The Last Independent Number: Measuring Legal AI Accuracy

Dr. Angela Liu

Seventeen percent. Thirty-three percent. Those are the last independent hallucination rates anyone has published for the two AI research platforms most American lawyers use. Stanford measured them in 2024: about 17 percent for Lexis+ AI, and about one query in three for Westlaw’s AI-Assisted Research.

Dr. Angela Liu

That was two and a half years ago. Since then, nobody outside the vendors has measured those platforms again, and the two largest vendors have not appeared in a single public benchmark of legal research accuracy.

Erick S. Robinson

Meanwhile, Damien Charlotin’s database of court decisions addressing AI-hallucinated material went from about 120 entries in May 2025 to 2,184 on October 10, 2026.

Erick S. Robinson

Error rates at the bottom of the stack fell. Consequences at the top of the stack rose. Both facts are true, and a firm that plans around only one of them is going to get hurt by the other.

Erick S. Robinson

Welcome to the AI Law Podcast. I am Erick Robinson, a patent trial lawyer and a partner at Cherry Johnson Siegmund James.

Dr. Angela Liu

I am Dr. Angela Liu.

Erick S. Robinson

This episode walks through an article I just published on The AI Law Blog called “The Last Independent Number.”

Dr. Angela Liu

The question it answers is easy to ask and hard to answer. Have Westlaw, Lexis, Harvey, and the other legal AI tools actually become more accurate since Stanford’s 2024 hallucination study? What problems are lawyers really having with them? And what should the profession do about it?

Erick S. Robinson

Two disclosures before we start. First, I build AI tools for my own firm, and I use several of the products I am going to name today. Second, everything here is current as of October 12, 2026, and this is education, not legal advice. Where the sources disagree, we will tell you.

Erick S. Robinson

I will give you the answer first. There are six conclusions, and they organize the rest of this episode.

Dr. Angela Liu

First, a hallucination rate is a property of the test as much as the product. Change the definition, the questions, or the checker, and the number moves more than product quality ever moves it.

Dr. Angela Liu

Second, the raw models got dramatically better at not inventing law, and only modestly better at everything else.

Dr. Angela Liu

Third, strict grading exposes a wall. Models that satisfy 94 percent of the individual requirements on a long legal task finish only 20 to 25 percent of those tasks without a miss.

Erick S. Robinson

Fourth, the most dangerous errors are no longer fake cases. They are real cases cited for things they do not say, invented facts about your own record, and legal judgment that is simply wrong while looking like boilerplate.

Erick S. Robinson

Fifth, the legal-specific wrappers, meaning Harvey, CoCounsel, Protégé, Legora, and the rest, earn their price through licensed content, citators, connectors, and workflow. They do not earn it through better reasoning.

Erick S. Robinson

Sixth, lawyers using AI are faster and more polished, but not measurably more accurate, and the junior lawyers who lean on it most appear to retain the least.

Erick S. Robinson

Trust has to be built into the workflow. It cannot be purchased with the license. Let me show you why.

Dr. Angela Liu

“Have the percentages gone down?” is a trick question, for three reasons. Each one explains why the vendors and their critics keep talking past each other.

Dr. Angela Liu

The first reason is that “hallucination” means at least two different things. The narrow meaning is fabrication: the case, the statute, or the quotation does not exist. The broad meaning, the one Stanford used, adds what researchers call misgrounding: the source is real, but it does not support the proposition it is cited for. When LexisNexis promised “hallucination-free linked legal citations,” it was promising the first thing. When Stanford reported 17 percent, it was measuring both. Those two statements can be true of the same product on the same day.

Erick S. Robinson

That distinction matters more now than it did in 2024, because the two errors have moved in opposite directions. Retrieval grounding, linked citations, and citators are built to catch fabrication, and they catch most of it. Misgrounding survives all three. A brief that cites a real Federal Circuit opinion for a holding the court never reached passes every existence check anyone has ever built.

Dr. Angela Liu

The second reason is that the yardstick moves the number. A 2026 study of Ukrainian legal questions scored the same answers from four commercial models against two reference databases, one sparse and one dense. Against the sparse one, 15 to 21 percent of the citations looked hallucinated. Against the dense one, 98.9 to 99.9 percent of the very same citations were grounded. All 54 citations the sparse database flagged turned out to be real. That is a 100 percent false-positive rate. The models did not change. The reference did.

Dr. Angela Liu

The same thing happens when AI checks AI. This month, when Artificial Analysis added a hallucination audit to Harvey’s Legal Agent Benchmark, it first compared six candidate checker models on identical deliverables. The most aggressive checker upheld 470 material hallucinations. The least aggressive upheld 28. The work product was the same, and the count swung by a factor of almost seventeen depending on who did the counting.

Erick S. Robinson

So any vendor that hands you a citation-validity rate owes you three facts: the database it checked against, the version of the product it tested, and whether a human or a model did the checking. Without those three facts, the number cannot be compared to anything, including the vendor’s own number from last year.

Erick S. Robinson

The third reason is that the stack has layers. A frontier model, a research platform built on licensed content, a workflow wrapper like Harvey, an agent working a twenty-step assignment, and a lawyer using any of them are five different things. The courts are a sixth layer. They do not measure accuracy at all. They measure consequences, and that measurement has moved fastest.

Dr. Angela Liu

Now the baseline. Stanford’s RegLab and its Institute for Human-Centered AI built a preregistered set of 202 queries, including 22 false-premise questions, and had human reviewers grade every answer for correctness and for whether the cited sources actually supported what the tool said. Lexis+ AI was accurate about 65 percent of the time and hallucinated on about 17 percent of queries. Westlaw’s AI-Assisted Research was accurate about 41 to 42 percent of the time and hallucinated on about one query in three. GPT-4, with no legal database at all, hallucinated on 43 percent.

