Generative AI has moved from novelty to fixture inside the legal profession, drafting memos, surfacing precedent, and reviewing contracts at a speed no associate can match. But the same tools that compress hours into seconds have also fabricated cases, misstated holdings, and led at least one legal team into court-ordered sanctions. The evidence points to a clear operating principle: AI is decision support, not decision replacement. It accelerates the search and the first draft; the lawyer verifies the law and stays accountable for what gets filed.
Key takeaways
- Hallucinations persist even in specialized tools: A 2024 Stanford RegLab/HAI study found leading legal research AI hallucinated on roughly 17% to 33% of queries, despite vendor claims of being “hallucination-free.”
- The cautionary precedent is real: In Mata v. Avianca (2023), a federal judge sanctioned two attorneys $5,000 after they cited six ChatGPT-fabricated cases.
- Adoption is accelerating fast: Survey data shows AI use among legal professionals jumping from roughly 11% in 2023 to about 30% in 2024, with later surveys reporting much higher figures.
- The market is consolidating around big players: Thomson Reuters acquired Casetext for $650 million in 2023; its CoCounsel assistant has since passed 1 million users.
- The throughline is human accountability: Bar associations and courts have responded with guidance reaffirming that lawyers remain responsible for verifying AI output.
Stanford RegLab/HAI, 2024
Mata v. Avianca, S.D.N.Y., 2023
Thomson Reuters, 2023
Grand View Research, 2024
The Stanford reality check: even specialized tools hallucinate
The most rigorous public test of legal AI to date came from Stanford’s RegLab and Institute for Human-Centered AI. Researchers led by Varun Magesh and Faiz Surani preregistered and ran a battery of more than 200 legal queries against retrieval-augmented tools from LexisNexis and Thomson Reuters. The headline finding: Lexis+ AI hallucinated on roughly 17% of queries and Westlaw’s AI-Assisted Research on roughly a third, even though vendors had marketed “hallucination-free” citations. The authors concluded bluntly that “providers’ claims are overstated.” The takeaway is not that the tools are useless, but that purpose-built, citation-grounded systems still require verification before their output reaches a brief. Source: Stanford RegLab and HAI, 2024. link
Mata v. Avianca: what happens when verification fails
The cost of skipping the verification step is no longer hypothetical. In Mata v. Avianca, a routine personal-injury dispute, the plaintiff’s attorneys submitted a brief citing six judicial decisions that did not exist; all had been generated by ChatGPT, which even produced fake quotations and citations on request. On June 22, 2023, U.S. District Judge P. Kevin Castel sanctioned the lawyers and their firm $5,000 under Rule 11 and ordered them to notify the real judges falsely named in the fabricated opinions. The case became the profession’s defining warning that an AI’s fluent confidence is not evidence of accuracy, and that the filing attorney, not the tool, bears responsibility. Source: Mata v. Avianca, Inc., S.D.N.Y., 2023. link
CoCounsel and Casetext: the incumbents’ bet on assisted drafting
The market response has been to embed AI inside trusted research platforms rather than replace lawyers with them. Casetext launched CoCounsel in March 2023 as one of the first GPT-4-based legal assistants for document review, research memos, and contract analysis. Months later, Thomson Reuters acquired Casetext for $650 million, folding its technology into the Westlaw ecosystem; by early 2026 CoCounsel reported passing 1 million users across more than 100 countries. The framing from these vendors is consistent and telling: the tools handle the first pass on review and drafting, while the lawyer directs the work and signs off. Source: Thomson Reuters and LawSites reporting, 2023–2026. link
Harvey and Luminance: contract review and the new entrants
Beyond the incumbents, well-funded specialists are pushing into drafting, due diligence, and contract review. Harvey, founded in 2022, reports use across thousands of organizations and a substantial share of large firms, and has raised capital at multibillion-dollar valuations, signaling deep investor conviction in legal-specific AI. Luminance, founded in 2015 out of University of Cambridge machine-learning research, applies a proprietary legal model to contract review and negotiation across more than 70 countries. Both position their systems as co-pilots that flag risky clauses and surface relevant language for a human reviewer rather than auto-executing legal judgment. Source: Company disclosures and CNBC reporting, 2025–2026. link
Market growth, with an honest caveat
The commercial momentum is real but should be read with caution. Grand View Research valued the global legal AI market at roughly $1.45 billion in 2024, projecting growth to nearly $3.9 billion by 2030; other firms publish materially different sizes and growth rates, so any single figure is best treated as an estimate, not a settled number. Adoption surveys show the same upward trend and the same variance: AI use among legal professionals appears to have climbed from around 11% in 2023 to roughly 30% in 2024, with some later surveys reporting far higher individual usage even as formal firm-wide deployment lags. The direction of travel is clear; the precise magnitudes are not, and reports that headline a single dramatic percentage should be read against differing methodologies.
The through-line: decision support, not replacement
Across every credible data point, the same pattern holds. AI is exceptional at the high-volume, first-pass work of legal practice, finding candidate precedent, drafting clauses, and flagging anomalies in a contract, but it cannot be trusted to be the final arbiter of what the law says. The Stanford findings and Mata v. Avianca are two sides of one lesson: the technology extends a lawyer’s reach without transferring the lawyer’s duty. Bar associations and courts have reinforced this by issuing guidance that places verification squarely on the practitioner. Used that way, as a tireless research assistant whose work is always checked, AI raises throughput and lets lawyers focus judgment where it matters. The human stays in the loop not as a formality, but because accountability cannot be delegated to a model.
Methodology & sources
- Legal AI tools hallucinate on roughly 17–33% of queries despite “hallucination-free” claims — Stanford RegLab and HAI (2024)
- Peer-reviewed version of the hallucination study — Magesh et al., Journal of Empirical Legal Studies (2025)
- $5,000 sanction for citing six AI-fabricated cases — Mata v. Avianca, Inc. (2023)
- Thomson Reuters acquired Casetext for $650 million — Legal Dive (2023)
- CoCounsel passes 1 million users — LawSites (2026)
- Harvey valuation and customer reach — CNBC (2026)
- Luminance legal AI for contract review — Luminance (2025)
- Global legal AI market size estimate — Grand View Research (2024)
Frequently asked questions
Can AI replace lawyers for legal research?
No; AI is best understood as decision support, not a replacement for a lawyer’s judgment. Stanford research found that even specialized legal research tools hallucinated on a meaningful share of queries, so a qualified attorney must verify citations and reasoning before relying on AI output. The technology accelerates research and drafting, but accountability for the result stays with the human.
What was the Mata v. Avianca case?
Mata v. Avianca was a 2023 federal case in which attorneys were sanctioned $5,000 for submitting a brief citing six nonexistent cases fabricated by ChatGPT. The judge found the lawyers had failed to verify the AI-generated authorities and required them to notify the real judges named in the fake opinions. It is now the leading cautionary example of why AI output in legal filings must be checked.
How accurate are AI legal research tools?
They are useful but imperfect, with a 2024 Stanford study reporting hallucination rates of roughly 17% to 33% for leading commercial tools. Accuracy has improved as vendors add citation grounding, but no tool eliminates errors, so professional verification remains essential. Treat these systems as fast first-draft assistants whose conclusions must be independently confirmed.
Part of our Real-World AI Use Cases series — how AI supports high-stakes decisions across surprising domains.