Lawyers will never eliminate AI hallucinations with today’s technology. But they can reduce them—sometimes substantially.
Hallucinations arise from fundamental characteristics of current large language models, which generate answers by predicting plausible language rather than independently determining whether every statement is true. Recent OpenAI research adds another wrinkle: AI systems can be rewarded for guessing rather than admitting uncertainty. OpenAI: Why language models hallucinate
Fortunately, there are several ways to reduce hallucination risk. Here are a few of the most promising:
Choose the Right Tool
Purpose-built legal systems such as Thomson Reuters CoCounsel Legal and Lexis+ with Protégé can ground their responses in curated legal databases rather than relying primarily on what the underlying language model learned during training. This is generally known as Retrieval-Augmented Generation, or RAG.
Grounding definitely helps, but it doesn’t make the AI infallible. The model can retrieve the wrong authority, misunderstand what a case says, overlook important contrary authority, or draw an unsupported conclusion from perfectly genuine sources.
The Stanford study Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools demonstrated this problem with earlier generations of commercial legal AI products. The specialized systems hallucinated substantially less than a general-purpose model, but they still produced incorrect or unsupported answers often enough to make lawyer verification indispensable. Stanford Law School: Hallucination-Free?
The products have evolved considerably since that research was conducted. Lexis+ AI, for example, has become Lexis+ with Protégé, and Thomson Reuters introduced a new generation of CoCounsel Legal in 2026. That makes the Stanford results important evidence about the problem, rather than a current scorecard for today’s products.
Use Good Prompts
Lawyers sometimes use instructions such as, “Act as a meticulous appellate attorney who values absolute factual precision over narrative flow.” More granular approaches may get better results. For example
“Do not guess or fill gaps in the evidence. If you do not have sufficient reliable information to answer a question, say so. Identify any material uncertainty. For factual or legal assertions, provide sources that I can independently verify.”
Jennifer Ellis’ Law Practice Today article, Four Strategies for Legal Professionals to Reduce AI Hallucinations, recommends several related techniques, including using specialized tools, restricting the sources the AI should rely on, providing more context, and independently checking its answers.
Other useful overviews include the ABA’s Generative AI for the Legal Profession: A Guide to Tool Selection, Risks, and Rewards and Nature’s excellent AI hallucinations can’t be stopped—but these techniques can limit their damage.
The common theme is that there is no single hallucination cure. Reliability comes from layers: better models, better sources, grounding, carefully designed prompts, opportunities for the AI to admit uncertainty, and—especially in legal work—independent human verification.
One technique deserves more attention than it usually gets: reducing what I call “answer pressure”—the subtle ways our prompts can encourage an AI to give us an answer even when it doesn’t have enough reliable information to do so.
I’ll discuss that in my next post.