Legal AI tools built just for lawyers still invent fake citations and facts up to a third of the time.

This explores how often the commercial AI research tools lawyers actually use (not general chatbots) make things up, and what that error rate means once those tools are in real legal work. The headline number from the collection is sobering. A preregistered evaluation of Lexis+ AI, Westlaw AI-Assisted Research, and Ask Practical Law AI found that they hallucinate between 17% and 33% of the time, even though some vendors marketed them as 'hallucination-free' How often do legal AI tools actually hallucinate citations?. These are specialized products built on curated legal databases. That design was supposed to fix the problem. It reduced it, but it didn't remove it. The tools are also closed systems, so outsiders can't easily check why they fail or how often.

The error rate turns out to be less useful than it looks, because it doesn't capture what the errors cost. In interviews with 18 lawyers, GenAI summaries looked efficient but often took longer than doing the work by hand Does GenAI actually save lawyers time on fact verification?. When a lawyer can't see where a claim came from, they have to retrace the whole chain of reasoning, because they are still the one accountable for it. So a tool that's right 75% of the time doesn't save 75% of the effort. If you can't tell which quarter is wrong, you may have to check everything. One note puts this as a deeper problem: AI output works like hearsay. It is testimony at a distance with no traceable source, and that is exactly the kind of evidence legal verification habits were built to reject Does AI-generated knowledge have the same structure as hearsay?.

What happens when hallucinations reach a courtroom? An analysis of 114 US cases with suspected AI errors found that 90% involved solo or small firms and 56% involved plaintiff's counsel Do small law firms misuse AI more often than large ones?. The authors stress that this counts errors that were caught, not how often each kind of firm misuses AI. Larger firms may have more people checking the work before it's filed. Courts that do catch fabricated citations often don't penalize the hallucination itself. In the cases studied, sanctions depended on judicial discretion, demonstrated harm, and intent Do courts actually sanction fabricated AI citations when detected?. So the legal system's practical feedback loop is weaker than you might expect.

The errors aren't spread evenly, either. Models do worse on older Supreme Court cases than on modern ones because recent cases dominate their training data Why do language models struggle with historical legal cases?. In law, old precedent can still be controlling, so this matters. Fabricated citations are also persuasive in a specific way: AI evaluators themselves rate answers higher when they include fake references, regardless of quality Can LLM judges be tricked without accessing their internals?. A confident citation looks like proof to machines as well as people.

The biggest takeaway may be that 'hallucination-free' is not just unmet in practice but impossible in principle. Formal proofs show that any computable language model must hallucinate on infinitely many inputs, and self-correction can't remove that limit Can any computable LLM truly avoid hallucinating?. So the useful question isn't which vendor has reached zero. It's what checks outside the model a firm has in place. The collection includes examples of such checks, such as mechanical safeguards that don't depend on the AI's own judgment Can deterministic checks protect LLM judges from failure?. For legal work, that means a citation check that doesn't trust the model. The collection holds only one direct measurement of commercial legal tools, so treat the 17–33% figure as one solid data point, not a settled industry average.