Lawyers will never eliminate all Gen AI hallucinations using current technology. They are inherent in the way Large Lanuage Models are built.
The best we can hope for is to reduce their frequency. This is very possible, and there are many ways to go about it.
One of the best approaches is using premium, purpose-built systems like Thomson Reuters CoCounsel or Lexis+ AI. These applications layer highly customized models over proprietary, closed-universe legal databases (a specialized type of Retrieval-Augmented Generation, or RAG), forcing the AI to reason over verified primary law rather than the general internet.
Even enterprise or premium, customized legal apps aren’t perfect, but they have improved since the well-known Stanford study. Today’s best products do a better job, as explained in the new book The AI-Ready Law Firm: How Lawyers Use Generative Tools to Work Smarter and Serve Clients Better (ABA 2026).
Some lawyers report success with prompts that include language like “Act as a meticulous appellate attorney who values absolute factual precision over narrative flow.”
I have yet to find a excellent article on this topic, but vendors Machine Learning Mastery, De Community, Voiceflow, and Parasoft have some pretty good additional suggestions.
I’ll be explaining my favorite hallucination reduction technique, reducing “answer pressure,” in a future post.
Recent Comments