The Courts Have Stopped Warning

The question to ask of any legal-AI tool is no longer how it performs on a benchmark. It’s whether you can stand behind its output in open court.

Earlier this month, a federal judge in the Northern District of Mississippi did something that should concentrate the mind of every legal team using AI. She removed all of the lawyers from a case — counsel on both sides — after finding that filings from each contained citations to authorities that did not exist. The dispute was an ordinary contract matter. The fabricated cases were not. One of the lawyers, by her own account, had not known that AI could invent sources at all.

It is tempting to file this under “careless lawyer, unlucky day.” That would be a mistake. It is the latest entry in a pattern that has hardened, over the first half of 2026, from cautionary tale into settled judicial expectation.

From warnings to consequences

The arc is easy to trace and hard to dismiss. In Oregon, a court imposed $110,000 in fines and fees — the largest AI-hallucination penalty in US legal history — after three filings carried more than twenty fabricated citations, and dismissed the underlying case with prejudice. In Alabama, a family lost a trust appeal outright because their lawyer had cited cases that were not real; the court called the conduct egregious and barred him from filing there again without a second signature. Federal appellate panels have moved in the same direction, attaching dismissal to financial penalty so that an AI hallucination no longer costs a fine alone — it can cost the client the case.

The scale is the part that resists the “isolated incident” reading. A public database maintained by the researcher Damien Charlotin now records more than 1,600 instances worldwide of a court or tribunal flagging AI-generated fabrications in a filing. It is described as a work in progress, and it expands most weeks. Whatever else this is, it is not rare.

The point the headlines keep missing

The easy conclusion — “lawyers should stop using AI” — is not the one the courts are reaching. This is worth dwelling on, because it is the whole argument. When the Ninth Circuit sanctioned two attorneys this year, it was unusually explicit that it was not punishing them for using generative AI as such. The misconduct was filing the output without checking it. The distinction is not a technicality. It is the entire lesson.

The reason general-purpose AI keeps producing convincing fake citations is architectural, not a bug awaiting the next model. These systems are built to generate text that looks like the right answer — which, in most fields, is most of the job. In law it is the wrong job. A model cannot confirm that the case it cited exists, that the case says what the brief claims, or that it remains good authority. The plausibility is the trap. A fabricated citation reads exactly like a real one until someone checks.

A hallucinated case in a brief is worse than no brief at all. It doesn’t only fail the client — it corrupts the record the whole system depends on.

What this changes for legal teams

If the problem is unverified output rather than AI itself, the response is not abstinence. It is verification — built in, not bolted on. And that reframes the buying question for any team bringing AI into legal work.

The benchmark score is close to irrelevant. The questions that matter are plainer. Can a human trace every claim back to a primary source? Is the status of each development stated precisely — what is actually in force, what is merely proposed, what is a political agreement still pending ratification? Did a qualified person review the output before it reached you, or did a model hand you a draft and a confident tone? When you repeat what the tool told you — to a client, a board, a court — can you stand behind it?

Those are not features. They are the difference between intelligence and a liability dressed up as one.

The honest version of the claim

It would be easy, in a market this anxious, to answer the moment with a slogan — “100% verified,” “guaranteed accurate.” We don’t, and we won’t. A profession watching colleagues get sanctioned for trusting a machine has every reason to distrust anyone promising perfection. The credible claim is narrower and, we think, stronger: there is a human review layer, status is flagged with care, and the primary source sits one click from every item — so you are never asked to trust an output you cannot check yourself.

The refusal to overclaim is not modesty. For this audience, it is the point.

The courts have stopped warning. The teams that come out of this period in good standing will be the ones that stopped treating AI output as an answer, and started treating it as a draft that a person has to be able to verify and own.