Legal Intelligence by Granid uses a large language model to answer legal research questions, draft documents, and work with your firm's files. It is built to work exclusively with official legal sources, but like every system built on this technology, it can make mistakes. This post explains why that happens, what we have engineered to reduce the risk, and how we make the remaining risk visible to you instead of hiding it.
What is an AI “hallucination”?
Large language models generate text by predicting, token by token, what is most likely to come next. That is what makes them remarkably fluent, and it is also why they sometimes produce statements that sound confident but are wrong. The industry calls these errors hallucinations: instead of saying “I don't know”, the model fills a gap with something plausible. In general-purpose chatbots this shows up as invented facts, mixed-up names, or, notoriously in the legal field, court decisions and statute citations that simply do not exist.
This is not a bug that a patch will remove. It is inherent to how the technology works, and any vendor claiming their AI never hallucinates is overpromising. What responsible engineering can do is make hallucinations rarer, and make sure that when one slips through, you can see it.
Why this matters more in legal work
In most professions a wrong AI answer costs a correction. In legal practice, a fabricated statute citation or an invented supreme court decision can flow into a client memo or a court submission, and courts around the world have already sanctioned lawyers for exactly that. For legal work, a made-up citation is worse than no answer at all. So the bar for a legal AI assistant has to be higher than for a general chatbot.
How Legal Intelligence reduces the risk
We did not solve this with a disclaimer alone. Legal Intelligence answers questions through a research pipeline built specifically to keep the model grounded, in five steps:
- Answers start from official sources, not from the model's memory.
When you ask a question, Legal Intelligence first retrieves the actual text of the law, court decisions, and legal commentary, exclusively from official sources: legislation as published by the state, decisions from the courts themselves. To this it adds relevant passages from your own documents. The model composes its answer from those retrieved texts. If you name a specific provision, the system fetches that exact article deterministically, character for character, together with a direct link to the official publication. The law is never reconstructed from the model's memory. - A structured research plan instead of free-form generation.
Every research question is first broken down into sub-questions and routed to the right sources: legislation, case law, commentary, your firm's files. Those searches run in parallel, and only their results reach the drafting step. The model works like an associate who is handed a file of sources, not one who answers from recollection. - Cross-references are looked up, never guessed.
The most common legal AI failure is invented case law. When Legal Intelligence retrieves a statute article, it expands it through a pre-computed citation graph: the court decisions that actually cite that article, and the commentary that discusses it, extracted from the real texts in advance. Connections between sources are looked up in that graph, not imagined by the model. - A second model double-checks relevance.
Before the answer is written, a dedicated ranking model re-scores every retrieved passage against your question, so the drafting step sees fewer, better sources instead of a pile of loosely related text. Less noise in means fewer errors out. - Every citation is verified after the answer is written, and problems are flagged, never hidden.
Once the answer exists, a deterministic verifier (no AI involved in this step) checks every statute, court decision, and commentary citation in it against the evidence the model was actually given. Citations confirmed by that evidence get a green badge. A citation that exists in our legal database but was not part of the evidence for this answer is marked as such. A citation that cannot be resolved at all is visibly flagged as potentially invented. The design is fail-closed: green appears only on confirmed evidence, and anything unconfirmed is shown to you rather than silently passed through.
What you should still do
- Click the sources. Every answer lists its sources with direct links to the official publication: the official legislation portal for statutes, the court's own website for decisions. Verification is one click away.
- Read the badges. A green check means the citation was confirmed against the sources the model saw. A flagged citation deserves your attention before it goes anywhere near a work product.
- Treat the assistant like a skilled junior. Fast, well-read, and useful, but its work is reviewed by a lawyer before it leaves the office. That is how Legal Intelligence is meant to be used.
The honest bottom line
No AI system today can guarantee a correct answer, and we will not claim otherwise. What Legal Intelligence guarantees is a process: answers grounded in official sources, citations checked against the actual evidence, and anything unconfirmed flagged to you instead of hidden. The professional judgment stays where it belongs, with you.