AI in the legal industry: what it does, what it breaks, and what it means for lawyers

Executive summary

AI in law is the use of machine-learning systems — especially large language models — to perform legal work such as research, contract review, drafting, and document discovery. As of 2026 it has moved from pilot to production: the Stanford AI Index 2025 found 31% of legal professionals personally using generative AI at work, up from 27% a year earlier, while Thomson Reuters reports 26% of legal organisations now actively deploy it, up from 14% in 2024. The economic pull is concrete — Thomson Reuters' 2025 Future of Professionals report estimates AI could free roughly 240 hours per lawyer per year, an average $19,000 in value each. Capital agrees: legal-AI startup Harvey reached an $11 billion valuation in March 2026 (CNBC), and Thomson Reuters' CoCounsel passed one million users by February 2026. But the technology's defining weakness is fabrication. A Stanford RegLab study (2024) found that even purpose-built tools from LexisNexis (Lexis+ AI) and Thomson Reuters (Westlaw AI-Assisted Research) hallucinate between 17% and 33% of the time. Those errors have reached courtrooms: Mata v. Avianca (SDNY, June 2023) drew a $5,000 sanction for six invented ChatGPT cases, and legal researcher Damien Charlotin's database counted 1,598 fake-citation filings worldwide by June 2026. The through-line for 2026 is supervised augmentation: AI accelerates discrete tasks, but licensed lawyers verify outputs and stay liable — a stance now reinforced by the EU AI Act, which treats AI in the administration of justice as high-risk.

Key findings

What is AI in law, and why now?

AI in law is the application of machine-learning systems — above all large language models (LLMs) like those behind ChatGPT, Claude, and GPT-4 — to legal tasks such as research, contract analysis, drafting, and document discovery. It spans two generations: older 'predictive' AI (technology-assisted review, clause classification) and the current wave of 'generative' AI that reads and writes natural-language legal text. The shift is why 2023–2026 became an inflection point rather than an incremental one.

The 'why now' is the arrival of usable generative tools. Casetext launched CoCounsel — billed as the first AI legal assistant — on 1 March 2023 using early, exclusive access to GPT-4; Thomson Reuters acquired Casetext for $650 million that August. Adoption followed the capability: the Stanford AI Index 2025 found 31% of legal professionals personally using generative AI at work, up from 27%, and Thomson Reuters reports organisational use rising from 14% (2024) to 26% (2025).

The economics explain the urgency. Thomson Reuters estimates AI could free roughly 240 hours per lawyer per year, worth about $19,000 each (Future of Professionals, 2025). Legal work is expensive, text-heavy, and billed by the hour — a near-perfect target for language models. That same text-density is why the technology's failure mode, fabricated citations, is so consequential in law specifically.

Can AI do legal research?

AI can do legal research, but it must be verified — the leading purpose-built tools still hallucinate 17–33% of the time, per Stanford RegLab (2024). Modern legal-research assistants such as Lexis+ AI, Westlaw AI-Assisted Research, and CoCounsel use retrieval-augmented generation (RAG): they retrieve real cases and statutes from a curated database, then have an LLM summarise and cite them. This grounding cuts fabrication sharply versus a raw chatbot but does not eliminate it.

The numbers set expectations. In Stanford's benchmark, Lexis+ AI answered 65% of queries correctly while Westlaw AI-Assisted Research was accurate only 42%, and Westlaw hallucinated at nearly double Lexis's rate. Both still outperformed general GPT-4, which fabricated far more often. The practical takeaway: AI is a powerful first-pass research engine that surfaces authorities and drafts memos in minutes, but every cited case must be pulled and read before it reaches a filing.

How is AI used for contract review and drafting?

AI is used to review, compare, and draft contracts by extracting clauses, flagging risks, checking against playbooks, and generating redlines and first drafts in minutes rather than hours. Tools including CoCounsel, Harvey, and dedicated contract-lifecycle platforms ingest an agreement, identify non-standard or missing terms (indemnity, liability caps, governing law, termination), and propose edits aligned to a firm's or client's standard positions.

This is one of the highest-ROI uses because contract review is repetitive, volume-heavy, and pattern-based — exactly where LLMs excel. Thomson Reuters' 240-hours-per-year estimate is driven substantially by these routine drafting and review tasks. The persistent caveats are confidentiality (client data must stay within controlled, enterprise-grade environments) and verification: AI-proposed clauses can be subtly wrong or omit protections, so a lawyer's judgement on materiality and risk allocation remains the sign-off step.

What is AI's role in e-discovery?

