Two thousand judgments that don’t exist

As of 10 October 2026, Damien Charlotin’s database of “AI hallucination cases” counts 2,184 decisions in which a judge found citations or precedents invented by a language model. United States 1,492, Canada 233, United Kingdom 76, Italy 15. In September Reuters reported that error-ridden filings “surge despite three years of court sanctions”.

Three years of sanctions, and the curve rises. It is worth asking why.

Not a technology problem

The comfortable explanation is technical: models “hallucinate”, they will improve. True, but not enough. A filing with a non-existent judgment is not a model error. It is a filing nobody read before submitting it.

For years we have studied how little people read what they sign: fewer than one in a thousand open a licence; it would take 76 working days a year to read every privacy policy. We always treated it as a client problem. Charlotin’s database tells us it is ours too. The legal profession suffers from the illness it diagnoses in others: the ritual of signature instead of understanding.

The expert’s paradox

One detail strikes me. An invented judgment sounds right to an experienced lawyer. Plausible name, plausible number, plausible headnote. The model produces exactly what we expect to read. And expectation is what switches off scrutiny.

Reading research shows this again and again: familiarity with a genre lowers attention, it does not raise it. That is why in my LX Complexity Score work I learned not to trust a single indicator. This site’s log has an entry marked “wrong”: I thought one sensor was enough. It isn’t. You need convergence between different signals. The same holds for a law firm: one check, human or automated, is not enough.

Six months for the bar

Italy’s AI Act implementing decree, published on 8 October, gives professional bodies six months to adapt their rules. A rare opportunity. The temptation will be a generic ban or a duty of “conscious use”. That would be another ritual.

I would try rules that can be verified:

1. every citation in a filing must be opened and read by a person in the official source, with a record of the check;
2. the signatory declares which parts were produced or reviewed with automated tools;
3. the firm keeps an error log of what was caught, because it is the only way to know whether the process works;
4. mandatory training includes exercises on plausible fake texts, not only lectures on rules.

The real question

I do not think the profession’s future is writing less with machines. It will be writing much more, and better. But a lawyer who does not read what they sign is selling something they no longer have: judgment. The 2,184 decisions do not tell the story of a failed tool. They tell how much value a signature loses when nobody behind it has understood.