Dr. Angela Liu

One detail from that paper matters more today than it did then. Westlaw’s answers averaged 350 words, and Lexis’s averaged 219. The authors pointed out that longer answers contain more falsifiable propositions, so they carry a greater chance of at least one hallucination, and they take longer to check. Hold onto that. It is the first statement of the arithmetic that now governs agentic legal AI, and we will come back to it.

Erick S. Robinson

The vendors answered with definitions, not data. LexisNexis said the study authors had “taken a wider position on what constitutes a hallucination.” Thomson Reuters said its “thorough internal testing of AI-Assisted Research shows an accuracy rate of approximately 90% based on how our customers use it.” Nobody outside Thomson Reuters has seen the questions, the rubric, or the definition of accuracy behind that 90 percent.

Erick S. Robinson

Since then, the vendors have described process instead of results. LexisNexis explained that each Protégé prompt is checked across five stages, with no error rate attached. In January 2026, the Thomson Reuters Institute published a 22-page report on managing hallucinations in courts. It said “AI must serve as an assistant, not an arbiter.” It reported no accuracy figure for any Thomson Reuters product.

Erick S. Robinson

For whatever one lawyer’s experience is worth, a little over a year ago I ran a trial of CoCounsel. On my matters, it performed well below the free version of ChatGPT.

Dr. Angela Liu

Then came the withdrawals. When Vals AI published its first legal benchmark in February 2025, the report stated that LexisNexis had withdrawn from everything except legal research and therefore did not appear. When Vals published its legal research study in October 2025, neither Thomson Reuters nor LexisNexis participated, and vLex withdrew.

Dr. Angela Liu

The later studies that did touch these platforms are not reassuring. A peer-reviewed study run between December 2024 and April 2025 coded 20 percent of Lexis+ AI’s answers as hallucinated, on a different and smaller question set. In 2026, Stanford’s Daniel Ho and colleagues tested the statutory-survey features of both platforms on 1,647 questions about state unemployment insurance law. Westlaw reached 58 percent. Lexis reached 64 percent. A basic retrieval baseline scored 70. Lexis found only 29 percent of the provisions that actually existed. A purpose-built academic system scored 83 percent, then 92 percent once the researchers discovered that the Department of Labor’s own answer key had left out valid statutes. That last detail is worth pausing on. The human gold standard was wrong often enough to move the score nine points.

Erick S. Robinson

The vendors’ best argument is that their products change monthly, so a 2024 test describes a product that no longer exists. That is true. It is also the strongest argument for continuous independent testing, not against it. Stanford’s dataset is public, with a random half held back as a clean test set. Rerunning it takes access to the current products and careful grading, and it takes nothing else. In more than two years, no one with that access has published the result.

Erick S. Robinson

**Bottom line:** the 17 and 33 percent figures describe 2024 products. They are still the last independent numbers because the vendors have chosen not to replace them.

Dr. Angela Liu

Now the good news, and lawyers should not pretend it is not real. In a study published in 2024, Stanford’s “Large Legal Fictions” project asked general-purpose models direct, verifiable questions about randomly selected federal cases. The models hallucinated between 58 percent of the time for ChatGPT 4 and 88 percent of the time for Llama 2.

Dr. Angela Liu

Since then, three independent lines of evidence agree that fabrication has fallen steeply. Linklaters put 50 hard English law questions to the 2023 models and found that in 47 of 150 answers, about a third, the models relied on case law or guidance that was fictitious. In the firm’s February 2025 rerun, fictional citations came back in only 9 percent of queries. OpenAI’s GPT-5 system card reports that on production traffic, the share of responses with at least one major factual error fell from 22.0 percent for o3 to 4.8 percent for GPT-5 thinking. That figure is self-reported, but it points the same way. On LegalBench, the open legal-reasoning benchmark, the top score rose from 77.7 percent in April 2024 to 88.56 percent as of October, and Vals now describes that benchmark as largely saturated.

Dr. Angela Liu

That is real progress. It comes with four asterisks, and every one of them maps onto a sanctions case we will get to later.

Erick S. Robinson

The first asterisk is that fabrication fell but did not vanish. Linklaters’ improved Gemini 2.0 cited a 2019 Court of Appeal case, “UBS AG v Rose Capital Ventures Ltd,” for a proposition about reasonable endeavours. That case does not exist. Nine percent is not a rounding error. A practice group that asks the tool a hundred research questions a month would, at that rate, receive about nine answers resting on authority that does not exist. Linklaters’ conclusion held across both rounds: there are “real risks to using them if you don’t already know the answer.”

Erick S. Robinson

The second asterisk is that progress is not a straight line. OpenAI’s own system card for o3 and o4-mini, released in April 2025, showed the newer models hallucinating more than their predecessor on a test called PersonQA: 33 percent for o3 and 48 percent for o4-mini, against 16 percent for o1. The same thing shows up in legal benchmarks. On Vals’ CaseLaw benchmark, the older GPT-4.1 outscored the newer GPT-5, 69.88 percent to 66.45 percent. A firm that validated a workflow on one model version has not validated it on the next.

Dr. Angela Liu

The third asterisk is that the models still defer to the user. Stanford’s 2026 AI Index reports that when a false belief was framed as the user’s own, GPT-4o’s accuracy fell from 98.2 percent to 64.4 percent, and DeepSeek R1’s fell from over 90 percent to 14.4 percent. Think about what that means in practice. A lawyer who types a conclusion and asks the tool to find support for it is running a false-premise test on every single query.

Dr. Angela Liu

The fourth asterisk is that knowing the law is not the same as reading the record. Vals explained that a common failure on its CaseLaw benchmark was models “relying more on their general knowledge rather than the specific document context (despite being instructed otherwise).” That is the model that knows the general rule and ignores your record. A model trained on the whole of the internet will always be tempted to answer the question it expected instead of the question your documents actually pose.

