AI for lawyers: use it like an associate, check it like an adversary
The associate that never sleeps also never checks its work. Where AI pays off in a law practice, and the two mistakes that end careers.
In February 2026, a federal judge in Manhattan handed prosecutors about thirty documents a defendant believed were private. Bradley Heppner, a corporate executive charged with securities fraud, had used an AI chatbot to think through his defense after receiving a grand jury subpoena: prompts, responses, strategy, all saved. His lawyers claimed privilege. Judge Jed Rakoff said no. A chatbot is not a lawyer, its terms of service promise no confidentiality, and talking to one is legally closer to thinking out loud in a crowded room than to calling your counsel.
That ruling is the sharpest version of a problem every lawyer now faces. AI is genuinely good at the drudgery of legal work: reading, summarizing, drafting, comparing. Most of the profession already uses it. But the two things that make legal work legal, verified authority and protected confidence, are exactly the two things consumer AI defaults get wrong.
Use AI like a tireless junior associate, and check it like an adversary: verify every citation as if opposing counsel wrote it, and keep client material on systems you contract with or hardware you own. This post maps the four jobs where AI earns its place in a practice, then the two mistakes that have already cost real lawyers real money, and in one case handed a defense strategy to the government.
Four in five legal professionals already use AI
Skip the question of whether lawyers should touch this technology. They already have. Clio’s Legal Trends Report, which surveys the profession annually, found AI use among legal professionals jumped from 19% to 79% in a single year. The same report estimated that up to 74% of hourly billable work, the information gathering, document drafting, and analysis that fills an associate’s day, could be automated with AI in some form.
The economics explain the speed. A profession that bills by the hour and drowns in text is the single best market for a machine that reads and writes at almost no cost. A solo practitioner gets the most out of it: the research assistant, the first-draft writer, and the summarizer a small firm could never hire are now a monthly software bill. Clients, for their part, have stopped minding. The same survey found 70% of them either prefer a firm that uses AI or are indifferent to it, which makes sense from the receiving end of an hourly invoice.
So the useful question is no longer adoption. It is discipline: which half of the work you hand over, and which half never leaves your desk.
AI earns its keep on four jobs, none of them judgment
Across practices, the wins cluster into four pillars, and every pillar has the same shape: the model does volume, you do judgment.
Researchis the pillar most lawyers try first. A model is excellent at mapping an unfamiliar area, explaining doctrine in plain English, and producing a starting list of authorities. Treat that list as leads, not law. The model’s account of what a case holds is a hypothesis until you have pulled the case and read it.
Draftingis where the hours actually come back. First drafts of contracts, engagement letters, demand letters, internal memos, and discovery requests are pattern work, and pattern work is what language models do best. The lawyers getting real leverage feed the model their own precedents and clause libraries so it drafts in the firm’s voice, then edit rather than compose. The rule that keeps it safe is old: nothing goes out unread, and the person who signs is the person responsible.
Review and summarization is the quiet workhorse. Triage in discovery, first-pass due diligence, comparing two versions of a contract, and summarizing a deposition are jobs where the model reads a hundred times faster than you and misses less through boredom. One caveat: on very long documents, models degrade quietly, so chunk the work and spot-check against the source rather than trusting one giant pass.
Client communication is the underrated fourth pillar. Turning a dense settlement letter into a plain-English summary for a client, drafting a status update, or structuring intake notes costs a model nothing and buys goodwill. The advice itself stays yours; the model just makes it readable.
Even the tools built for lawyers invent cases
The reflexive answer to hallucination is “use a legal AI tool, not ChatGPT.” It is half right. Purpose-built legal research tools ground their answers in real case databases, and they are measurably better than general chatbots. They are not fixed.
Stanford’s RegLab ran the first pre-registered evaluation of these tools, putting 200+ legal research questions to the paid AI products from LexisNexis and Thomson Reuters and hand-scoring the answers. The results: Lexis+ AI gave false or misgrounded answers more than 17% of the time, and Westlaw’s AI-Assisted Research more than 34%. That is far better than a bare chatbot, which the same group found hallucinated on 58% to 82% of legal queries. It is nowhere near the “hallucination-free” language the marketing used.
The study’s subtler finding matters more than the headline number. The tools failed in two ways: stating the law wrong, and stating it right while citing a source that does not support the claim. That second failure, the misgrounded citation, is the dangerous one for a lawyer, because the cite looks real, the case exists, and only reading it reveals the mismatch. Why models make things up is its own explainer, but the practical version is one line: a language model is a guessing machine, and retrieval narrows its guesses without ending them.
It helps to see why grounding falls short in legal work specifically. Retrieval fetches documents that look relevant, and law is a domain where looking relevant and being controlling are different things: a case can match every keyword and still be from the wrong jurisdiction, overruled, or distinguishable on its facts. The model then summarizes whatever was fetched with the same fluent confidence either way. Grounding fixed “the case does not exist.” It did not fix “the case does not say that,” and the second error is harder to catch.
