Use of AI in legal work

Language models can assist lawyers, but their answers should not be used without independent verification. NAMM attorney Simone Eelmaa discusses this topic in her article “Language Models in Practice: Relevance, Reliability and the User’s Verification Burden”, published in Juridica 2026/4, pp. 342–351.

In the article, Simone examines how the wording of a prompt affects the relevance and verifiability of a language model’s answer. She compared the results of two widely used language models in searching for Estonian criminal procedure case law and assessed, among other things, whether the cited judgments actually exist, whether they are substantively relevant, and whether the quotations and summaries correspond to the content of the judgments.

The tests showed that model outputs can differ significantly. One model mostly found existing and relevant judgments, while the other more often provided fabricated references or attached correct principles to the wrong judgments. Simone emphasises that stricter instructions help only if the model is also able to state honestly that it cannot provide a reliable answer. Otherwise, requiring a precise format may instead increase the risk of fabricated quotations and references.

The article is available in Juridica 2026/4.