Erick S. Robinson

**Bottom line:** the models are much better at not inventing law. They are not much better at resisting a confident user, at reading the record in front of them, or at staying consistent from one release to the next.

Erick S. Robinson

The most important development in legal AI evaluation since Stanford is not a new product. It is a new grading standard.

Erick S. Robinson

The 2026 benchmarks score a deliverable against a rubric of specific yes-or-no criteria, and the deliverable passes only if every criterion passes. That is how a partner reviews an associate’s draft, and it is how a court reviews a brief. Under that standard, the same models that saturate LegalBench fail most long legal tasks.

Dr. Angela Liu

Harvey released its Legal Agent Benchmark, called LAB, in May 2026. It has more than 1,200 tasks across 24 practice areas, graded against more than 75,000 rubric criteria. Harvey’s own initial results were candid. In Harvey’s words, “the frontier models we evaluated complete less than 10% of tasks end-to-end in aggregate,” and the top performer cost roughly 50 dollars a task with more than 20 minutes of latency.

Dr. Angela Liu

Vals now runs LAB on a held-out set. As of mid-October, the leading models satisfy about 94 to 95 percent of the individual criteria but complete only 20 to 25 percent of tasks with every criterion met. The top model completes 25.42 percent. Claude Fable 5 completes 11.25 percent, Claude Opus 5 completes 6.67 percent, GPT-6 Astra completes 5.42 percent, and GPT 5.5 completes 3.75 percent.

Dr. Angela Liu

On October 8, Artificial Analysis, working with Harvey, released version 1.1 of its independent implementation of the benchmark. Every deliverable is now audited against the task’s own source documents, and a task earns no credit if it contains a single material hallucination. The evaluators define that as a claim that would mislead a reader on a substantive point: a wrong party, amount, date, or obligation, or a fact that changes a conclusion. Under that gate, the best model scores 9.4 percent. Without the gate, the leader would have scored 26.7 percent, but two thirds of those passes contained a material hallucination. Across the models tested, more than 60 percent of otherwise passing results contained one, and several models scored zero.

Erick S. Robinson

Two findings from that release deserve more attention than they will get. First, finishing the work and not hallucinating are different skills. The model that passed the most rubric criteria, 96 percent of them, averaged 1.68 material hallucinations per task. The model with the fewest hallucinations averaged 0.03 per task and still passed fewer than one task in ten. Second, the hallucinations were not exotic. They were wrong numbers, wrong dates, and statements that something had happened when the source said it still needed to be confirmed. Those are the errors that end up in a closing memo, a damages model, or a declaration.

Dr. Angela Liu

Legal research looks the same under strict grading. On Vals’ Legal Research Bench, Claude Opus 5 earns 90.58 percent of the weighted rubric points but answers only 55.29 percent of the questions entirely correctly, and that is tied for first. Vals explains why it reports the strict number, and the explanation repays quotation: “in legal work, a partially correct answer can be more dangerous than a wrong one: it may read as sound while omitting a critical point.” In the Netherlands, a public benchmark called DELTA tested 23 models on research assignments built from practitioner work. The best model passed 15 percent of tasks while satisfying 75.7 percent of the criteria. The benchmark’s conclusion is blunt: “No model is ready to conduct Dutch legal research independently.”

Dr. Angela Liu

None of this is mysterious. It is arithmetic. If a deliverable contains fifty checkable claims, and each claim is independently 95 percent likely to be right, the chance that all fifty are right is about 8 percent. At 99 percent per claim, it is about 61 percent. Real model errors are correlated, so real models do better than that math, but the principle holds. Reliability per step does not translate into reliability per deliverable, and the longer the deliverable, the worse the translation. Stanford saw the early version of this in 2024, when the tool with the longest answers had the highest hallucination rate.

Erick S. Robinson

There is one hopeful finding, and it comes from Harvey’s own analysis of what separated passing runs from failing ones. Agents that added a validation step after drafting scored about 0.8 points higher. Agents that revised after checking scored about 1.5 points higher. Agents that drafted without any later validation scored about 1.2 points lower. The machines are learning to cite-check their own work. The lesson for lawyers is the one every senior associate learned the hard way. The draft is not the product. The checked draft is.

Erick S. Robinson

**Bottom line:** assume every long AI work product contains at least one error until someone has ruled it out. On the current evidence, that assumption is correct more often than not, even for the best models.

Erick S. Robinson

This is the part of the episode I most want litigators to hear. The fabricated case is now the easier error to catch. Three harder errors have taken its place.

Dr. Angela Liu

The first is the real case cited for support it does not give. A 2026 study presented at the ICML AI for Law workshop tested fourteen model configurations on controlled citation corruptions. When the researchers swapped in the wrong case entirely, the models caught it 93 to 100 percent of the time. When they kept the right case and changed the pinpoint, the models caught only 37 to 61 percent of the errors on court opinions and 52 to 83 percent on legal briefs. Even GPT-5.4 at high reasoning effort still missed 40 percent of the pinpoint mismatches on court opinions. The reason is the one any appellate clerk would predict. The models accept a citation because the topic overlaps, not because the page supports the point.

Dr. Angela Liu

That finding reframes the whole “use AI to check the AI” strategy. Princeton researchers assembled more than 1,000 real filings that contained fabricated citations and built a benchmark of 1,300 brief excerpts with injected errors. GPT-5, working as an agent, caught 84.4 percent of the errors, but only 40.8 percent of what it flagged was actually wrong. On incorrect pincites specifically, the researchers reported that GPT-5 “catches only 18.2% of incorrect pincites.” Automated cite-checking is a reasonable first filter. It is not a substitute for a human being reading the cited page, because the errors it misses are exactly the ones a court will notice, and the false alarms it raises cost review time of their own.

