Will AI Replace Lawyers? What the Evidence Actually Shows

AI is not on track to replace lawyers, but it is already replacing a large share of the tasks lawyers used to bill for. Research, first drafts, document review and contract comparison are increasingly done by software, while the parts of the job that carry legal responsibility, such as advising a client, signing a filing, negotiating and arguing in court, remain with licensed humans. The rules of the profession point the same way: courts have sanctioned lawyers who filed AI-invented case citations, and the American Bar Association's ethics guidance makes the lawyer, not the tool, answerable for every word.

The evidence also cuts against the "no change" camp. A widely cited 2023 Goldman Sachs analysis estimated that 44% of legal tasks in the United States could be automated by AI, one of the highest exposures of any occupation. Yet the Bureau of Labor Statistics still projects lawyer employment to grow 5% between 2025 and 2035. Both things can be true, and understanding why tells you a lot about where legal work is heading and what it means if you need to hire a lawyer.

What AI already does in law

Software has been creeping into legal work for more than a decade. In 2012, a federal magistrate judge in New York, Andrew Peck, approved the use of computer-assisted document review in Da Silva Moore v. Publicis Groupe, one of the first judicial endorsements of letting algorithms sort millions of documents in litigation. Generative AI, which arrived in the mainstream with ChatGPT in late 2022, widened the range of tasks dramatically.

Today the main uses fall into a few groups:

• Legal research. Tools from LexisNexis and Thomson Reuters, the two dominant legal research companies, now answer questions in plain English and point to supporting cases and statutes.
• Document review and e-discovery. AI classifies emails and files for relevance and privilege far faster than teams of junior lawyers.
• Contract analysis. Software flags unusual clauses, compares drafts against a company's standard terms and summarizes long agreements.
• Drafting. First drafts of letters, memos, discovery requests and routine contracts can be generated in minutes, then edited by a lawyer.
• Summaries and transcripts. Depositions, hearings and case files can be condensed for quick review.

The productivity claims are large. Thomson Reuters' 2025 Future of Professionals report estimated that AI could free up about five hours a week, or roughly 240 hours a year, for each legal professional. For a profession that sells time, that is a structural change, not a gadget.

Can AI pass the bar exam?

In 2023, researchers Daniel Martin Katz, Michael Bommarito and colleagues reported that GPT-4 passed the full Uniform Bar Exam, including the essay and performance sections, with a score above the passing line in every jurisdiction that uses the test; the study was later published in the journal Philosophical Transactions of the Royal Society A. Later critics argued that the headline percentile comparisons were overstated, but the core point stands: the written knowledge tested on bar exams is no longer a moat. Passing an exam, however, is not the same as practicing law, as the next section shows.

When AI gets it wrong: the fabricated citation cases

The most important lesson of the past few years came from courtrooms, not labs. Generative AI systems can produce confident, well-formatted, completely fictitious legal authorities, a failure known as hallucination. Several lawyers learned this in public.

Mata v. Avianca (2023)

In a personal injury suit against the airline Avianca in federal court in Manhattan, lawyers for the plaintiff filed a brief citing court decisions that did not exist. One of the lawyers, Steven Schwartz, had used ChatGPT for research and testified that he was operating under the false perception that the tool "could not possibly be fabricating cases on its own." On June 22, 2023, U.S. District Judge P. Kevin Castel imposed a $5,000 penalty, jointly and severally, on Schwartz, his colleague Peter LoDuca and their firm, Levidow, Levidow & Oberman. The judge stressed that using AI was not itself the problem; the lawyers' failure to check the citations, and to come clean promptly once questioned, was.

Other cases that followed

Case	Court and date	What happened	Outcome
Mata v. Avianca, Inc.	U.S. District Court, Southern District of New York; sanctions order June 22, 2023	Brief cited fake decisions generated by ChatGPT	$5,000 penalty on two lawyers and their firm
United States v. Cohen	U.S. District Court, Southern District of New York; opinion March 20, 2024	Michael Cohen, Donald Trump's former lawyer, gave his attorney citations produced by Google Bard that turned out to be fake	Judge Jesse Furman declined to impose sanctions, finding the episode embarrassing and negligent but not done in bad faith
Wadsworth v. Walmart Inc.	U.S. District Court, District of Wyoming; February 2025	Motions from lawyers at the firm Morgan & Morgan cited cases invented by the firm's in-house AI tool	Fines of $3,000 for the drafting lawyer and $1,000 each for two others

In the Wyoming case, the judge noted that the fake cases lacked pinpoint citations and proper formatting, which a careful reader could have caught, and that the firm had since added a warning that users "must independently verify" AI output. The pattern across these cases is consistent: courts do not punish the use of AI, they punish signing work nobody checked.

Even legal-specific AI tools make mistakes

It is tempting to assume the problem is limited to general chatbots. A 2024 study by Stanford's RegLab and Institute for Human-Centered AI tested the AI research tools sold by LexisNexis (Lexis+ AI) and Thomson Reuters (Westlaw AI-Assisted Research and Ask Practical Law AI). The researchers found that these tools hallucinated less than GPT-4, but still produced incorrect information between 17% and 33% of the time on their benchmark questions. Specialized tools reduce the risk; they do not remove the need for a lawyer to read the sources.

What the rules say: ABA Formal Opinion 512 and court orders

On July 29, 2024, the American Bar Association's Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, "Generative Artificial Intelligence Tools," its first comprehensive guidance on the subject. ABA opinions are not binding law, but state bars and courts rely on them heavily. The opinion does not ban AI; it maps existing duties onto it.

