Legal reasoning and research · Contested (since 2023)

GPT-4 passes a simulated Uniform Bar Exam; within a year, a Stanford study finds leading models produce false statements about federal case law in 58 to 88 percent of verifiable queries, depending on the model.

Exam questions and issue-spotting are largely ceded; drafting and first-pass research are used daily in firms. What is contested is reliability at the level a court requires: a 2024 audit of commercial legal research tools built on retrieval still found fabricated or misapplied citations in a substantial share of answers, and courts have sanctioned lawyers for filing them. Legal reasoning also ends in a signature. The lawyer who signs and the judge who rules are answerable in a way no system yet is.