Dr. Angela Liu

The second shift is from invented law to invented facts. The hallucination audit on Harvey’s benchmark did not look for fake cases. It looked for statements that contradicted the source documents, made up their content, or asserted specifics the record did not support. Those are the errors that matter in transactional work, in damages, and in patent litigation, where the most important citations are often not to case law at all.

Erick S. Robinson

That brings me to an order every patent litigator should read twice. On July 27, 2026, the USPTO’s Office of Enrollment and Discipline publicly reprimanded a registered practitioner, in a proceeding called In re Mitchell, for AI-generated errors in a joint claim construction chart filed in a patent case in the Eastern District of Pennsylvania. The errors were not citations to case law. They were citations to the patent’s own specification and prosecution history. The order warns that “the potential pitfalls of AI-generated citation errors are not limited to extrinsic sources such as statutes, regulations, and case law, but may also extend to the intrinsic evidence.” According to opposing counsel’s count, 47 quotations in that chart did not exist, and 26 of them were attributed to the file history. The practitioner had used one AI tool to pull the citations and a second AI tool to verify them.

Erick S. Robinson

Think about what we build in patent cases: claim construction briefs, infringement and invalidity contentions, expert reports, and claim charts. All of it rests on column-and-line citations to the intrinsic record and pinpoint citations to product documentation. A model that is 95 percent reliable per citation will, across a 200-row claim chart, produce errors that look exactly like real evidence. A second model checking the first is not independent verification. It is a second draw from the same kind of process.

Dr. Angela Liu

The USPTO said as much in its 2024 guidance: “Simply relying on the accuracy of an AI tool is not a reasonable inquiry.” The patent research points the same way. In PANORAMA, a NeurIPS 2025 benchmark built from 8,143 U.S. examination records, the best model classified novelty and nonobviousness correctly less than half the time, only modestly better than a random baseline. Unreviewed AI text has already made it into the patent record itself. An Australian patent application published in March 2025 contains this sentence in its background section, and I am reading it exactly: “As of my last knowledge update in September 2021, I do not have access to specific patent numbers.”

Erick S. Robinson

The third error is the one no checker can catch. It is the most dangerous error I have seen this year, and it was not a citation at all.

Erick S. Robinson

I recently asked Harvey, ChatGPT, and Claude to draft the same document: a standstill agreement that would let a patent owner open discussions with an accused infringer without creating declaratory judgment jurisdiction.

Erick S. Robinson

The stakes come from two cases every patent litigator knows. Under the Supreme Court’s 2007 decision in MedImmune versus Genentech, the question is whether the facts show “a substantial controversy, between parties having adverse legal interests, of sufficient immediacy and reality to warrant the issuance of a declaratory judgment.” Under the Federal Circuit’s decision the same year in SanDisk versus STMicroelectronics, jurisdiction arises “where a patentee asserts rights under a patent based on certain identified ongoing or planned activity of another party, and where that party contends that it has the right to engage in the accused activity without license.” The entire purpose of a standstill agreement is to keep the patent owner on the right side of that line.

Erick S. Robinson

Harvey’s draft piled on whereas clauses that added nothing, and one that was worse than nothing: a recital that Party X had accused Party Y of infringing patents A, B, and C. Presenting that draft to the other side would have supplied, in the patent owner’s own words, the assertion of rights that creates the very controversy the agreement existed to avoid. Every citation in that draft could have been real. No citator, no linked-citation check, and no automated hallucination detector would have flagged it, because nothing in it was false. It was simply wrong, and it looked like boilerplate.

Erick S. Robinson

That is the category of error the strict-grading benchmarks are only beginning to measure, and the fabrication-focused safeguards are not built to see it at all. It is why the reviewer has to be someone who knows what the document is for.

Erick S. Robinson

**Bottom line:** check pinpoints, propositions, and facts against the source, not just whether the authority exists, and do not mistake a second model for a second reviewer.

Dr. Angela Liu

Harvey, CoCounsel, Protégé, Legora, and their competitors sit on top of models built by OpenAI, Anthropic, Google, and others. The buyer’s question is what the layer in between adds. The evidence gives a clear answer. It adds licensed content, citators, connectors, and workflow. It does not measurably add better legal reasoning.

Dr. Angela Liu

Start with the independent benchmarks. In Vals’ February 2025 study, Harvey had the top AI score on five of the six tasks it entered, including 94.8 percent on document question-answering against a lawyer baseline of 70.1 percent. That study had no general-purpose model as a control, and it remains the only independent public benchmark ever to score Harvey.

Dr. Angela Liu

Vals’ October 2025 legal research study did include a control. The legal-specific tools scored 78 to 81 percent on accuracy, and ChatGPT scored 80. On the authoritativeness of sources, the legal tools averaged 76 percent against ChatGPT’s 70. Licensed sources bought a modest edge in citing good authority. They did not buy better answers.

Dr. Angela Liu

Legal Benchmarks, an independent evaluator founded by former BigLaw associate Anna Guo, ran 13 AI tools and a panel of human lawyers through 30 real contract-drafting tasks. Legal-specific tools were reliable on 57.6 percent of tasks, general-purpose tools on 58.3 percent, and the human lawyers on 56.7 percent. Guo’s summary this May was direct: “the ‘purpose-built is automatically better’ claim does not survive contact with independent testing.”

Erick S. Robinson

The vendors’ own data, read carefully, says the same thing, and that is to their credit. Harvey reported in February that in its SEC EDGAR research product, lawyers preferred Harvey over GPT-5 with web search 51.73 percent of the time. That is a statistical coin flip. After Harvey rebuilt its data connector, the preference rate rose to 76.7 percent. In Harvey’s Vault product, the share of valid citations ran from 62 percent on one base model to 97 percent on another. The edge came from the plumbing, and it moved with the model underneath.