• Competence. Lawyers must understand the capabilities and limits of any AI tool they use well enough to use it responsibly, and must check its output to an appropriate degree.
• Confidentiality. Before putting client information into a tool that may learn from or retain it, a lawyer generally needs the client's informed consent. The opinion says boilerplate language in an engagement letter is not enough.
• Communication. In some situations lawyers must tell clients they are using AI, for example when it affects the representation or when a client asks.
• Candor to the court. Lawyers must verify that citations and legal arguments are real and sound before filing them.
• Supervision. Managing lawyers must set policies for AI use by other lawyers and staff in their firm.
• Fees. Lawyers who bill by the hour may charge only for time actually spent, so the time AI saves cannot be billed as if the work were done by hand, and general time spent learning a tool they will use routinely should not be charged to the client.

Courts have added their own requirements. On May 30, 2023, U.S. District Judge Brantley Starr of the Northern District of Texas began requiring lawyers to certify either that no generative AI was used in their filings or that any AI-drafted text had been checked by a human. That court later adopted local rules requiring disclosure of generative AI use in filings, and many other judges and courts have issued similar orders.

Which legal work is most exposed to AI

Studies of "jobs most threatened by AI" consistently put legal work near the top, but exposure is uneven within the profession. The Goldman Sachs estimate placed legal work second only to office and administrative support, at 44% of tasks automatable, while physically demanding fields such as construction sat at 6%. Exposure is a measure of tasks, not jobs: it says how much of the work AI could do, not how many people will be let go.

Type of legal work	Exposure to AI	Why
Document review and e-discovery	Very high	Repetitive classification at scale, already automated for years
Routine contracts and forms	High	Standardized language that AI drafts and compares well
Legal research memos	High, with verification	AI finds and summarizes authority quickly but can invent it
Transactional deal work	Medium	Drafting speeds up; structuring and negotiating stay human
Client counseling and strategy	Low	Depends on judgment, trust and accountability
Trial advocacy and courtroom work	Low	Requires a licensed advocate, live persuasion and ethical responsibility

The pressure falls hardest on entry-level work. Much of what junior associates and paralegals have traditionally done, such as reviewing documents and producing first-draft research, is precisely what AI does best. That raises a long-term question the profession has not answered: if beginners no longer learn by doing the routine work, how will they gain the experience to do the hard work later? Corporate and transactional lawyers will feel the speed-up sooner than trial lawyers, but none of these roles disappear outright.

What stays human in legal work

Several features of law make full replacement unlikely in any near-term horizon:

• Licensing and accountability. In the United States, only licensed lawyers may practice law, and someone must sign each filing and answer for it. An AI system cannot be disbarred, held in contempt or sued for malpractice the way a lawyer can.
• Judgment under uncertainty. Deciding whether to settle, which argument to lead with, or how much risk a client can tolerate depends on facts, people and priorities that rarely appear in a prompt.
• Fact-finding. Many cases turn on facts nobody has written down yet: interviewing witnesses, spotting inconsistencies, knowing what to ask for in discovery.
• Persuasion and negotiation. Reading a judge, a jury or an opposing lawyer in real time is a social skill, not a retrieval task.
• Trust and confidentiality. Clients share their worst moments with lawyers under the protection of privilege. That relationship does not transfer to a chatbot.

The labor statistics reflect this. The Bureau of Labor Statistics counted 863,700 lawyer jobs in 2025 and projects about 904,300 by 2035, with a median annual wage of $159,670 in May 2025. That projection could shift as AI spreads, but nothing in the official data so far suggests a collapse in demand for lawyers.

Will AI replace lawyers in 10 years?

A realistic ten-year picture looks less like replacement and more like reorganization. Firms will likely need fewer hours for routine work, which puts pressure on the billable-hour model and on the size of junior classes. Flat fees and subscription pricing for routine matters may spread. Lawyers who can supervise AI tools, check their output and explain the results to clients will be more valuable, not less. The same pattern shows up in medicine, where people ask whether AI will replace doctors: software is taking over reading, sorting and drafting, while diagnosis, treatment decisions and responsibility for the patient stay with licensed professionals.

The unknown is how quickly reliability improves. If hallucination rates fall to near zero and courts grow comfortable with AI-assisted filings, more routine matters, such as simple wills, uncontested filings or small claims preparation, could shift to self-service tools. Even then, the rules on who may give legal advice would have to change before software could stand in for a lawyer in any formal sense.

What this means if you are hiring a lawyer

For clients, AI is mostly good news, provided you ask the right questions. Faster research and drafting should mean lower costs for routine work, and the ABA's guidance means those savings should not vanish into inflated hourly bills.

1. Ask whether the firm uses AI, and how. A good answer names the tasks (research, first drafts, document review) and explains how a lawyer checks the output.
2. Ask about confidentiality. Find out whether your information goes into tools that retain or train on data, and expect to be asked for consent if it does.
3. Ask about billing. If AI does in minutes what used to take hours, you should not be billed for the hours. Flat fees for routine work are worth requesting.
4. Do not rely on a chatbot for your own case. People representing themselves have also filed fake AI citations. If you use AI to understand your situation, verify every legal claim against official sources or with a lawyer before acting on it.
5. Value judgment over speed. The cheapest draft is worthless if the strategy is wrong. You are paying for the lawyer's judgment and accountability, not their typing.

For more on how technology is reshaping everyday life, explore our tech section, including our guides to passkeys and passwordless sign-in and how to tell if your phone is being monitored.