"The AI Did It" Just Stopped Working in Court. Your Field Is Next.

I have signed things I only skimmed. Early-career me initialed a forty-page deployment runbook after reading the headings, because the meeting was in ten minutes and the document looked like it knew what it was doing. Nothing went wrong that day, which was the worst possible outcome. It taught me that skimming works, a lesson I then spent several production incidents unlearning.

The legal profession is currently unlearning that lesson in public, with judges grading the homework. What the courts decided this month should worry every profession that certifies work with a signature. Last time I checked, that is all of them.

The month the excuse died

On July 16, the Arizona Court of Appeals put lawyers and self-represented litigants on notice: cite fake AI-generated cases and you can be sanctioned no matter your intent. Honest mistake, sloppy delegation, malicious fabrication, the court does not care. The question is no longer whether you meant to deceive anyone. The question is whether you checked.

The rest of the month kept the theme. The Fifth Circuit fined a lawyer 2,500 dollars for a reply brief full of AI-hallucinated quotations, then added punishment for her evasive response when caught. The Eleventh Circuit referred Anthony Sabatini, a Florida county commissioner and former congressional candidate, for discipline over AI-generated errors in his briefs, with a judge observing that AI is "no substitute for actual intelligence." A Pennsylvania attorney was suspended by a federal court in June over hallucinated citations. A Texas federal judge issued sanctions in a TCPA case for the same offense. One attorney in a suit against Roc Nation is now facing his third sanction for quoting AI inventions, which suggests the first two lessons did not fully land.

The genre has a database now

None of this is anecdote anymore. A public database maintained by researcher Damien Charlotin tracks court decisions that address AI hallucinations in filings, and it has become a growth chart. It listed roughly 1,624 cases in mid-June and roughly 1,725 by early July, about a hundred new entries in three weeks. The genre that began with the famous 2023 New York case, Mata v. Avianca, now grows almost daily, and one legal-tech commentator summed up the trend with the driest line of the month: the one thing we are getting faster at with AI is hallucinations in cases.

Sit with the shape of that curve for a second. Three years of warnings and headline sanctions have not slowed it down. Every lawyer in the country has heard of the case where this went badly. They keep filing anyway.

Why competent people file fiction

The lazy explanation is that these are careless people, and the numbers say otherwise. The wave includes experienced appellate litigators and at least one elected official. Something structural is going on, and I think it is a simple asymmetry: generation became free while verification stayed expensive.

A model produces a confident, perfectly formatted brief in forty seconds. Verifying its thirty citations means pulling thirty cases and reading them, which is hours of the exact tedious work the model was hired to eliminate. The economics practically beg you to skip it. The output looks like diligence, reads like diligence, and is priced like a lottery ticket where the losing numbers get you suspended.

The signature was supposed to be the checkpoint. A professional signature has always meant one specific thing: a competent human stands behind this. Somewhere along the way it quietly became a formality stapled to whatever the machine produced.

The courts have now noticed the difference, and they have decided the signature still means what it always meant. You signed it, you own it. "The AI did it" has roughly the same legal weight as blaming your dog.

Your signature is next

It would be comfortable to file this under lawyer problems, and it would be wrong. Every serious profession runs on the same ritual: engineers stamp drawings the way accountants sign audits, and each signature certifies that a qualified human verified the work. A physician's chart note carries the same promise. In every one of those fields, AI is now drafting the work product while the verification budget stays flat.

The Arizona standard travels well: intent is irrelevant, verification is the job. When that logic reaches your industry, and it will, the question your regulator asks will not be whether your team meant well. It will be whether anything in your process actually checked the machine's claims before your name went on them.

What we built instead of willpower

Disclosure time: my company works on exactly this problem, so discount what follows as much as you like. I would. But the mechanism matters more than the pitch.

The failure mode above is a human being asked to summon willpower at 11pm to re-verify work that looks finished. Willpower loses that fight often enough that courts built a database about it. So at Coheria we moved the check into the pipeline itself. No single model is ever trusted alone; specialized experts from different model families review the same output and have to corroborate it before it moves. Where a claim can be verified mechanically, our Truth Oracle verifies it with real tools rather than asking a model to grade itself. And every decision is sealed into an immutable hash-chain audit log, so when someone eventually asks what was checked and by whom, the answer is a record instead of a memory.

None of that eliminates hallucination, and anyone who claims otherwise is auditioning for the database. What it does is make verification the default path instead of the virtuous one, so the unchecked confident answer never gets the chance to look finished.

The part i keep thinking about

The tragedy of the sanctions wave is that every one of those briefs was checkable. The cases either existed or they did not. The quotes either appeared in the opinion or they did not. This was the easiest possible verification problem, binary and mechanical, and the database has still logged more than 1,700 court matters involving AI hallucinations.

Now imagine the claims that are not binary, like the load calculation buried in an engineering report or the revenue assumption holding up a financial model. Signatures are riding on unverified machine output in those fields right now. The only difference is that their database of consequences has not been built yet.

I got lucky with my skimmed runbook all those years ago. The profession that just spent a month in the sanctions headlines did not, and the lesson is the same one I keep relearning the hard way: trust is not a workflow. If your process depends on a tired human choosing to double-check something that looks done, you do not have a verification step; you have a hope step. Tell me I am wrong in the comments, ideally with citations that exist. More at coheria.ai.