Legal AI is the use of artificial intelligence — mostly large language models — to research case law, draft and review contracts, sift through discovery documents, and monitor regulatory compliance. It is reshaping how much legal work costs and how lawyers spend their time, while introducing a risk specific to the profession: software that writes with total confidence and occasionally cites cases that do not exist.
The jobs legal AI actually does
Legal research is the most mature use case. Tools like Lexis+ AI, Thomson Reuters’ CoCounsel (built into Westlaw), and Harvey combine a language model with retrieval-augmented generation — meaning the system searches a real database of case law and statutes and grounds its answer in what it finds, rather than generating an answer purely from memory. This cuts down on errors compared with a general-purpose chatbot, though it does not eliminate them (see below).
Contract review and drafting is the other widely adopted use. Products such as Spellbook work inside Microsoft Word, comparing a contract’s language against a firm’s playbook, flagging non-standard clauses, and drafting a first-pass revision. Law firms report this cuts the time spent on routine reviews from hours to minutes, freeing lawyers for negotiation and judgment calls a model cannot make.
Document review and electronic discovery — sorting through the hundreds of thousands of documents produced in litigation or M&A due diligence — was one of the first legal tasks handed to machine learning, well before generative AI. Newer generative tools now summarize and cross-reference that material in plain language instead of just tagging it as relevant or not.
Compliance monitoring and “agentic law” is the newest and most ambitious layer. Instead of a lawyer looking up a regulation on demand, a company called Norm builds AI agents that have the relevant law encoded into them in advance, so they can flag a compliance problem — say, in a financial product’s marketing copy — automatically, as it happens, with attorneys supervising the output rather than producing it themselves.
Why it still gets things wrong
A 2024 Stanford RegLab study — the first controlled test of the leading legal-research tools — found that even RAG-grounded systems still hallucinate: Lexis+ AI and Ask Practical Law AI produced incorrect information in more than 17% of test queries, and Westlaw AI-Assisted Research in more than 34%. General-purpose chatbots without legal grounding fared far worse, hallucinating on 58% to 82% of legal questions in earlier benchmarks. Vendors that had marketed their tools as “hallucination-free” walked that language back after the study’s publication.
The consequence shows up in court. A database maintained by legal researcher Damien Charlotin has tracked over 1,700 cases worldwide where fabricated, AI-generated citations turned up in filings. A New York court ordered an attorney and his firm to pay $10,500 in sanctions in 2026 for a brief containing fake citations; the Alabama Supreme Court sanctioned a lawyer earlier the same year who cited a nonexistent case again in the very next filing after being caught once. The American Bar Association’s Formal Opinion 512, issued in 2024, is the leading professional guidance on the issue: it requires lawyers to independently verify anything a generative AI tool produces before relying on it, and to assess confidentiality risks before entering client information into a tool at all. AI hallucination is not unique to law, but a fabricated citation in a court filing carries consequences — sanctions, malpractice exposure, disciplinary referral — that a wrong answer in a chatbot conversation does not.
In the news
The scale of investor interest in this shift was visible when Norm raised $120 million at a $1.2 billion valuation, backed by Khosla Ventures and clients including Blackstone and Vanguard, to expand its compliance-focused legal AI agents.
FAQ
Can AI replace lawyers?
Not for judgment, negotiation, or representing a client’s interests, which remain the core of legal work. It can replace much of the repetitive research and drafting that used to consume junior lawyers’ time.
Is it safe to use ChatGPT for legal research?
General-purpose chatbots without a grounded legal database hallucinate on the majority of legal questions in benchmarks. Purpose-built legal research tools perform meaningfully better, but every citation still needs to be checked against the primary source before it goes in a filing.
What is “agentic law”?
A term for AI agents that carry out legal or compliance tasks autonomously — reviewing a document or flagging a rule violation without a person initiating each step — typically under attorney supervision rather than attorney execution.
Do bar associations allow lawyers to use generative AI?
Yes, with conditions. The ABA and most state bars permit it but require lawyers to understand the tool’s limits, verify its output, and protect client confidentiality — the same duties that apply to any other technology used in practice.
Sources: Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools — Stanford RegLab; ABA Formal Opinion 512 summary — The Bar Examiner; AI Hallucination Cases Database — Damien Charlotin; company reporting on Norm’s Series C round.