Erick S. Robinson

Thomson Reuters launched its own model, called Thomson, on August 24 and reported that Thomson working over Westlaw and Practical Law scored 0.83 on factuality, against 0.65 and 0.68 for leading frontier models given unrestricted access to the open web. Look at how that comparison is built. It gives Thomson the Westlaw content and denies it to the rivals. As The Decoder reported from the same data, with web access alone, Thomson scored 0.53 against GPT 5.4’s 0.65. When both had Thomson Reuters content, Thomson edged GPT 5.4 by 0.83 to 0.82. The content did nearly all the work.

Dr. Angela Liu

OpenAI made the same point from the other direction when it released Astra for Law on September 17. OpenAI reports that GPT-6 Astra with its new legal search index passed the overall correctness check on 54.0 percent of questions, against 38.7 percent using web search alone. That is self-reported, but the direction is the lesson: a fifteen-point gain from better sources, on the same model.

Dr. Angela Liu

Harvey’s own help center is admirably candid about limits. It lists use cases “which may produce less reliable results,” and one of them is “Exhaustive ‘scorched earth’ legal research.” For a litigation team, exhaustive research is not an edge case. It is the job.

Dr. Angela Liu

What firms actually experience is more mixed than either the demos or the critics suggest. Harvey announced a 550 million dollar round at a 15.5 billion dollar valuation on September 9 and says that 80 percent of Am Law 100 law firms use Harvey. The International Legal Technology Association’s 2026 survey of more than 500 firms found Microsoft 365 Copilot in use at 76 percent of firms, CoCounsel and Anthropic’s Claude at 44 percent each, Harvey at 43 percent, Lexis+ AI at 26 percent, and Legora at 16 percent.

Dr. Angela Liu

Licenses and use are different things. Harvard’s Center on the Legal Profession describes the product as “there and available” at major firms but “not heavily used yet.” A Harvey-commissioned report in June found that 48 percent of in-house teams required their law firms to use AI, but “only one reported receiving cost savings as a result.” Artificial Lawyer’s Richard Tromans put it plainly on October 12: “Most major law firms use legal AI tools, yet there has been little significant change to how those firms operate.”

Erick S. Robinson

Then there is the first court record I have found that ties Harvey to a filing a judge had to question. On September 22, Judge Arun Subramanian of the Southern District of New York ordered counsel in Hill versus Foundation Media to explain whether certain quotations and a citation in their filing came from “unchecked use of AI.” On September 29, he declined to impose sanctions, called the episode “a wakeup call,” and warned that “over-reliance on AI risks stunting the training of young lawyers.” On October 1, the firm disclosed that it had used Harvey to help prepare the filing. Harvey declined to comment, and nothing in the record says which tool produced which error. The episode illustrates the one risk a wrapper cannot remove, which is the risk that lives in the review.

Erick S. Robinson

I want to be clear about the weight of what comes next. It is the experience of one litigation team, not a benchmark. Several lawyers on my team have watched Harvey return erroneous information and hallucinated authority on a regular basis, including when its answers were grounded in LexisNexis content. Grounding in a licensed database is supposed to prevent exactly that. In our experience, it has not. In the standstill test, three systems got one prompt for one deliverable that a patent litigator can grade in ten minutes. The wrapper was the one that got it wrong, and it got it wrong in the place a wrapper is supposed to add value: the legal framing around the operative terms. The templates, recitals, and house style layered over the model are themselves a source of error. More text is more surface area. That is Stanford’s 2024 finding about long answers, in a new setting.

Dr. Angela Liu

The market has also moved underneath the buyers. In June 2025, LexisNexis and Harvey announced an alliance to bring Lexis primary law and Shepard’s Citations inside Harvey. On February 3 of this year, after Anthropic released a legal plugin for Claude, shares of Thomson Reuters fell 14.2 percent, RELX 14 percent, and Wolters Kluwer 10.5 percent. In August, Thomson Reuters launched its own model. In September, OpenAI began selling Astra for Law directly to selected firms. Harvey itself now routes work across models from several labs, and in August it released Tenet, a post-trained open-weight model built on Moonshot AI’s Kimi K3. That architecture only makes sense if the model is interchangeable and the layer around it is the product.

Dr. Angela Liu

For a buyer, that means the four things bundled into a legal AI subscription, which are model access, licensed content, workflow, and verification tooling, should be priced and tested separately, because each one is now available separately. It also means you should know which base model touched your client’s data, and on whose infrastructure.

Erick S. Robinson

One last point on vendor numbers. Every self-run benchmark crowns its sponsor. In October 2024, Harvey’s own BigLaw Bench reported sentence-level hallucination rates of 0.2 percent for Harvey, against 0.7 percent for Claude, 1.3 percent for ChatGPT, and 1.9 percent for Gemini. Ivo’s benchmark ranks Ivo first. As Anna Guo puts it, “these are vendor benchmarks, and vendor benchmarks tend to make the vendor look good.” The right discount on a vendor benchmark is not zero. It is also not the vendor’s.

Erick S. Robinson

**Bottom line:** buy a wrapper for licensed law, citators, connectors, and workflow, and test it on exactly those things. Then run the same matters through the product and through the model underneath it. If the product cannot beat its own base model on your work, you are paying for packaging.

Dr. Angela Liu

While the benchmarks improved, the court record went in the opposite direction.

Dr. Angela Liu

Damien Charlotin’s database listed about 120 decisions in May 2025. On October 10, 2026, it listed 2,184. Of those, 1,259 involve self-represented litigants, 864 involve lawyers, 32 involve judges, and 15 involve expert witnesses, across more than 40 jurisdictions. An independent analysis of the October 4 export found that decisions dated January through September 2026 ran about 2.7 times the number for the same months of 2025.

