Legal Services and Artificial Intelligence: Real Productivity Gains, Fabricated Citations, and a Profession Learning to Verify

Abstract

Purpose. Law is a profession built on text, which makes it an obvious target for language models, and one in which a confident error can end up in a court filing. This article asks where artificial intelligence stands in legal services in 2026 and where the profession may be by about 2035. Design and methods. A review combining the share of the legal research literature that involves AI (OpenAlex, 2010–2025), Eurostat's survey of AI use in EU professional-services enterprises (2021–2025), randomised trials with law students, audits of the accuracy of legal AI tools, a public database of court decisions on fabricated citations, and the regulatory position in the EU and the United States. Claims taken from the literature were checked against the sources cited. The scenarios are judgements built on this evidence. Findings. The share of legal works that mention AI rose from 0.7% in 2018 to 7.0% in 2025, 10.6 times as much and the second-largest increase of the 20 industries, and language-model works from 93 to 3,598. In the EU, 37.9% of enterprises in the wider professional-services sector used AI in 2025, up from 18.1% in 2021 and nearly twice the 19.9% for all enterprises. Randomised trials with law students show real gains: a recent trial found productivity gains of 50% to 130% on five of six tasks and better quality of work with a retrieval-based legal tool and a reasoning model, while an earlier trial with GPT-4 found large and consistent speed gains but only slight, inconsistent gains in quality. The risk is equally documented. Leading commercial legal research tools hallucinated between 17% and 33% of the time in a preregistered audit, general-purpose models at least 58% of the time on questions about federal cases, and a public database had logged 2,202 court decisions involving hallucinated content by October 2026. We expect AI-assisted drafting and research to be routine in firms by 2028, the profession to settle norms of verification and disclosure in 2029–2032, and the allocation of work between junior lawyers, software and clients to be the open question after that. Originality. The article sets randomised evidence of productivity beside audit evidence of error and a count of court decisions, and places both next to an enterprise survey whose sector is wider than law, with that limit stated.