specifically to keep out of a jury's ears.

"Show how 3M is 0% at fault" inverts the ordering explicitly, in writing, in a discoverable log. It is the conclusion, stated first, with the methodology commissioned afterward to match. The prompt isn't evidence about the opinion. In any sense that matters, the prompt is the opinion. Everything the model produced downstream is decoration.

The U.S. Chemical Safety Board, doing this the slow way, found the explosion traced to a degraded and poorly crimped rubber welding hose leaking propylene. Plaintiffs argued 3M had failed to properly service the facility's gas detection system. A jury eventually assigned 3M 30 percent of the fault and returned $61 million.

Thirty percent. Not zero. The number the report was commissioned to produce was never on the table.

ii · the friction was the ethic

Here's where I have to say the unpopular thing, which is that none of this is a new sin.

The hired-gun expert is older than most of the technology in this article. The "battle of the experts" — each side buying a credentialed professional who arrives at exactly the conclusion their retainer implies — has been a structural feature of American civil litigation for over a century. Nobody in the field is shocked that a company paid a firm to produce a report favorable to that company. Reporting puts Autenrieth's rate at $475 an hour and his total north of $90,000. That is not a scandal. That is Tuesday.

What changed is the price of the lie.

The old system was held together not by virtue but by effort. Producing a forty-page report that looked defensible — that had the structure, the citations, the technical register, the appearance of a methodology honestly applied — took a competent engineer real weeks. And in those weeks, something useful happened. Somebody had to actually engage the physical evidence long enough to build a plausible story around it. Somebody had to look at the hose. Somebody had to know what the gas detection system did, because you cannot fake the vocabulary for forty pages without absorbing some of the reality underneath it.

That friction was doing enormous unacknowledged moral work. It wasn't a conscience. It was a tax — a cost high enough that most people, most of the time, found it cheaper to shade an honest analysis than to fabricate a dishonest one from scratch. The system's integrity was never primarily a function of individual character. It was a function of the fact that lying well was expensive.

The obvious objection is that this brake had already been released. A $475-an-hour expert has not personally typed forty pages in decades; junior engineers, staff, and firm boilerplate have been doing that work since long before anyone had a chat window, and the hired-gun economy ran at industrial scale anyway. Fair. But delegation is not elimination, and the gap between them is where the whole argument lives. When juniors wrote the forty pages, somebody still had to hold the physical facts in their head — read the CSB findings, look at the hose, learn what the detection system was supposed to do — because human beings cannot generate forty pages of technical register without absorbing the domain. The contact with reality was distributed, not deleted. And it left witnesses: people who could be deposed, timesheets that could be subpoenaed, a chain of humans each with a license, a conscience, and a name on the invoice.

What's new is that the number of people required to touch reality went to zero, and the number of witnesses went with it. That's the categorical part. Not that fabrication got cheaper. That the loop closed with nobody inside it.

Remove the expense and you remove the brake. That's not a moral failure of one engineer in Houston. It's an environmental change, and environments beat intentions every time. You can write all the professional-ethics guidance you want; what actually governs behavior is the relative cost of the available paths. We just made one of those paths approximately free.

This is what people mean, or should mean, when they worry about AI and epistemics — and it's almost never what they say. The fear is usually framed as the machine will generate falsehoods. The real exposure is that the machine collapses the cost of manufacturing the appearance of rigor, and appearance-of-rigor was the currency the entire professional-expertise economy was denominated in. Credentials, formatting, citation density, technical register: these were proxies for effort, and effort was a proxy for engagement with reality. Both links in that chain just snapped, and we're still pricing the currency as if they hadn't.

iii · thirty percent is the counterargument

The honest problem with everything I just said is sitting in the verdict.

The jury assigned 3M thirty percent. The counterfeit lost. If appearance-of-rigor is the currency and it just became free to mint, why did freshly minted rigor fail in front of twelve people with no engineering background?

Because juries were never running on appearance-of-rigor. Cross-examination was. The adversarial system does not evaluate a report by admiring it — it puts the author in a chair and asks him questions he did not write, and that procedure is entirely unchanged by the existence of ChatGPT. You cannot prompt your way through a deposition. Whatever the document looked like, the man behind it had to answer for it out loud, and under oath he conceded the prompts were biased toward 3M to help 3M win. The system worked.

I want to concede that completely, because the concession is where the real story is.

It worked here. In a courtroom — the single environment on earth purpose-built on the assumption that everyone in it is lying, staffed by a professional whose entire compensation depends on catching it, and equipped with compulsory disclosure that hands him his opponent's files. Courts have an immune response. It is slow, adversarial, absurdly expensive, and it functions.

