Legal AI Adoption Reaches 49% as Courts Raise the Stakes

A milestone with an asterisk

Generative AI is no longer a pilot project in legal practice. It is now a routine tool, though one that is unevenly governed and increasingly risky. That is the main finding of Everlaw's 2026 Legal AI Adoption & Impact Report, released September 24. The report was produced with the Association of Certified E-Discovery Specialists and the International Legal Technology Association, and it draws on more than 250 legal professionals.1 Its central number is that 49% of respondents now actively use generative AI in their work. That is 12 percentage points higher than last year and 15 points higher than in 2024.1

The report arrived during an unusually busy few weeks for AI and the law. Five days later, the Third Circuit issued the first federal appellate ruling on whether training AI on copyrighted material can be fair use. The case involved a legal research tool.16 Courts are also punishing lawyers who file AI-fabricated citations more harshly than before.46 Taken together, these stories show a profession that is adopting AI faster than its rules, its training, and its liability framework can keep up.

What the survey actually shows

Everlaw presents the shift as a sprint compared with the long, slow move to cloud ediscovery. The company says generative AI reached about the same adoption level as cloud ediscovery software in roughly four years after ChatGPT's release.2 Daily use is rising. Twenty-one percent of respondents use generative AI several times a day, up 11 points from last year, and another 19% use it at least once a day.2 The share reporting weekly savings of five to 10 hours more than doubled.1

The business model is starting to feel the effects. Nineteen percent say generative AI has already significantly changed conventional billing, up eight points. Another 31% expect a significant change within a year, up 11 points.2 Disclosure practices are split. Thirty-one percent tell clients about AI use for all substantive work, while 26% do not because they treat AI as an internal tool.2

The sharpest criticism came from Above the Law, which called the report a "glass half full" reading and laid out the half-empty version. The columnist pointed out that 49% is still less than half, so most respondents are not using the technology at all. Of the 46% who think generative AI will soon be standard, it is unclear whether that share covers all respondents or only current users. The column also argued that the 42% who "plan" to adopt should not count toward momentum.3 The time-savings figure looks smaller in that breakdown too: 17% of users report saving five to 10 hours a week, and 21% are not sure AI saves them any time.3

The biggest gap is on agentic AI. Everlaw's announcement says nearly a third of respondents are piloting multi-agent systems.1 Above the Law's reading of the same survey found that only 3% currently use AI agents.3 Both can be true, since piloting and real deployment are different things. But that difference is the gap between hype and practice. The fairer reading is that agentic AI in legal work is still mostly at the experimental stage.

Governance is the real story

The most important finding has nothing to do with productivity. It is about who controls how AI gets used. One-third of respondents say their organizations have no formal generative AI policy.1 Above the Law's breakdown shows how thin governance is. Only 48% have written policies. Twelve percent have none, 21% are still planning one, and 20% have policies that cover employees but apparently not partners. Just 34% are rolling out generative AI across the whole organization.3

Other data suggests the problem may be worse than Everlaw's sample shows. Newsweek reported on a separate 2026 survey of more than 1,300 legal professionals. In it, 69% used general-purpose AI for work, but only 9% of firms had a written AI policy that was actually enforced.8 Different samples and definitions make exact comparisons unreliable. Still, both surveys point the same way: use is running ahead of oversight.

Output verification is where weak governance turns into liability. Only 41% of Everlaw respondents named output verification as a key future AI skill, and 14% said no AI skills are needed at all.3 The courtroom record shows the cost of that attitude.