Dr. Angela Liu

That count has to be read carefully. It measures adoption and detection, not an error rate, because nobody knows how many AI-assisted filings there are, and most filings with errors are presumably never flagged. More than half the entries involve people without lawyers. Among lawyers, the problem is concentrated in small practices: in a 2025 Stanford analysis, 90 percent of the law firms involved were solo practices or small firms. Still, the lawyer cases include some of the most sophisticated firms in the country, and the patterns in those cases are what matter for practice. There are five of them.

Erick S. Robinson

First, legal-specific tools are in the record. CoCounsel, Westlaw Precision, Lexis+ AI, Protégé, vLex, Paxton, and Harvey all appear in sanctions proceedings or in counsel’s explanations to the court. In United States versus Farris, decided by the Sixth Circuit in April, the principal brief’s file name was “CoCounsel Skill Results,” and three quotations attributed to real sources did not appear in those sources. The court held that the lawyer’s duty to verify “remains true even when new tools are sponsored by trusted legal technology providers.”

Erick S. Robinson

Second, blaming the tool does not work. In Cartagena versus Dixon, in the Southern District of New York, a lawyer blamed LexisNexis’s Protégé. LexisNexis then wrote the court that he “was never an authorized user or subscriber” to either Lexis+ AI or Protégé. In Lnu versus Blanche, the Ninth Circuit suspended counsel from practice before that court for six months after what it called a “repeated failure of candor,” and it stated the governing principle as precisely as any court has: “the rules are not violated at the point of research and drafting, but at the point of signing and filing.”

Dr. Angela Liu

Third, the sanctions are escalating in kind, not just in amount. In 2023, Mata versus Avianca produced a 5,000 dollar penalty. Since then, courts have dismissed claims with prejudice, entered default judgment against the client, revoked pro hac vice admissions and barred lawyers from a district for two years, and suspended lawyers from appellate practice. Sanctions now follow lawyers across jurisdictions. A Morgan & Morgan partner sanctioned in Wyoming in 2025 was denied permission to appear in a Massachusetts matter in 2026 on the strength of that sanction.

Dr. Angela Liu

Fourth, policies are not controls. Sullivan & Cromwell’s April letter to Chief Judge Martin Glenn of the bankruptcy court in the Southern District of New York acknowledged that “The Firm’s policies on the use of AI were not followed in connection with the preparation of the Motion.” The firm’s review process did not catch the errors. Opposing counsel did. Butler Snow had a policy too, and three of its lawyers were reprimanded, disqualified, and referred to the bar in Johnson versus Dunn.

Dr. Angela Liu

Fifth, the client often pays first. In Couvrette versus Wisnovsky, in Oregon, fifteen nonexistent cases and eight fabricated quotations across three summary judgment briefs ended with the client’s claims dismissed with prejudice, a 15,500 dollar penalty, and 94,704 dollars and 38 cents in fees split between lead and local counsel. In Flycatcher versus Affable Avenue, the client lost by default judgment. Malpractice exposure follows, and the insurance may not.

Erick S. Robinson

Patent practice and Texas courts are squarely in this. In Lexos Media versus Overstock, a patent infringement case in the District of Kansas, Judge Julie Robinson found eleven misstatements of authority in a brief opposing the exclusion of the plaintiff’s technical expert. One cited case did not exist, and the court wrote: “Neither this case nor the quotation exists.” All five signing lawyers were publicly admonished, four were fined a total of 12,000 dollars, and one lawyer’s pro hac vice admission was revoked.

Erick S. Robinson

Here in Texas, the Eastern District’s Local Rule AT-3(m) requires lawyers who use these technologies to “review and verify any computer-generated content.” In Gauthier versus Goodyear, Judge Marcia Crone sanctioned a lawyer 2,000 dollars, ordered generative AI training, and required him to give his client a copy of the order, after a brief drafted with Claude cited two nonexistent cases that a Lexis AI checking feature failed to flag. Notice the pattern: one AI drafting, and a second AI missing the problem. In June, in McCormick versus Texakoma Financial, Judge Amos Mazzant publicly reprimanded a lawyer, imposed a 5,000 dollar penalty, required her to review and certify every filing she had signed in 2026, and required a verification certificate on her filings in that court for a year. The Tyler Division’s April 2025 standing order requires a certificate of generative AI usage with every filing. The Beaumont Division’s July 2026 standing order says it as plainly as it can be said: “The fact that AI was used to create the offending document will not excuse or absolve ethical infractions.”

Dr. Angela Liu

The judges are in the database too, all 32 of them. In 2025, two federal judges withdrew orders after errors were traced to a clerk’s use of Perplexity and an intern’s use of ChatGPT. An Oklahoma state judge admitted to investigators that a custody order citing at least two nonexistent cases was drafted with ChatGPT. A survey of federal judges published in March 2026 found that 61.6 percent of respondents had used at least one AI tool, and 45.5 percent had received no training.

Dr. Angela Liu

Abroad, England’s Divisional Court found that 18 of the 45 authorities cited in one case did not exist. In April, South Africa withdrew its draft national AI policy over fictitious sources, and the minister explained that “The most plausible explanation is that AI-generated citations were included without proper verification.” That is a national policy on artificial intelligence, undone by artificial intelligence, and nobody could ask for a cleaner illustration.

Dr. Angela Liu

The costs are spreading beyond sanctions. In EPIC’s 2026 survey of lawyers’ professional liability insurers, seven of thirteen carriers reported an increase in AI-related claims, and malpractice policies typically do not cover court-imposed sanctions or fines. California’s SB 574, signed on September 30, now requires attorneys to verify “the accuracy of all case and statutory citations” in AI-assisted work. On September 18, the Tenth Circuit proposed a rule requiring a certification that a human “has independently verified the existence of all cited authorities.” In United States versus Heppner, Judge Rakoff held that a defendant’s reports generated with a consumer AI tool, using facts learned from counsel, were neither privileged nor work product, in part because the provider’s terms allowed disclosure and training.