Now name one other institution that runs on the same proxy — a credentialed professional producing a document that looks like considered judgment — and has any of that machinery. Peer review doesn't. A reviewer reads the paper; nobody deposes the author and nobody subpoenas the lab notebook. Audit opinions don't. Due-diligence memos, analyst notes, credentialing bodies, consulting deliverables, the entire apparatus of institutional trust: all of it verifies by reading the artifact, and the artifact just got cheap.

Expert testimony isn't the disease. It's the first symptom that turned up somewhere with an X-ray machine.

There's also a version of this that's genuinely good news, and skipping it would be exactly the sin I'm describing. A $90,000 expert report is not only a weapon. It's a wall. It's the reason a family whose house came apart at 4:30 in the morning cannot put a competent-looking technical rebuttal in front of a jury and 3M can. Collapse the cost and the floor rises — the same tool that let KnightHawk manufacture a defense lets an underfunded plaintiff finally afford one. That's a real gain and I'm not going to wave it away.

But it arrives with an asymmetry that eats most of it. When production quality is cheap for everybody, it stops carrying information about anybody. The plaintiff gets a report; the jury loses whatever weak signal document quality used to provide about which side had actually touched the evidence. Access goes up, signal goes to zero, and what remains is whoever can afford the better cross-examination — which is, once again, 3M. The wall didn't come down. It moved to a room with worse acoustics.

iv · the logs remember

And now the genuinely satisfying part, which I'd enjoy more if three people hadn't died.

He got caught because of the tool.

Plaintiffs' attorney Will Moye noticed a five-page "Citation Overlay" document in discovery that had the unmistakable smell of machine generation. So he searched its contents on the open web — and found Autenrieth's ChatGPT conversations sitting there, publicly accessible. Shared chat links. The deposition reportedly stopped for three hours while roughly 350 pages of prompt logs were produced. 3M then tried to disqualify its own expert; Moye called him to the stand anyway.

An engineer used a system that keeps a perfect, timestamped, verbatim record of every instruction he gave it, to do something he would not have wanted recorded, and then left the record where a search engine could reach it.

The ledger cuts both ways. That's the thing about building your fraud on infrastructure — infrastructure logs. And the precedent now established is bigger than one report: chatbot session logs are discoverable. They are work product. If you used a model to reach a conclusion, opposing counsel gets to read every word of how you asked. Every defense engineering firm that has run a live case through a chatbot is currently having a very quiet, very expensive week.

So here's my weary prediction, and I'd like the record to show I'd rather be wrong.

Inside litigation, the correction is already underway, and it's the boring kind that works. Courts have been issuing AI-disclosure and certification requirements since the Mata v. Avianca sanctions in 2023. And "run it locally, keep no history" is not the clean escape it sounds like — deliberately failing to preserve records after a duty to preserve has attached is spoliation, which draws an adverse-inference instruction and is a substantially worse day than the one Autenrieth had. The courtroom will metabolize this. It metabolizes everything, slowly, at enormous cost, which is more or less the definition of an immune system.

The problem is that the courtroom is a rounding error.

Everywhere the same architecture operates without discovery — which is everywhere else — the selection pressure we just applied doesn't select for honesty. It selects for opacity. Nobody subpoenas a peer reviewer's chat history. Nobody deposes a consultant. The lesson available to every profession that watched Houston and does not operate under compelled disclosure is not "form your opinion before you write it down." It's "don't use the tool that keeps a receipt" — and unlike the legal version, there is no spoliation doctrine waiting on the far side of that choice, because there was never a proceeding to spoliate.

We caught this one because a plaintiff's lawyer had good instincts, a search bar, and subpoena power. Remove the third and the first two find nothing. Outside a courthouse that isn't luck. It's just quiet.

The uncomfortable truth underneath all of it: the model did nothing wrong. It was asked to make a predetermined conclusion look like a considered judgment, and it succeeded, because that is a language task and language tasks are what it is superb at. We built a machine that is world-class at producing the texture of expertise, handed it to a set of institutions whose credibility rested on that texture being expensive to fake, and then acted surprised.

But the counterfeit isn't the failure. The failure is older and it's ours: we agreed, nearly everywhere, to verify by resemblance. Rigor's texture was never rigor. It was a proxy for contact with reality that held only while the proxy was costly, and it stood in for the thing so long and so reliably that we forgot it was standing in for anything. AI didn't break that. AI is the first tool cheap enough to make the gap visible — a signal decoupling from what it signified, in public, with a timestamp.

The hose was degraded and poorly crimped. That was true before anyone opened a chat window, and it stayed true through 350 pages of prompting designed to make it not matter. Reality was never in the document. We just kept checking the document.

And then, in one of the last buildings in America that still opens the document and makes the author explain it out loud:

Thirty percent.