Courts turn verification into a liability rule

A public database run by researcher Damien Charlotin tracks about 1,490 decisions worldwide in which a party relied on AI-hallucinated material and the court responded. More than 1,000 of those are in the United States.4 Reuters reported in September that problem filings keep increasing even after three years of sanctions.6

Penalties have risen accordingly. In March 2026, the Sixth Circuit ordered each of two lawyers in Whiting v. City of Athens to pay $15,000. They also had to cover the other side's appellate fees and double costs, and were referred for discipline.4 The court sanctioned the fake citations without ever confirming whether AI had produced them.6 In June, the Ninth Circuit suspended two immigration lawyers from practicing before it for six months. It held that the violation happens "at the point of signing and filing," not during research.9 Illinois's appellate court fined a lawyer $1,500 for each false citation or misquotation, $15,000 in total.11 Nebraska issued what is reported as the first indefinite suspension tied to AI filings.4

The October cases make clear that expensive tools offer no protection. In Boston, Judge Angel Kelley revoked a California lawyer's pro hac vice status and ordered up to $10,000 in fees. The lawyer had relied on what he called the "highest-grade enterprise version of Claude." Kelley wrote that the use of AI does not reduce an attorney's Rule 11 obligations.7 The trend is international. England's Solicitors Disciplinary Tribunal struck off a registered foreign lawyer in September, and India's Supreme Court has called citing unverified AI authorities misconduct.6

Not every court is cracking down. A Los Angeles judge fined a State Farm lawyer $999.99, one cent below the amount that would have triggered mandatory notice to the California State Bar.12 Even so, the overall direction is clear. Courts are not banning AI. They are making lawyers personally responsible for everything it produces. For the half of the profession that now uses these tools, and the third working without a policy, that is the most immediate risk.

The Ross ruling: a legal-tech case with industry-wide reach

The copyright fight landed close to home for legal tech. On September 29, the Third Circuit upheld Thomson Reuters' win against ROSS Intelligence. ROSS was a now-defunct startup that built an AI legal search tool using material derived from Westlaw's editorial headnotes.1619 After Thomson Reuters refused to license its content to a competitor, ROSS hired LegalEase. That firm produced about 25,000 "Bulk Memos" using the headnotes as a reference, and the trial court found that 2,243 headnotes were infringed.19

Coverage agrees on the core holding. The headnotes are original enough to be protected by copyright. ROSS's use was commercial and only minimally transformative. And it harmed two markets: legal research, and an emerging market for licensing headnotes as AI training data.20 Reports differ on one procedural point. One outlet says the opinion was unsealed September 30.15 A law firm analysis dated October 1 said it was still sealed pending redactions.19

The weight of the ruling depends on how narrowly it is read. Ballard Spahr stresses that the panel distinguished ROSS, which retrieved existing judicial text, from systems that generate new expression. That leaves the central generative AI questions open.20 Fox Rothschild takes a broader view. It argues that recognizing a potential training-data licensing market gives rights holders a theory that reaches "well beyond the legal-research context."19 Music industry groups clearly agree. The RIAA and NMPA backed Thomson Reuters and cited labels' licensing deals with AI companies as evidence that such a market exists.18

Our reading is that both views are partly right, but the market-harm reasoning will matter more than the generative carve-out. The carve-out protects the biggest model builders for now. The licensing-market logic, though, will affect every vertical AI product trained on curated professional content. That includes many legal-tech startups that survey respondents now use every day.

A patchwork of precedents and mandates

ROSS adds to a mixed set of trial-level outcomes. In Bartz v. Anthropic, the court held that training on books was fair use but keeping pirated copies was not. That led to a $1.5 billion settlement that received final approval in July 2026.1623 Kadrey v. Meta went Meta's way because the plaintiffs failed to show market harm.16 In Germany, courts found OpenAI and Suno liable over song lyrics memorized by their models. The Suno court added that complying with EU AI Act transparency rules does not, by itself, excuse infringement.1621

Policy is moving in parallel. Since January 1, 2026, California has required public summaries of training datasets. The EU AI Act has required general-purpose model providers to have copyright policies since August 2025.23

Why it matters

The Everlaw report shows a profession that has already committed to AI.1 Over the past month, courts have defined what that commitment involves. Lawyers personally answer for every citation, whatever tool produced it.79 Vendors answer for where their training data came from.1923 Firms that still treat AI as an unregulated internal tool, without policies, verification training, or disclosure rules, are running the highest risks on both fronts.