A thousand attorneys have been sanctioned over AI. They all made one mistake.

In 2023, a $5,000 sanction against two New York lawyers for filing a brief full of invented cases was remarkable enough to make national news.

By the first quarter of 2026, courts issued around $145,000 in AI-related sanctions in three months. One Oregon attorney accounted for $109,700 of it. The Sixth Circuit added $30,000 for a set of briefs a court described as containing over two dozen fake citations.

A public database maintained by legal researcher Damien Charlotin now tracks roughly 1,490 court decisions worldwide involving AI-hallucinated material, more than 1,000 of them in the United States. About 496 involved licensed attorneys rather than people representing themselves. (Sourced, alongside the rest.)

What is striking is not the number. It is how similar the cases are.

What every one of them shares

Different tools. Different practice areas. Different levels of seniority, including attorneys at firms with review processes that were supposed to catch exactly this.

In every case, a statement reached a court without anyone tracing it back to a source.

That is the whole failure. Not that the tool was used. Not that it was the wrong tool. Somebody read a plausible sentence, found it plausible, and filed it.

The plausibility is the trap. A fabricated case citation looks precisely like a real one. It has a party name that sounds like a party, a reporter volume that sounds like a volume, a year that fits. There is nothing on the face of it to catch, which is why the error passes through associate review and supervisor review and lands in front of a judge.

Why “be careful with AI” is not the lesson

The obvious takeaway is to use it less. That reading does not survive contact with the cases.

Careful people were sanctioned. Firms with review layers were sanctioned. The Sixth Circuit case involved briefs in three consolidated appeals, which means multiple opportunities for someone to look.

Care is not the variable, because the failure is invisible to care. You cannot look harder at a sentence and see that the case behind it does not exist. You have to go and check whether it exists, and that is a different action.

The useful lesson is narrower and duller:

If you cannot point at the source, it does not go in the filing.

That rule would have caught every case in the database. It requires no judgment about which tools are trustworthy, and it does not get less effective as the tools improve.

Where this touches a plaintiff practice

Almost all the reported sanctions involve legal citations, because a fake case is something a court can spot. Opposing counsel looks it up. The clerk looks it up. The judge looks it up.

Nobody in the courtroom looks up whether page 604 of a medical record says what your chronology claims it says.

That does not make records work safer. It makes the failure quieter. A chronology with a wrong date, a misattributed provider, or a treatment gap that is an artifact of a missing production will not get you sanctioned. It will get you a demand package built on a fact that is not true, and an adjuster who finds it before you do.

The exposure is different. The discipline is identical.

What this means when you are buying

Most legal AI is now sold on accuracy percentages. Ninety-something percent, on some benchmark, measured by the vendor.

Accuracy figures are close to useless for this problem. A tool that is right 97% of the time produces three wrong statements in every hundred, and gives you no way to tell which three. What you need is not a higher percentage. It is a cheap way to find the three.

So the question to ask a vendor is not how accurate it is. It is: can I check this, and how long does checking take?

• Does every statement cite the specific page, document and line it came from?
• Can I open that source in one click, or do I search the record myself?
• When it is unsure, does it say so, or does it produce its best guess in the same confident voice as everything else?
• What does it reliably miss, and will you tell me before I ask?

A vendor who answers those four cleanly is offering something you can actually use. A vendor whose answer to all four is a benchmark score is offering you the exact shape of the problem in the database.

The one thing verification will not do

Checking a source protects you from statements that are wrong. It does nothing about statements that were never made.

In legal research this is the case the tool did not find. In records it is the provider whose file never arrived, producing no entry, no flag and no anomaly — nothing on the page to catch.

That failure needs a different check: comparing what you have against an independent list of what should exist. It is the step people skip, and it is the only one that can detect an absence. The procedure for both takes about ten minutes on a routine file.

Both matter. Only one of them is what the courts are currently fining people for.

The rule, stated plainly

Use the tools. They are useful, and firms that refuse them are going to be slower than firms that do not.

Then apply one rule with no exceptions: nothing goes into a filing, a demand, or a letter to an adjuster unless somebody has looked at the thing it came from.

Roughly 1,490 court decisions exist because that rule is easy to state and easy to skip when a deadline is close. It is not sophisticated advice. Neither is checking that the fire door is unlocked.