AI in Law
AI in law is the use of machine learning and, more recently, large language models to perform or assist work that lawyers and courts do with text: finding relevant documents among millions, extracting clauses from contracts, researching precedent, drafting memos and briefs, and answering legal questions for the public. Law is unusually text-heavy, rule-bound and expensive per hour, which makes it an obvious target, but it is also a field in which a confident error is not a typo but a sanctionable event, so the central engineering and professional question is not what the models can write but what a human can verify.
The oldest and best-established application is e-discovery, where technology-assisted review (TAR, often called predictive coding) ranks documents by likely relevance so that humans review far fewer of them. In a typical workflow, senior lawyers label a seed set, a classifier (classically logistic regression or a support vector machine over word features, later neural models) scores the remaining collection, and the process iterates until sampling shows acceptable recall, the fraction of truly relevant documents found. Courts began endorsing the method in the early 2010s; in Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012) Magistrate Judge Andrew Peck issued a widely cited opinion accepting computer-assisted review as permissible in appropriate cases, and later decisions such as Rio Tinto v. Vale (2015) addressed how much transparency about the process is required. The defensible claim is about measured recall and precision with documented sampling, not that the machine is smarter than a reviewer.
Contract analysis applies the same pattern to a different object: extracting and classifying clauses, parties, dates, termination rights and indemnities, then comparing them with a playbook or flagging deviations. Due diligence in a merger, where thousands of agreements must be checked for change-of-control or assignment clauses, was an early commercial use of machine-learned extraction, and large language models now handle the long tail of unusual phrasing far better than keyword rules did. The pitfall is silent omission: an extractor that misses one unusual non-compete clause looks identical, on the dashboard, to one that found everything, so serious deployments sample outputs against human review and keep a lawyer accountable for the conclusion.
The most famous documented cautionary example is Mata v. Avianca, decided in the Southern District of New York in 2023, in which lawyers filed a brief containing case citations that ChatGPT had invented. The plaintiff's attorneys, Steven Schwartz and Peter LoDuca, submitted opinions that did not exist, including a purported decision styled Varghese v. China Southern Airlines, and when the court and opposing counsel could not find them, the lawyers initially stood by the citations; Schwartz later said he had used ChatGPT for research and had not understood that it could fabricate. Judge P. Kevin Castel sanctioned the two lawyers and their firm, imposing a fine of five thousand dollars and requiring notices to the judges falsely identified as authors of the fake opinions, while separately dismissing the underlying personal-injury claim as untimely. The episode is a hallucination failure in its purest form: a language model optimizes for plausible legal-sounding text, not for retrieving verified authority, and a fluent citation with a realistic docket number proves nothing until checked in a real database. Many courts and judges have since issued standing orders or rules requiring disclosure or certification of AI use in filings.
Unauthorized practice and confidentiality are the two structural constraints that do not depend on model accuracy. Most jurisdictions restrict the practice of law to licensed lawyers, and consumer products that present themselves as a robot lawyer raise questions about where information ends and legal advice begins; the US Federal Trade Commission brought an action against DoNotPay in 2024 over claims about the service's capabilities, which concerned advertising substantiation rather than a ruling on the practice of law as such. Confidentiality is the other limit: a lawyer owes duties to keep client information private and to supply competent representation, so pasting privileged material into a third-party tool requires understanding its retention, training and access terms. The American Bar Association's Formal Opinion 512 (2024) addressed generative AI along exactly these lines, covering competence, confidentiality, communication with clients, supervision and reasonable fees.
Retrieval-grounded legal research tools reduce but do not eliminate fabrication, and a common misconception is that attaching a citation database makes a system reliable. Systems that retrieve real cases and then draft text from them can still misread a holding, cite a case for a proposition it does not support, or rely on overruled authority; a 2024 Stanford study of commercial legal research assistants reported that such tools still produced incorrect or misleading answers at a nontrivial rate, which is why vendor claims of hallucination-free output deserve scrutiny. The relevant discipline mirrors review of a junior associate's work: every authority is pulled and read, and the human signs the filing.
Read AI in law as a force multiplier for the tedious, measurable parts of legal work, such as ranking documents and extracting clauses, paired with a professional duty that cannot be delegated: the lawyer who signs remains responsible for every citation, every confidence kept and every conclusion drawn, which is the lesson Mata v. Avianca made public.