AI in Legal: Applications, Risks, and How Law Firms Actually Deploy It

AI in legal is the use of large language models for legal research, contract review, first-draft drafting, and document-intensive litigation work — always with a lawyer accountable for the output. The distinguishing fact of the domain is measurement: independent testing has found the leading legal research tools hallucinating in roughly 17% to 33% of responses, which is why supervision is an ethics rule here and a best practice everywhere else.

Why Legal Work Behaves Differently Under AI

Legal output is unusually verifiable and unusually consequential. A citation either exists or it does not; a deadline either was met or it was not. That makes some tasks — finding every termination clause in a contract set, extracting parties and dates, summarizing a document set — well suited to machines, because the result can be checked mechanically. And it makes the rest dangerous in a way that generating marketing copy is not: a fabricated case cited in a filing has already produced judicial sanctions.

Three constraints shape every deployment. First, confidentiality: client information cannot go into systems whose training and data-handling terms are unknown. Second, competence: lawyers must understand the tools they use well enough to catch their errors. Third, attribution: someone — a named attorney — is responsible for what gets filed, signed, or sent. AI changes the cost of producing the work; it does not move the responsibility.

Legal Research and the Citation Problem

Retrieval-grounded legal research assistants search a corpus of statutes, regulations, and case law, then answer in prose with citations. In principle the grounding fixes the fabrication problem. In practice, the model still composes authority that matches the shape of what it found — right court, right year, plausible paragraph — without the case existing.

The canonical incident is Mata v. Avianca, Inc., decided by the Southern District of New York in June 2023. A plaintiff's lawyer submitted a brief citing six cases generated by ChatGPT, none of which existed; when the court asked for copies, the fabricated opinions contained invented quotations and docket numbers. The court sanctioned the lawyers under Rule 11 for failing to verify what a tool produced. The opinion is now the reference point for the proposition that "the tool made it up" is not a defense — verification was the lawyer's job before the tool existed, and it remains the lawyer's job after.

The operational fix is unglamorous and non-negotiable: every citation is opened and read before it is relied upon, with particular attention to subsequent history — whether the case was distinguished, reversed, or criticized, which a summary sentence rarely tells you. Treat generated citations the way you would treat a first-year associate's: plausible until confirmed, never filed unconfirmed.

What the Accuracy Benchmarks Actually Show

The most-cited independent evaluation is the Stanford study of leading AI legal research tools, which tested Lexis+ AI, Westlaw-assisted generation, and general-purpose assistants against bar-exam-style and open legal queries with expert grading. The headline numbers explain why the profession is cautious rather than hostile.

Verdict: these tools are force multipliers for issue spotting and for drafting the frame of an analysis, and they are not authorities. The correct mental model is a diligent researcher who occasionally invents a case — useful for every draft, trustworthy for nothing you sign. Accuracy also varies by task type: the same tool that is solid on well-posted doctrine is weakest on niche statutes and jurisdiction-specific procedure, which is exactly where the stakes concentrate.

Contract Review and Abstracting

Contract work is where legal AI is most comfortably productive, because the source document is in hand and the output can be checked against it. Three patterns dominate.

1. Clause extraction: Pull governing law, term, renewal, liability caps, and indemnity language from hundreds of agreements into a structured table. The machine does the reading; a human confirms a sample and every flagged anomaly.
2. Deviation review: Compare a draft against the firm's playbook and list departures — auto-renewal removed, cap moved, venue changed. This is classification against a standard, and models handle it well when the playbook is explicit.
3. Diligence sweeps: Mass review for a specific issue across a data room — change-of-control clauses, assignment restrictions, MFN terms — with recall measured by seeding documents that contain the known answer.

The risk is asymmetric in an easy-to-miss way: models are better at finding what is there than at certaining what is absent. "No termination-for-convenience clause in this agreement" is a negative claim, and a model that read hurriedly will make it confidently. Evaluate extraction tools on recall against seeded ground truth, and require a human to confirm every absence before it drives advice.

E-Discovery and Technology-Assisted Review

Litigation's document problem predates the current wave of LLMs by fifteen years, and the field already has its accepted answer: technology-assisted review. TAR uses supervised classification to rank documents by responsiveness so reviewers read the relevant ones first instead of the whole collection.

The empirical anchor is Grossman and Cormack's study of the TREC 2009 Legal Track, which showed that technology-assisted processes achieved recall and precision superior to exhaustive manual review while reviewing only a small fraction of the collection — on the order of two percent, against the alternative of reading every document. Courts in the United States have accepted TAR where the producing party can explain the methodology, and proportionality under Rule 26 of the Federal Rules of Civil Procedure points in the same direction: the cost of review is weighed against the needs of the case.

Generative models now sit on top of that stack for summarization and issue coding, but the discipline carries over unchanged. Validate the pipeline with a known set, document the method for the record, and never let a summarization step silently become the destruction step — electronically stored information that the process cannot see is a spoliation question, not a performance optimization.

Drafting: First Passes, Not Final Words

The drafting use cases that work are the ones where the lawyer already knew what the document should say and the model removes the blank page: first drafts of client memoranda, routine engagement letters from a precedent, demand letters from an approved template, motion outlines built from an outline the lawyer supplied. Throughput rises because editing a plausible draft is faster than composing one.

The failure pattern is drafting that inherits confident wrongness — a motion citing the wrong jurisdiction's standard, a contract clause imported from a different transaction type, boilerplate that contradicts the negotiated deal three sections later. The controls are structural: generate only from the firm's own precedents rather than from the model's memory, keep a diff against the source template so every addition is visible, and reserve the final read for the person whose name goes on it.

The Ethics Framework in One Table

ABA Formal Opinion 512, issued in July 2024, and NYC Bar Formal Opinion 2024-5, issued in August 2024, reach the same place through existing rules: no new AI-specific duties, strict application of the old ones. The practical controls below map each duty to something a firm can actually implement.

On confidentiality specifically, the opinions distinguish "open" systems — general chatbots whose inputs may be retained or reused — from enterprise deployments with no-training commitments and contractual data terms. The rule of thumb the opinions endorse: do not enter information you would not want disclosed, and understand your provider's data terms before the first client file reaches it.

What a Deployment Playbook Looks Like

Firms that moved past pilot purgatory converged on the same operating model:

• Approved tools only: One list, maintained by risk and IT, with data terms reviewed — enterprise legal research products for research, contracted platforms with no-training clauses for everything else.
• Task boundaries in writing: What the tools may be used for, what they may never be used for (final authority, unreviewed client advice, filing), and what must be logged.
• Verification with evidence: The citation checklist is completed and retained with the work product, so the file shows the review happened rather than relying on memory.
• Seeded testing before adoption: Give the tool documents with known answers and measure recall and fabrication on your own matters before any workflow depends on it.
• Training that teaches failure: Associates learn to recognize confident fabrication, not just which button to press — the skill is skepticism with a procedure attached.

Measure the outcome in hours returned per matter and in the error rate caught at review. If errors are not being caught, that is not evidence the tool is working; it is evidence the review step is decorative.

Where Not to Use It

Certain uses remain inappropriate regardless of tool quality: anything filed or signed without a human reading the entire document; strategy advice built on generated summaries of authorities that were never opened; client-specific analysis run through consumer chatbots; and deadline calculations delegated to a model rather than a docketing system with an owner.

The common structure of these exclusions is that they remove the accountable human from the loop. The technology can improve; the allocation of responsibility cannot.