In e-discovery, AI reviews and prioritises large document sets in litigation and investigations, with generative tools now explaining why a document is relevant rather than only scoring it. E-discovery was the earliest home of legal AI: 'technology-assisted review' (predictive coding) has been court-accepted since the early 2010s. The generative wave, led by Relativity's aiR for Review, adds transparent rationales and counterpoints for each decision, addressing the 'black box' objection that limited earlier adoption.

The efficiency gains are large and measurable. Relativity reports that a government agency used aiR for Review to run issues review across 650,000 documents in a single week, identifying key documents in roughly 20% of the time of traditional review. Relativity, alongside vendors such as Everlaw and DISCO, dominates this segment. Defensibility — the ability to demonstrate to a court that the review process was reasonable and auditable — is the metric that matters, which is why transparent, explainable AI is winning here.

What are generative AI copilots for lawyers?

Generative AI copilots for lawyers are assistants that sit across a firm's work — research, drafting, review, and knowledge — and execute multi-step legal tasks from natural-language instructions. Harvey and Thomson Reuters' CoCounsel are the two most prominent. Harvey, built on frontier models and used across law firms and corporate legal departments, reached an $11 billion valuation in a March 2026 round (CNBC), after a $300 million Series E at $5 billion in June 2025. CoCounsel passed one million users by February 2026 across 107 countries.

These copilots are moving from single-task tools toward 'agentic' workflows that chain steps: read a matter, pull authorities, draft the memo, then a review pass. LexisNexis' Lexis+ AI and Protégé compete directly. The commercial signal is strong — capital and users are flowing in — but the reliability signal from Stanford's study tempers it: copilots amplify a lawyer's throughput, not their licence to skip verification.

Where exactly is AI used in legal work?

AI in law concentrates in five recurring, text-heavy tasks: legal research, contract review and drafting, e-discovery, litigation and brief drafting, and knowledge management. The table below maps each use case to what the AI does and a named, real-world example as of 2026.

How widely is AI adopted in the legal industry?

AI adoption in law is now majority-curious and minority-operational: about 31% of legal professionals personally use generative AI at work (Stanford AI Index 2025), while organisational deployment reached 26% in 2025, up from 14% in 2024 (Thomson Reuters). Openness is far higher than deployment — a defining 2026 gap. The table below compiles adoption figures by source.

The pattern across sources is consistent: individual use and enthusiasm outrun firm-wide, policy-governed rollout, held back by ethics, confidentiality, and reliability concerns.

What are the real AI hallucination court cases?

The landmark AI hallucination case is Mata v. Avianca, Inc. (S.D.N.Y., June 22, 2023), in which Judge P. Kevin Castel sanctioned two attorneys and their firm $5,000 for submitting a brief citing six wholly fabricated cases invented by ChatGPT. Attorney Steven Schwartz of Levidow, Levidow & Oberman had asked ChatGPT for supporting authority; it produced fake decisions — including 'Varghese v. China Southern Airlines' and 'Martinez v. Delta Airlines' — with fake quotes and citations, and even confirmed they were real when asked. The court found the lawyers acted in bad faith by not verifying them.

Mata was the first, not the last. Damien Charlotin's AI Hallucination Cases database, started in April 2025, documented roughly 200 such filings by mid-2025 and 1,598 worldwide by June 2026 — spanning general chatbots and, in some instances, even the specialised legal tools. Penalties have escalated from Mata's $5,000 into the tens of thousands of dollars, alongside referrals to bar discipline and, in the most serious cases, attorney suspensions.

The lesson courts keep restating: a lawyer's duty of candour and competence is non-delegable. Using AI is permitted; filing its output unverified is sanctionable. This is why every credible legal-AI workflow ends with a human pulling and reading each cited authority.

What are the risks, ethics, and regulation of AI in law?

The main risks of AI in law are hallucinated authority, breach of client confidentiality, bias, and lack of accountability — each with a concrete mitigation and a tightening regulatory backdrop. Bar associations (including the American Bar Association's Formal Opinion 512, 2024) require competence, confidentiality, communication with clients, and reasonable fees when using generative AI. The EU AI Act classifies AI used in the administration of justice as high-risk under Annex III, with full obligations applying from August 2026; lawyers acting merely as end-users of general-purpose models are generally not high-risk providers themselves. The table maps risks to mitigations.

Will AI replace lawyers?

AI will not replace lawyers in 2026; the evidence points to augmentation, not automation — AI performs discrete tasks while judgment, advocacy, negotiation, and legal accountability remain human. The tasks AI does well (research first-drafts, clause extraction, document review) are inputs to legal work, not the whole of it. A licensed lawyer must still exercise judgement, sign filings, appear in court, and bear liability — a boundary the hallucination cases have hardened, not softened.

