In one sentence
The Kammergericht in Berlin has rebuked a lawyer because her pleading cited two court decisions that do not exist — presumably the work of a “fantasising” AI — making one thing clear: using AI is allowed, adopting AI citations unchecked is not.
The case: two sources that never existed
In a family-law matter (urgent proceedings with legal-aid funding), a lawyer's pleading relied on an order of the Federal Court of Justice (BGH) “of 14 November 2007 – XII ZB 183/07” and on a decision of the Higher Regional Court of Brandenburg. Neither exists. The Kammergericht admonished the lawyer by order of 20 November 2025 (ref. 17 WF 144/25); the case has been reported by beck-aktuell and the German Federal Bar Association (BRAK), among others. The court's central message is remarkably level-headed: lawyers may use AI — but they must verify every case-law citation before it leaves the office. This is long past being an isolated case: internationally, around 1,600 hallucination incidents have been documented across 35 countries, and US courts have already imposed sanctions.
Why language models invent judgments
A language model is not a database but a machine for completing patterns. It has learned what a BGH case number looks like, which senate sounds plausible for family matters and how a citation is phrased. Where the real decision is missing, it produces one that looks formally perfect — court, date, case number, all consistent. It is precisely this formal perfection that makes hallucinations so treacherous: they do not stand out when you skim, only when you look them up. Anyone who does not look them up notices nothing.
Professional conduct: responsibility stays with the human
The Kammergericht spells out what professional conduct rules require in any event: whoever is responsible for a pleading is responsible for every source in it — regardless of which tool produced the draft. That aligns with our product philosophy: AI supplies groundwork and assessments; checking and deciding is done by people. A tool that demanded blind trust would be the wrong tool for this profession.
How SmartLegalPro tackles the problem at the root
Hallucinated citations arise where the model phrases freely although looking things up is what is actually called for. That is why we intervene at three points:
- Rechtsbibliothek instead of model memory: Our knowledge graph holds real statutes and decisions with real sources. Citations are checked against this body of material, not “recalled” from the model.
- Source attribution: Where the AI carries out research, it discloses its sources — and marks honestly what is an official source and what is merely web research.
- Second check on deadlines: For deadlines, a second, independent calculation runs alongside — precisely because numbers and dates are among the most expensive hallucinations.
This, too, does not replace the lawyer's review. But it makes that review faster and gives it something solid to check against.
For consumers: “ChatGPT said so” is not an argument
What is a professional risk for lawyers often remains invisible to laypeople: a chatbot that invents a provision, a deadline or a judgment sounds just as convincing as one that is right. Anyone who lets a deadline lapse or gives up a claim on that basis has a real problem. Our initial assessment is therefore deliberately framed as an assessment — and when things get serious, the path leads straight to real lawyers who can take responsibility for the case.
Five points before an AI citation goes into a pleading
- Look up every source in the original: court, date, case number, headnote.
- Existence is not enough — check whether the decision actually supports what it is being cited for.
- Accept AI research only with disclosed sources, never on mere assertions.
- Independently recalculate numbers, deadlines and amounts as a matter of principle.
- When in doubt, leave it out: a missing citation is awkward, an invented one is a professional risk.
