Lawyers are trained to trust confident, well-structured writing that cites authority. Generative AI produces exactly that style, whether or not the content is true. This is why AI mistakes are dangerous in law.
Where the errors come from
- Prediction, not lookup. A model writes what is likely to come next. If the true answer is rare or the question is unusual, it may produce something plausible but wrong.
- Out-of-date knowledge. A model's training stops at a date. New cases, amendments and guidance may be missing.
- Misgrounding. Legal research tools retrieve real documents and ask the model to answer from them. The model can still misread a source, cite the wrong one or state a proposition the source doesn't support.
- Jurisdiction blur. Models mix rules from different jurisdictions unless you are specific.
- Agreeableness. Ask a leading question and the model tends to agree with you.
How often?
A 2024 Stanford RegLab and HAI study tested leading legal research tools on a set of legal questions. Lexis+ AI and Ask Practical Law AI gave incorrect information more than 17% of the time; Westlaw AI-Assisted Research more than 34%. General chatbots did worse. The tools have changed since, and the vendors disputed parts of the method, but the lesson stands: even purpose-built legal AI makes mistakes that need a lawyer to catch.
What reduces the risk
- Use tools that show their sources and click through to them.
- Ask narrow questions with the jurisdiction and date.
- Give the tool the actual documents rather than asking from memory.
- Verify every authority before it is relied on.
- Match the review level to the stakes (more on this in L2).