Legal AI Adoption Hits Every Big Law Firm as Court Risks Mount

The profession that resisted technology now can't get enough of it

Lawyers have a long reputation for resisting new technology, and generative AI has broken that pattern. Bloomberg Law's Leading Law Firms survey found that all 40 firms with at least 500 attorneys that broke down their technology use reported using legal-specific AI tools in 2025. That figure is slightly higher than the year before, and it puts AI on the same footing as long-established attorney software.1 Firm leaders describe the change as coming from lawyers themselves. Katrina Dittmer of Eversheds Sutherland said that for years, getting attorneys to adopt technology felt like pushing a rock uphill, and that now lawyers ask to join AI pilots.1 Norton Rose Fulbright said 80% of its attorneys have adopted legal-specific AI tools. Its managing partner noted that most clients now expect AI to be part of the work.1

The wider industry data shows the same jump. The 8am 2026 Legal Industry Report, which surveyed more than 1,300 legal professionals, found that 69% now use general-purpose generative AI for work, up from 31% the year before.56 The report says reaching most practitioners took three years, not the decades earlier legal technologies needed.5 Clio's surveys found adoption at 71% of solo practitioners, 75% of small firms and 86% of mid-sized firms.7 Litify's report put overall adoption at 78%.4

Where the numbers disagree, and what that reveals

The surveys do not fully agree, and the gaps matter. Bloomberg Law reports that every responding large firm uses AI. 8am's broader sample found that only 46% of firms have rolled out general-purpose AI tools and only 34% have adopted legal-specific ones, though that 34% was up from 21%.56 Firm size explains most of the difference: among firms with more than 20 lawyers, the general-purpose figure rises to 58%.5 Earlier research also found that firms with 51 or more attorneys used AI at about twice the rate of smaller firms, likely because enterprise systems are expensive.10

The more important gap is between individual lawyers and their firms. 8am reports that 54% of firms offer no AI training and 43% have no AI governance policy. It concludes that many lawyers are likely using AI without their firm's approval.6 The North Carolina Bar Association added that only 9% of respondents said their firm has a written AI policy that is actually enforced.7 Big firms look better on this measure: Bloomberg Law found that, on average, 78% of attorneys at firms with more than 500 lawyers had completed AI training.1 Even so, adoption among individual attorneys at large firms still trails mature tools like e-billing. One BCLP leader said his firm's own vetting and governance work, not lawyer reluctance, is what slows things down.1

Put together, the data does not show a profession adopting AI carefully. Enthusiasm is ahead of oversight almost everywhere, and the biggest firms are the ones closest to catching up.

The courts are pricing in the risk

That gap is already costing lawyers in court. A public database kept by researcher Damien Charlotin listed 2,184 decisions worldwide involving AI-fabricated material as of October 10, 2026, including 1,492 in U.S. courts.23 One count of that data found that U.S. decisions involving lawyers rose from 14 in 2024 to 207 in 2025 and 361 so far in 2026.23 Another tracker reported about 2,046 cases in September, up from roughly 200 a year earlier.28 The totals vary with when and how each tracker counts, but all of them show steep growth.

The penalties are getting heavier. In Whiting v. City of Athens, the Sixth Circuit ordered two attorneys to pay $15,000 each, cover the other side's appellate fees and pay double costs, and referred them for discipline. The panel said lawyers must personally read and verify every citation they file, whether it came from AI or anywhere else.2225 In June, Judge Sharion Aycock in Mississippi sanctioned all four lawyers on both sides of a civil case and canceled the trial. Two of the lawyers were barred from the Northern District of Mississippi for two years.30 One of them said she had not known AI could invent cases.30 Nebraska reported the first indefinite bar suspension tied to AI filings, after a brief in which 57 of 63 citations were defective.22 A combined sanction of about $109,700 in an Oregon winery dispute still stands as the record.28