Erick S. Robinson

**Bottom line:** courts treat a fabricated or misrepresented citation as the signing lawyer’s failure to verify, whatever tool produced it. Using a legal-specific product does not change that analysis, and blaming one makes it worse. The consequences now reach clients, careers, and coverage.

Dr. Angela Liu

The question that matters most to a managing partner is not whether a model hallucinates in a lab. It is whether a lawyer using the tool produces better work than a lawyer without it. The randomized evidence now answers that question consistently. Lawyers using AI are faster and more polished, they are not measurably more accurate, and there is a hidden cost to skill formation.

Dr. Angela Liu

Schwarcz and his co-authors assigned 137 upper-level law students to six realistic legal tasks using Vincent AI, a retrieval-based legal tool, or OpenAI’s o1-preview reasoning model, or no AI at all. Productivity rose by 50 to 130 percent on five of the six tasks. In the pooled results, though, the authors found no evidence that either tool improved accuracy. Across 768 completed tasks, graders found 18 hallucinated citations: 11 in the o1-preview group, 3 in the Vincent group, and 4 in the group with no AI at all. Retrieval grounding did roughly what the vendors claim on fabrication. It did nothing measurable for legal accuracy.

Erick S. Robinson

The study closest to patent practice is a preregistered field experiment by Autor and six co-authors, released as NBER Working Paper 35720 in September and revised in October. Two disclosures matter: Google funded it, and five of the seven authors disclose Google employment. The researchers gave 133 practicing patent lawyers at eleven U.S. intellectual property firms three months of access to an unreleased Google Labs patent-drafting assistant, and blinded expert patent attorneys scored the work. AI-assisted drafts scored about a third of a standard deviation higher. Then the researchers took the tool away and asked everyone to redline a patent draft unassisted. The treatment group still scored higher overall, but the gain was concentrated among lawyers with seven or more years of experience. Junior lawyers showed no average gain. The September version of the paper put it bluntly: “The largest gains from AI thus accrued to the lawyers who retained the least.” The measured time savings on the drafting task were about ten minutes, against a baseline of roughly 110 minutes.

Erick S. Robinson

Vals had measured raw AI output at anywhere from six to 80 times faster than lawyers. The gap between 80 times and about 9 percent is the time a competent lawyer spends checking and integrating the output. That time is not overhead. It is the product. Firms that budget for the 80 times and not the 9 percent are going to be surprised, and so are their clients.

Dr. Angela Liu

One more study is worth remembering, and it is not a legal one. Dell’Acqua and colleagues gave 758 Boston Consulting Group consultants GPT-4. On tasks inside the model’s capability frontier, the consultants completed 12.2 percent more tasks, 25.1 percent faster, at higher quality. On a task designed to sit just outside that frontier, they were 19 percentage points less likely to reach the correct answer than consultants without AI. The frontier is jagged, and it is invisible from the inside. A wrong pinpoint, a false premise, a fact buried in the record, and a recital that concedes your client’s position all sit just outside it, and none of them announces itself.

Dr. Angela Liu

The surveys describe a profession using tools it does not trust. LexisNexis’s 2026 survey of 543 UK legal professionals found 94 percent using AI for legal work, up from 11 percent in 2023, while the share worried about inaccurate or fabricated output rose to 83 percent. Morae’s July survey of 850 senior legal professionals found that only 33 percent trust the results of AI-assisted legal work, and 67 percent worry that the cost of human verification and oversight might outweigh the efficiency AI provides. The 8am Legal Industry Report found 69 percent using general-purpose AI at work, but only 9 percent with a written and actively enforced AI policy. In a 2026 report from Everlaw, ACEDS, and ILTA, only 41 percent of respondents named output verification as a needed skill. That may be the most worrying number in this entire episode.

Erick S. Robinson

The vendors have a serious case, and it deserves to be stated at full strength.

Dr. Angela Liu

First, humans make mistakes too. In Vals’ October 2025 research study, the lawyer baseline scored 71 percent on accuracy, below every AI tool tested. In the Legal Benchmarks drafting study, the human lawyers came in slightly below the AI tools and raised no explicit risk warnings at all. The honest comparison is to the lawyer the client would otherwise get, not to perfection.

Dr. Angela Liu

Second, the products change monthly, so even a 2026 benchmark will be stale by spring.

Dr. Angela Liu

Third, zero is the wrong standard. As Harvey’s cofounder put it, “One common misconception in legal is that an AI system needs to be 0 hallucinations to be useful.”

Dr. Angela Liu

Fourth, the court record is skewed. More than half the entries involve self-represented litigants using consumer chatbots, most entries never identify the tool, and the lawyers involved are disproportionately solo and small-firm practitioners without enterprise tools or review layers.

Dr. Angela Liu

Fifth, the strict benchmarks are harsher than practice. Lawyers do not hand an agent a twenty-step assignment and walk away.

Erick S. Robinson

Each of those points is correct as far as it goes. None of them changes the answer, for four reasons.

Erick S. Robinson

The first is detectability. A human associate’s mistake usually has a traceable reason and looks like a mistake. An AI’s wrong pinpoint, invented quotation, or unsupported date is formatted exactly like a right one, delivered with the same fluency, and placed exactly where reviewers skim.

Erick S. Robinson

The second is accountability. The vendor’s terms disclaim responsibility. The lawyer’s signature does not. Under the Ninth Circuit’s formulation, every error the tool makes is legally the lawyer’s error, and every error the lawyer misses is a candor problem waiting to happen.

Erick S. Robinson

The third is cost. Zero hallucinations is indeed the wrong standard. The right standard is the total cost of producing correct work, including verification, and on the measured evidence that cost has not fallen nearly as far as the demos suggest.

