The short answer: whether you can trust a legal AI has less to do with which model it runs and more to do with where its answers come from. Grounding, wiring the system to real, cited legal sources, is what separates output a lawyer can rely on from output that reads well and cites cases that do not exist.
The hallucination problem is really a citation problem
Language models are trained to produce fluent, plausible text. In most settings that is enough. In legal work it is not, because a plausible citation to a case that was never decided is worse than no citation at all. Across the category, even well funded platforms have had to confront the same issue. Independent reviews of leading legal AI tools, including well known names like Harvey, have flagged the risk of incorrect citations when outputs are not verified.
This is not a knock on any one product. It is a property of how large language models work. If the answer is generated from the model’s memory, the model will sometimes remember something that is not there.
What grounding actually means
Grounding means the system does not rely on the model’s memory for the law. Instead, it retrieves from a defined set of cited legal authorities at the moment of use, and every answer traces back to a real source. Qanooni is wired directly into more than 5,000 cited legal authorities and works from them in real time. The model is used to read, summarise and draft, not to remember the law.
The practical effect is simple: you can click through to the source behind any statement, and the system stays current as the law changes, because it is reading live sources rather than a snapshot frozen at training time.
Why this matters more than model size
A larger model is more fluent. It is not more accurate about the law. Fluency without grounding produces confident, well written answers that a busy lawyer is more likely to trust and less likely to check. That is the dangerous combination. Grounding inverts it: the system is only as authoritative as the sources it can point to, which is exactly the standard legal work is held to.
What to ask any legal AI vendor
- Where does each answer come from, and can I trace every citation to a real source?
- Does the system read live authority, or is it limited to what the model learned in training?
- What happens when the law changes, and how quickly does that flow through?
If the honest answer to the first question is the model, treat every citation as unverified. If the answer is these cited sources, with a link, you have something you can rely on.