Artificial intelligence can help a lawyer search a large record, organize discovery, compare contracts, and start a draft. Those uses are real, but they do not make a legal conclusion correct. AI can introduce new errors, and a prediction based on past cases cannot reliably determine the outcome of a new one. The useful question is narrower: which part of a legal task can a particular system assist with, and how will the lawyer check its work?
AI in law is more than a chatbot
Legal teams have used forms of AI and statistical software for years. In electronic discovery, technology-assisted review helps classify large collections of documents. Contract analytics can identify or compare clauses across an agreement set. Generative AI is different: it produces new text, such as a research summary, draft letter, or proposed contract language. The American Bar Association's Formal Opinion 512 describes these uses and the professional duties that apply when lawyers use generative tools.
The distinction matters because the risks vary. A search tool can omit a relevant document. A clause extractor can misread a defined term. A language model can give a fluent explanation supported by a nonexistent case. A firm should evaluate the system and the task together, rather than assume that an "AI-powered" label says anything about accuracy.
Where AI can assist legal work
Research and first-pass analysis
A legal research tool can surface potentially relevant authorities, summarize a long opinion, or suggest questions for further research. It is useful as a starting point when the attorney can open the underlying sources. The lawyer still needs to confirm that each authority exists, is accurately described, remains good law, and controls in the relevant jurisdiction. A plausible citation or quotation in an AI answer is not itself a source.
This is a measurable limitation, not merely a theoretical concern. In a 2024 Stanford study of selected commercial legal research tools, researchers found hallucinations in 17% to 33% of responses to their test questions. That range describes the systems and test set studied in 2024; it is not an error rate for every product today. It does show why a source-linked answer still calls for independent review.
Discovery and document review
AI can help sort, tag, deduplicate, and summarize large sets of documents. That may help a review team focus its attention, but it does not guarantee that responsive or privileged material will be found. The team should test the tool against a representative sample, check misses as well as correct classifications, and preserve a review process that can explain how documents were handled. Formal Opinion 512 notes that the appropriate level of human review depends on the tool and task; for a tested contract summarization tool, a lawyer might verify a sample rather than manually reread every document.
Contracts and drafting
Clause extraction can make it easier to compare indemnity, termination, or notice provisions across agreements. A generative system can propose language or organize a first draft. The attorney must still read the actual agreement, check definitions and cross-references, and decide whether the language serves this client's objectives. A clause that works in one transaction may conflict with another contract, a statute, or the allocation of risk the client negotiated. Drafting speed is useful only when the review needed to make the result dependable is included in the workflow.
Can AI predict what a court will do?
Analytics can describe historical patterns, such as a judge's past rulings on a category of motion. They may help frame questions for litigation strategy. A past pattern, however, is not a reliable promise about a particular case. A new dispute has its own record, legal issues, procedural posture, and decision-maker. The underlying data may be incomplete or unrepresentative, and the law can change. The ABA opinion acknowledges predictive analytics as one possible use; it does not say that such tools can determine a case's result.
Any proposed outcome score needs a stated data source, a defined target, and validation for the type of matter in which it will be used. The NIST AI Risk Management Framework calls for testing AI under conditions similar to deployment and documenting where results may not generalize. For a client deciding whether to settle or try a case, an opaque score should not replace an attorney's assessment of the evidence, applicable law, uncertainty, and the client's goals.
The lawyer's responsibilities remain
Verify the work. Formal Opinion 512 says uncritical reliance on an AI output can lead to inaccurate advice or misleading statements to a court. Verification should match the stakes: check quotations and citations against primary law, compare summaries with source documents, and have the responsible lawyer approve advice and filings. Court-specific rules can add requirements. For example, U.S. Magistrate Judge Peter H. Kang's July 2025 standing order, applicable to civil cases before him in the Northern District of California, addresses disclosure and verification of AI-assisted filings. Counsel should check the applicable court's current orders before filing.
Protect client information. Before putting matter facts into a tool, determine who can access the inputs and outputs, how they are retained, and whether they may be used to improve a provider's system. The answer depends on the product's actual settings and agreement, not its marketing description. Formal Opinion 512 requires a fact-specific confidentiality assessment and explains when informed client consent is necessary. Do not assume that every tool treats privileged or sensitive material alike.
Keep professional judgment and fees honest. A lawyer remains responsible for advice, strategic choices, supervision, candor to the tribunal, and reasonable fees, regardless of how a draft was produced. AI may free time for analysis and client communication, but claimed savings should be measured on the complete task, including review and correction. There is no universal time-saving or error-reduction figure for legal AI.
The practical path is to use AI for a defined step, evaluate it on representative work, and keep a lawyer accountable for the result. For a concrete research-to-drafting example, see AI in legal workflows. For the broader competence question, read why lawyers need to understand AI.
Editorial disclosure: This revision was prepared with AI assistance. It offers general educational information, not legal advice.