Premium tools don't prevent these failures. Bloomberg Law notes that Sullivan & Cromwell apologized to a bankruptcy judge after filing documents full of errors.1 In Farris, a lawyer admitted using Westlaw's CoCounsel to draft both briefs. The Sixth Circuit warned that the risk remains even with tools from trusted legal technology providers.23 Lawyers keep repeating the mistake despite years of warnings. Scientific American described an Alabama case in which a lawyer promised to stop citing fake cases and then cited nonexistent cases in the very next sentence.21 The rule the courts are applying is consistent: using AI is not sanctionable, but filing output nobody checked is.24

Legal AI sits at the center of the copyright fight

The legal industry is also where AI copyright law is being written. On September 29, the Third Circuit upheld the ruling that ROSS Intelligence infringed Thomson Reuters' copyrights when it used Westlaw headnotes to train a competing legal research tool. It was the first federal appeals court to directly decide whether AI training can be fair use.14 The court found that ROSS competed with Westlaw for the same customers. It also held that the developing market for licensing content as AI training data deserved protection, and pointed to Thomson Reuters' own use of its headnotes in AI search.14

How far the ruling reaches is the main point of disagreement in the coverage. One advisory firm argues that courts in 2025 and 2026 have effectively ended the fair-use defense for AI training.20 Most other analyses read the decision more narrowly, and so do we. The court stressed that ROSS was a non-generative system that could only retrieve existing opinions, and it expressly left open whether training generative AI could be fair use.1417 Commentators also note that ROSS copied headnotes nearly word for word, which makes the case easier to distinguish.11 The ruling also does not bind the Ninth Circuit, where most generative AI cases are being heard.14

The district courts that have ruled on generative models mostly point the other way. In Bartz v. Anthropic, the court held that training on lawfully obtained books was fair use but that storing pirated copies was not. The case then settled for $1.5 billion, about $3,000 per work.13 Final approval came on July 20, 2026, but approving a settlement does not decide whether the conduct was lawful.12 In Kadrey v. Meta, the court ruled for Meta on fair use for training, while claims over seeding pirated files remain alive.1316 The consolidated OpenAI cases in New York, including The New York Times' suit, are still in discovery, and OpenAI has been ordered to produce tens of millions of output logs.16 New cases keep coming, including publishers suing Google over Gemini training.16

The pattern across these cases is fairly clear. Courts are separating three questions: how the data was obtained, how the model uses it, and whether the product competes with the source.18 Pirated data and direct competition have been the losing facts so far. Training a generative model on lawfully acquired material has not lost at any appeals court.1317

Why the two stories are really one

Adoption, sanctions and copyright usually get covered as separate stories, but they are linked. Thomson Reuters won a landmark copyright case protecting the editorial content behind its research products.14 Its CoCounsel tool has also appeared in sanctions cases.23 The ROSS ruling strengthens publishers and database owners who license their content for AI. That favors established vendors at a time when firms are sticking with legal-specific tools built into software they already use. 52% of firms that adopted legal AI cite that integration as their main reason.514

For law firms, there are two practical lessons. First, sourcing legal AI is not legally neutral. The Third Circuit's reasoning raises the risk for any tool trained on a competitor's curated database without a license, so vendor due diligence now includes questions about training data.14 Second, liability is landing on the lawyer who signs the filing, not on the software maker. Courts have consistently held that responsibility for verifying citations stays with the lawyer, whatever tool produced them.23 Many firms keep vendor contracts to one year because the market changes so quickly.1 That flexibility makes sense, but it does not reduce the lawyer's personal exposure.

The business pressure keeps building alongside the legal risk. AI is squeezing the billable hour and raising questions about how many associates firms will hire.1 Only about a third of solo and small firms report revenue gains from AI.7 In our view, the profession has finished deciding whether to use AI. Its remaining question is whether its oversight can keep up with how fast lawyers are using it, and at the moment, sanctions are growing faster than enforced policies.