Verify every citation as if opposing counsel wrote it
The profession learned this the loud way. In June 2023, Judge P. Kevin Castel fined two New York lawyers $5,000 and dismissed their client’s case after they filed a brief citing precedents ChatGPT had invented outright, fictitious airlines and all. When challenged, they had asked ChatGPT whether the cases were real, and it assured them they could be found in reputable legal databases. The judge called the resulting legal analysis “gibberish.”
Mata v. Avianca was supposed to be the cautionary tale that ended the practice. Instead it was the first entry in a database. By May 2026, a tracker maintained by legal researcher Damien Charlotin had logged more than 1,300 court decisions worldwide commenting on AI-invented authority in filings. The penalties have escalated with the count. An Oregon federal court ordered two lawyers to pay a combined $110,000 after briefs with 15 nonexistent cases and 8 fabricated quotations, conduct the judge called “a notorious outlier in both degree and volume.” In Alabama, a court dismissed a family’s appeal over their lawyer’s fake citations. The clients lost the case, not just the lawyer’s fee.
The fix is not a better model. It is a rule you never break: every authority in anything you sign gets pulled from Westlaw, Lexis, or the reporter and read before filing. Not confirmed to exist. Read. The misgrounded-citation failure means an existence check is not enough; the case must actually say what the brief claims it says. Treat AI output exactly like a first-year associate’s memo, because that is what it is: fast, useful, confident, and unsigned. Rule 11 does not have an AI exception, and judges have stopped treating “the software did it” as mitigation.
A consumer chatbot can waive privilege
Hallucination embarrasses you in public. The confidentiality failure is quieter and worse, because it can hurt the client directly.
The Heppner ruling spelled out the mechanics. Privilege protects communication with a lawyer; a chatbot holds no license and owes no loyalty. Confidentiality requires a reasonable expectation of privacy; the consumer terms Heppner used allowed his prompts to be reviewed, used for training, and shared. Feeding a defense strategy into that, the court held, is disclosure to a third party, and disclosure waives protection. OpenAI’s chief executive Sam Altman had said the same thing plainly months earlier: there is no legal privilege in a ChatGPT conversation, and the company can be compelled to produce user chats in litigation.
The ruling left one door open, and it is the door that matters for practitioners. Judge Rakoff noted that if counsel had directed the AI use, the tool might arguably have functioned as the lawyer’s agent, the way privilege already extends to a retained expert or a paralegal. Structure matters: the same query, run inside a lawyer-supervised workflow on properly contracted systems, sits on much safer ground than a client or lawyer freelancing in a consumer app.
Now run the same logic from the other side of the desk. A lawyer who pastes a client’s merger terms, medical records, or litigation strategy into a consumer chatbot has sent privileged material to a third party whose systems may retain it, train on it, and surrender it under subpoena. Courts have already ordered AI vendors to preserve user conversations wholesale in unrelated litigation, a story covered in what you actually own in SaaS AI. The convenience is real. So is the trail.
Confidential work belongs on hardware you control
The bar’s guidance points the same direction. ABA Formal Opinion 512, the first national ethics opinion on generative AI, issued July 29, 2024, ties AI use to the duties lawyers already carry: competence requires understanding what the tool does with your inputs, confidentiality requires safeguarding client data inside it, and supervision makes you responsible for the nonlawyer assistance you employ. In practice that sorts AI setups into three tiers.
Consumer chatbot accounts are the bottom tier: fine for public-facts work, never for client material. Enterprise legal AI with a signed agreement is the workable middle: a contract that excludes training and limits retention is what makes the vendor a proper agent instead of a stranger. The top tier is the one most lawyers still have not priced in: open-weight models running locally, on a machine the firm owns, where the matter file never crosses the wire at all. Capable open models now run on an ordinary workstation, and for a solo practitioner the entire setup costs less than one month of a research platform seat.
Local is not automatically better lawyering; a frontier cloud model with a real enterprise agreement is stronger on hard reasoning than anything you can run on a desktop today. But for the material where disclosure itself is the harm, privileged strategy, unfiled deals, anything covered by a protective order, the strongest confidentiality control is the one that never transmits. Regulators are converging on the same instinct, which is why regulated industries keep pulling AI on-premise.
Start where mistakes are cheap
Distilled to a card you could tape to a monitor, the whole post is three rules. AI drafts, you sign: every authority pulled and read before it carries your name. Client material only goes to systems you have a contract with or hardware you own; a consumer chat window is neither. And nothing the model says is advice until a lawyer has made it theirs.
Then sequence the adoption by the cost of an error. Start with internal summaries, plain-English client letters, and first drafts of low-stakes documents, where a mistake costs an edit. Research assistance comes next, protected by the verification rule. Filed work product comes last, and only after the checking habit is automatic. Worked this way, the technology is exactly what it looks like: the most productive associate you have ever hired, one who works nights, bills nothing, occasionally lies with total confidence, and tells anyone who asks. Staff accordingly.
Disclaimer: This article is general information about technology, not legal advice, and reading it creates no attorney-client relationship. Cases, sanctions, ethics guidance, and AI tool behavior described here reflect sources available as of July 2026 and may have changed. Consult your own counsel, your jurisdiction’s bar guidance, and your court’s standing orders before relying on AI in practice.
