Skip to content

Guide · AI for lawyers

AI for lawyers without invented citations: every answer with its source

A general-purpose chat can attribute to a ruling something it never said, or make up the ruling entirely, with the same confidence it has when it is right. This guide covers what that has cost real lawyers, why it happens, and how Custodio works so that every statement can be checked against the document it came from.

Citing a ruling that doesn't exist is no longer an anecdote: Spanish-speaking courts have fined lawyers for doing it, and in every case the tool that made the ruling up answered for nothing. Whoever signed the brief did.

The problem

What happened to lawyers who trusted the citation

In February 2026, the criminal chamber of the High Court of Justice of the Canary Islands fined a lawyer who had built an appeal on as many as 48 citations of Supreme Court rulings, plus a report of Spain's judicial council, none of which existed. A general-purpose AI tool had suggested them. The court didn't fault him for using AI: it faulted him for not checking that the numbers, dates and identifiers were real, when a single search in a free case-law database would have done it. It called that a breach of the basic duty of human oversight.

That same month, in order AC739-2026 of 13 February, the civil cassation chamber of Colombia's Supreme Court of Justice sanctioned a lawyer for this conduct, in what the press described as the country's first such case: an extraordinary appeal for review citing rulings, case numbers and reporting judges that the court's records office could not find anywhere. When the Court asked him to correct the filing, he added new fabricated citations. The fine was 15 monthly minimum wages, and the Court made the point that AI isn't banned, but no tool replaces the lawyer's personal responsibility.

The cause

Why it happens

A language model doesn't look the ruling up: it writes the most likely continuation of the text it is given. A case number, a date and a reporting judge have a very recognisable shape, and the model can produce that shape flawlessly with no ruling behind it. Colombia's Constitutional Court explained it in ruling T-323 of 2024: the next most likely word is not always the truest one, and among the risks of generative AI it named hallucinations explicitly. That is why it asked users to review, confirm and, where needed, correct what the tool delivers.

The problem isn't that AI gets things wrong now and then. It's that it gets them wrong in the same tone it uses when it's right, and a general-purpose chat doesn't show the lawyer where each sentence came from. Without that reference, checking means redoing the whole research.

How we approach it

How Custodio approaches it: the source travels with the statement

Custodio doesn't ask the AI to remember case law: it has it read the documents of your matter, and requires that everything it says stays tied to a passage in those documents.

  • Every fact keeps the exact fragment that supports it. When Custodio reads a document in the matter, each fact it extracts — a party, a date, an amount, an obligation, a decision, a legal citation — is stored with the literal fragment of the document and, when it is known, the page. A fact whose fragment doesn't appear verbatim in the document is discarded before it is stored.
  • The matter state accepts no statement without evidence. The current situation, the last relevant change, pending items, risks and upcoming events must cite facts, documents or court actions of that same matter. If the AI returns a statement without evidence, or with a reference the matter doesn't have, that result is rejected and not stored.
  • What is only inferred is not presented as confirmed. A date the document suggests but doesn't set stays as a candidate for you to review, never as a confirmed event.
  • Every answer comes with the documents it came from. When you ask something about the matter, the answer carries its sources: the document's name and, when it is known, the page, with a link that takes you there. Those sources are the passages the search returned, not a list the model drafts on its own.
  • If there's no source, it says so. An answer that cited no document from the matter shows exactly that, instead of looking like a backed-up answer.
  • What's missing gets flagged, not filled in. When a document names another one the matter relies on — a contract, a promissory note, a report — and that document isn't in the matter, or two documents disagree about a date, an amount or a party, Custodio flags it for your review instead of picking a version.

Run the test on your own file

Upload a document, let Custodio extract its facts and ask it something specific: every answer comes back with the document it came from.

Create your free account

Use Claude? Ask it from there

Your judgement

What this doesn't replace

Custodio doesn't promise an infallible AI, and its Terms of Service say so in Section 21: AI output is support material, it can contain errors, and you review it before using it in any filing. What changes is the cost of that review. Instead of researching from scratch whether a statement is true, you open the passage that supports it and read it.

This is one of the four problems the AI for lawyers guide works through. The other three are professional secrecy when you use AI, AI that doesn't know your case file, and AI that acts without asking you first.

Try it

How to check it yourself

  1. Create the free account, no card, and open a matter with a document you know well.
  2. Review the facts Custodio extracted and compare each one with its fragment.
  3. Ask it something you already know the answer to, and open the sources it returns.
  4. Ask it something the document doesn't answer, and see whether the answer cites any document.
FAQ

Frequently asked questions about citations and AI

Why does an AI invent rulings that don't exist?

Because a language model writes the most likely continuation of a text; it doesn't look anything up in a case archive. A case number, a date and a reporting judge have a very recognisable shape, and the model can produce that shape without any ruling behind it. Colombia's Constitutional Court warned about this in ruling T-323 of 2024: the most likely word is not always the true one.

Does Custodio guarantee the AI is never wrong?

No. No AI can guarantee that, and Custodio's Terms of Service say so in Section 21: AI output is support material and you review it before you use it. What Custodio does is make that review possible: every fact and every answer points to the document it came from, so you check it in one click instead of researching from scratch.

Where do the sources of an answer come from?

From the documents of the matter you are working on. The sources attached to an answer are the passages the search returned from those documents — with the document's name and, when it is known, the page — not a list the model writes on its own.

What if the answer isn't backed by any document?

The screen says it in so many words: "This answer cited no document from the matter." An answer without a source isn't drawn the same way as one with three, so an inference can't pass for a reading of the file.

Try it on a document you know by heart

Create your free account, upload a real document and ask it something you already know the answer to. Check which passage the answer comes from. No card.