Erick S. Robinson

The fourth is that the vendors’ strongest argument cuts against their own behavior. If the products improve monthly, the answer is to test them monthly, in public, on frozen question sets. A vendor that has declined to be tested cannot also claim that the only available independent numbers are out of date.

Erick S. Robinson

That leaves one question: why, two and a half years on, the most basic fact about the most consequential tools in legal practice is still unknown. The answer is incentives. No vendor gains from a public number unless it wins. Independent benchmarks depend on vendor opt-in, so the vendors with the most to lose can decline, and they have. Buyers rarely insist, because procurement runs on security, integrations, and existing contracts, and the people who sign the contracts are seldom the people who read the cited pages. The vendors are right that a fair test is hard. They are wrong that it is impossible, because Stanford, Vals, Linklaters, Legal Benchmarks, the DELTA project, and Artificial Analysis have each run one.

Erick S. Robinson

A credible test is not mysterious. It uses a frozen, public question set with a held-out portion. It reports fabrication, misgrounding, wrong pinpoints, contradictions of the record, omissions, and refusals as separate rates rather than one blended accuracy number. It names the reference database, the product version, and the base model. It uses blinded human graders and reports how often they agreed. It compares the product to its own base model on identical prompts. It reports cost and time, including verification time, and it gets rerun on every major release. Until a vendor submits to that, the right posture toward its accuracy claim is the posture a court takes toward an unverified citation.

Erick S. Robinson

**Bottom line:** the vendors are right that their tools are valuable, and wrong that their value can be assessed without independent measurement.

Erick S. Robinson

The article closes with 34 recommendations for lawyers, patent practitioners, firm leaders, clients, vendors, and courts. We are not going to read all 34. Here are the ones I would put on the wall of every litigation team, and several of them are what my own firm already does.

Dr. Angela Liu

For anyone who signs a filing, start here. Read the cited page, not the citation. The live risk in 2026 is a real case that does not say what your brief says, or says it on a different page. Existence checks and citators do not catch that. A human reading the pinpoint does.

Dr. Angela Liu

Verify against an independent source, never with a second model alone. In In re Mitchell, a second AI tool verified citations to a file history that did not contain them. A different model is not a different source. Pull the opinion, the specification, the transcript, or the exhibit.

Dr. Angela Liu

Treat facts about your own record as citations. Dates, amounts, parties, deposition pages, column-and-line cites, and prosecution history statements are where the 2026 hallucinations live.

Erick S. Robinson

Read the recitals, definitions, and boilerplate as carefully as the operative terms. In my standstill test, the dangerous sentence was a recital, which is exactly where a tired reviewer skims.

Dr. Angela Liu

Never give the tool your conclusion and ask it to confirm. Deference to a confident user is the failure that has improved least since 2024. Ask open questions, then ask for the strongest contrary authority.

Dr. Angela Liu

If an error gets through, own it immediately and completely. Courts have been harshest on the lawyers who minimized, blamed the tool, or misstated how they used it. In Cartagena, the vendor itself supplied the evidence that the lawyer’s explanation was false.

Erick S. Robinson

For patent litigators and prosecutors, build intrinsic-record verification into every claim chart, claim construction brief, contention, and expert report. Check every column-and-line cite, figure reference, and file-history quotation against the PDF before anything is served or filed. A 200-row chart is exactly the long work product the strict benchmarks say will contain errors. Keep patentability and invalidity judgments with the experts, because model performance on novelty and nonobviousness remains only modestly better than chance.

Erick S. Robinson

For firm leaders, run your own frozen test set. Take twenty to fifty matters from your practice, have your partners build the answer keys, and rerun the set on every major product release and on the base model alone. Include at least one task like my standstill test, where the failure is judgment rather than citation. The cost is a few partner-days. The alternative is discovering the answer in a show-cause order.

Dr. Angela Liu

Budget verification as a named line of work, not as overhead. The measured time savings for practicing lawyers is on the order of 9 percent, not 80 times, and the difference is review. Write the AI policy down, train to it, and audit compliance, because Sullivan & Cromwell and Butler Snow both had policies. Revalidate on every model change, and require your vendors to tell you when the model changes.

Dr. Angela Liu

Finally, protect the skill that makes these tools safe. In the patent-drafting trial, junior lawyers who used AI for three months showed no average gain on unassisted work, and a federal judge has now warned in a written order that over-reliance on AI risks stunting their training. Schedule AI-off work for juniors. Make unassisted cite-checking a required rotation. Evaluate associates on their ability to find the error in an AI draft, not just to produce one.

Dr. Angela Liu

For in-house counsel and clients, put verification requirements, not just disclosure requirements, in your outside counsel guidelines. Ask your firms which AI products they use, how they validated them, and when they last revalidated them. If the firm cannot answer the third question, the first two do not matter.

Erick S. Robinson

Here is where I come out. Reliability is a property of a workflow, not of a product. The same model scores 94 percent per criterion and 25 percent per task. The same outputs score 15 percent or 1 percent depending on the yardstick. The same lawyer is faster with the tool and no more accurate. What changes the outcome is the structure around the tool: the quality of the question, the length of the deliverable, the source the output is checked against, and whether a competent person actually reads the result.

Erick S. Robinson

The labs now sell legal products. The content owners now build models. The wrappers now post-train their own. Whatever you validated this year, you will need to validate again next year. The firms that win the next five years will not be the ones with the most licenses. They will be the ones that know, on their own matters, exactly how often their tools are wrong.

Erick S. Robinson

The full article, with every source linked, is on The AI Law Blog at ailawblog.ai, under “The Last Independent Number.” I am Erick Robinson.

Dr. Angela Liu

I am Dr. Angela Liu. Thank you for listening to the AI Law Podcast.