Why AI Invents Statutes and Judgments

A language model isn't trained on facts but on likely continuations of text. When you write "according to the judgment of the Supreme Court, case no.", the model fills in something that looks like a case number, because it has seen thousands of such patterns. It has no way to verify that the decision actually exists. The same applies to section numbers, deadlines and the wording of provisions.

The less data on a topic the model has seen, the more confidently it makes things up. Czech legislation and case law are weakly represented in training data, so the risk is greatest precisely there. An answer in English or German may be more accurate than one in Czech, which is, paradoxically, the worst possible combination for Czech lawyers.

The first known case dates back to 2023 in New York, where lawyers submitted a filing to the court containing six non-existent decisions generated by ChatGPT and were fined. Since then, similar cases have been multiplying in the United States, Canada, the United Kingdom and Europe, and courts are beginning to require declarations on the use of AI. In the Czech environment, the first filings citing sources that cannot be traced are starting to appear.

An instruction in the prompt reduces the model's willingness to fabricate citations, but gives it nothing to replace them with. The model either refuses to answer, answers vaguely, or still makes things up anyway. The only reliable solution is to give it a source it can actually read from: this principle is called grounding.

Grounding works by having the model call a search in a real database before answering, retrieve the text, and answer based on it. The citations are then not generated but copied from the result, including a link you can open yourself.

• The AI receives results with the case number, court, date and a URL to the full text. It has no reason to add anything on its own.
• The section-lookup tool returns the exact wording, so the AI doesn't cite from memory even for numbers and deadlines.
• Citation links between provisions are only verified ones: the connector never returns a reference that isn't actually in the text of the regulation.
• When there's nothing in the corpus, the connector says so. The AI then has nothing to invent from and should admit that the source is missing.

Ask with an explicit request for a source. Have the AI search first, then summarise. Open every citation you use in a document via its link. For key statements, ask again with a differently worded query to check whether the same result comes up. And keep track of what the AI has searched for; the Europaius connector logs every query under your account, so you can always return to the research.