Looking for precedent is the first use a firm finds for a language model, and it almost always starts the same way: paste the statement of facts and ask whether there is case law going the right way. It works. It is also the most expensive possible way to get that answer.
The model needs the question, not the parties
A statement of facts holds names, dates of birth, addresses, account numbers, the subject of the dispute and the docket number. Of all that, finding a precedent requires exactly one thing: the legal question, which usually fits in three lines and which a lawyer can frame better than any automatic summary.
“Construction contract, termination for breach, defects raised after the time limit: authorities on when time starts running where the defects are latent” is a better question than the forty lines it came from, not merely a safer one. A model handed the whole file spends half its attention working out who the parties are, and gives back a summary of your case when what you asked for was a list of authorities.
How to write a question that finds something
Framing tightly is not an exercise in caution, it is the technique that produces the better answers. It has a recurring shape, and it is four pieces.
The institution, called by its name (“termination for breach”, “dismissal for objective justified reason”, “avoidance in insolvency”). The provision, with the article and the code. The contested point, meaning the single thing you need authority on: not the case, the question. And the shape of the answer: “conflicting lines of authority”, “the Supreme Court’s position after 2020”, “whether there is anything going the other way”.
What this shape does not contain matters just as much: it does not contain a person. No name, no date of birth, no docket number, and therefore nothing to anonymise, because there was nothing to take out. That is why this is the one use of artificial intelligence in a law firm that does not raise the secrecy problem: it does not solve it, it sidesteps it.
When the document really is needed
Some requests do need the whole text: summarising a sixty-page pleading, comparing two versions of a contract, pulling out the defences the other side raised. Removing the document is not an option there, and the question becomes a different one: what leaves along with it.
This is where substituting beats deleting. A text where every party has a stable placeholder stays readable to the model: [FULLNAME_1] is the same person on page two and on page fifty, so the summary still tells claimant from defendant. A text full of black boxes does not, and the summary mixes up the roles. The difference between the two routes is in its own piece.
The answer still has to be checked, and there is a ruling that says so
The Court of Florence, in an order of its business division dated 14 March 2025, dealt with a filing citing judgments that did not exist, generated by ChatGPT during research carried out by an assistant and confirmed on a second query. The court held that the fact alone did not amount to aggravated liability under Article 96 of the Italian code of civil procedureopens in a new tab, but it wrote the part that matters: using these tools does not replace the professional’s duty to check and verify the information.
In practice: every citation goes into the real database before it goes into a filing. A model inventing a case number does it with exactly the confidence it cites a real one, and nothing in the text lets you tell them apart.
The check that works is shorter than it sounds: three tests in thirty seconds. The number and the year exist in the database. The division and the date match the ones cited. And, the one almost everybody skips, the judgment actually says what the model has it saying: the nastier hallucination is not the invented number, it is the real number attached to a proposition that ruling never stated. In Florence, precisely, the numbers pointed at Supreme Court decisions whose content had nothing to do with the issue in the case.
Not an isolated case, which is the surprising part
The first case anywhere was in New York in 2023: a lawyer sanctioned for filing citations generated by ChatGPT. The Italian reaction at the time was that it could not happen here. Three years on, the known Italian cases number at least five (alongside Florence, decisions in Turin, Latina, Milan and Syracuse are cited), and in one of them the costs order against the party that filed reached the order of thirty thousand euro.
The interesting figure is not the count, it is how consistently the same mechanism repeats: nearly always the research was done by somebody else, nearly always the person who signed did not know a model had been used, and nearly always nobody opened the citations. Those are three separate failures and they close in three separate ways (a written rule, a duty to declare the tool used, a check before filing), and none of the three requires giving up the tool. The page that writes them down is shorter than the pleading it takes to put things right.
Two minutes less, and one question fewer
Framing the question tightly has an effect nobody expects: it saves time. The answers are more relevant because the question is sharper, and there is nothing to clean up afterwards.
For the times the whole document is needed, the substitution has to happen before the text leaves, on the firm’s own computer. That is the work Omissia does, showing every value it finds before replacing it and putting the real names back into the answer when it returns.