What changes is the shape of the work and its economics. Thomson Reuters' estimate of ~240 freed hours per lawyer implies a shift away from the billable-hour model toward value and outcomes, pressure on entry-level research-heavy roles, and rising demand for lawyers who can supervise, verify, and deploy AI competently. The likely 2026–2030 trajectory: fewer hours on routine text, more on strategy, counsel, and client trust — with 'AI-literate lawyer' becoming a baseline skill rather than a specialty.

How is AI in law regulated in Europe and the Nordics?

In Europe, AI in law is governed primarily by the EU AI Act, which took full effect for high-risk systems in August 2026 and classifies AI used in the administration of justice — such as tools that assess evidence or predict case outcomes — as high-risk under Annex III. Providers of such systems face obligations on risk management, data governance, technical documentation, logging, transparency, and human oversight. Ordinary law-firm use of general-purpose AI to draft or review documents generally falls outside high-risk provider status but remains subject to GDPR and professional-conduct rules.

In the Nordics, national guidance is filling in around the EU framework. In Sweden, the Agency for Digital Government (DIGG) and the privacy regulator IMY issued generative-AI guidance in January 2025 covering information security, copyright, data protection, and ethics, and the Swedish Bar Association (Sveriges advokatsamfund) has moved to shape how firms use AI consistent with confidentiality and independence duties. The direction across the region is cautious adoption under explicit human-oversight and data-protection guardrails.

What is the outlook for AI in law?

The outlook for AI in law through 2027 is rapid, supervised expansion: capability and adoption keep climbing while accountability tightens. Expect agentic copilots that chain multi-step tasks, deeper integration of research, drafting, and review in single platforms (Harvey, CoCounsel, Lexis+ AI), and continued vendor competition funded by valuations like Harvey's $11 billion (March 2026).

Two forces will discipline that growth. First, reliability: independent benchmarks like Stanford's keep vendors honest and verification norms strict as long as hallucination rates stay in the double digits. Second, regulation: the EU AI Act, bar-association opinions, and a growing case record — 1,598 fake-citation filings and counting — make human sign-off a permanent feature, not a transitional one. The winning model for 2026 and beyond is the same one the data supports: AI does the volume, lawyers own the judgement.

Scoreboard (machine-readable data)

Every headline indicator with its value, period, source and confidence. Free to reuse under CC BY 4.0.

↓ CSV · ↓ JSON

Methodology & verification

This report synthesises public, named-source data on AI in the legal industry as of July 2026. Every quantitative claim is attributed inline to its primary or authoritative source: the Stanford RegLab reliability study (2024, peer-reviewed in the Journal of Empirical Legal Studies, 2025), the Stanford HAI AI Index 2025, Thomson Reuters' Future of Professionals 2025 report (survey n=2,275), LexisNexis Future of Work 2025, Damien Charlotin's AI Hallucination Cases database, primary court records (Mata v. Avianca, S.D.N.Y.), the EU AI Act (Annex III), and vendor and financial-press disclosures (Harvey, Thomson Reuters/CoCounsel, Relativity). Figures are reported as stated by their sources; where a source gives a range (e.g., 17–33% hallucination), the range is preserved. No statistics were estimated or interpolated by Affärslivet. Source URLs were verified to resolve (HTTP 200) as of 2026-07-30.

Data dictionary

Frequently asked questions

Glossary

Hallucination

When an AI model generates false information — such as a non-existent case or citation — presented as fact; measured at 17–33% for leading legal-research tools. ↗

Retrieval-augmented generation (RAG)

A technique where an AI retrieves real documents from a curated database and grounds its answer in them, reducing but not eliminating hallucination. ↗

Technology-assisted review (TAR)

Predictive coding used in e-discovery to prioritise relevant documents, court-accepted since the early 2010s and now augmented by generative AI. ↗

High-risk AI system

Under the EU AI Act (Annex III), AI systems — including those in the administration of justice — subject to strict obligations on oversight, documentation, and data governance. ↗

Sources

• Stanford RegLab — Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (2024)
• Stanford HAI — 2025 AI Index Report
• Thomson Reuters — Future of Professionals 2025
• Damien Charlotin — AI Hallucination Cases database
• Mata v. Avianca, Inc. (S.D.N.Y., 2023)
• EU Artificial Intelligence Act — Annex III (high-risk systems)

Related reports & data

• AI in Finance
• The EU AI Act, explained
• AI and Jobs