AI hallucination rates have dropped about 96% since 2021 (!), from roughly 20% down to under 1% for the best models today. In three years, AI went from getting 1 in 5 things wrong to getting 1 in 100.
However, here’s the problem: legal information still has a 6.4% hallucination rate even among top models, versus 0.8% for general knowledge questions. Westlaw AI and Lexis+ AI, the two tools specifically built to solve this for lawyers, still hallucinate between 17-33% on legal questions (according to a 2025 empirical study out of Stanford).
AI is getting dramatically better, and legal AI is getting better, too. Just more slowly. The gap between what vendors claim and what independent research finds is also still pretty wide.
Today, I’m diving into the stats, explaining why legal is harder, and talking about what actually helps you get more accurate legal answers when you’re using AI.
I’ll break down which legal tasks are most prone to hallucination, why giving AI your actual documents dramatically changes the accuracy equation, and what verification steps you should be building into your workflow right now.
Read more
