## The Technology Blind Spot John C. Farris pleaded guilty to drug trafficking in a Kentucky federal court. His appeal came next. The court appointed Steven N. Howe, a Kentucky practitioner with four decades of clean practice. Howe’s office had recently acquired Westlaw’s CoCounsel platform. He directed a staff member to upload the district court documents into the tool. The platform generated draft briefs. Howe worked the file for approximately six hours. That subscription renewed despite a documented one-in-three hallucination rate — measured by Stanford, unknown to Howe, invisible to the court that sanctioned him. The Sixth Circuit noticed the file name first: “CoCounsel Skill Results.” On closer examination, the court found fabricated quotations in three cited cases and misrepresented holdings in two more. *United States v. Washington*, 715 F.3d 975 (6th Cir. 2013), had upheld the exact sentencing enhancement Howe was attacking. CoCounsel’s output said otherwise. On April 3, 2026, a three-judge panel removed Howe from the case, denied him compensation for the entire appeal, and referred the matter for disciplinary proceedings. Farris continues to wait for resolution of his appeal. Thomson Reuters renewed Howe’s subscription on schedule. That last sentence is not a punchline. It is the problem. **The Asymmetry No One Examines** Every AI hallucination sanctions case in American courts follows the same procedural sequence: the AI output fails, the court identifies the failure, and the attorney bears the consequences alone. The vendor, whose product generated the fabrication, faces no proceeding, no consequence, and no interruption to its revenue. That sequence has held through *Mata v. Avianca, Inc.*, No. 22-cv-1461 (S.D.N.Y. June 22, 2023), where attorney Steven Schwartz submitted six hallucinated citations and paid a $5,000 fine. It held through *ByoPlanet Int’l, LLC v. Johansson*, 792 F. Supp. 3d 1341 (S.D. Fla. July 17, 2025), the largest AI hallucination sanction in American legal history at $85,567.75. It held through *Johnson v. Dunn*, No. 2:21-cv-1701-AMM (N.D. Ala. July 23, 2025), where Butler Snow partners were disqualified and referred to bar regulators despite firm-wide AI governance policies. It held through *Farris*. No proceeding has examined the product. The question of whether the tool performed as marketed has not been before any court. That will change. **What Stanford Measured** In 2024, researchers at Stanford University’s RegLab and Human-Centered AI Institute ran the first preregistered empirical evaluation of AI-driven legal research tools. The study, published in the *Journal of Empirical Legal Studies* in 2025, tested Lexis+ AI, Westlaw AI-Assisted Research, Ask Practical Law AI, and GPT-4 against a carefully constructed dataset of legal queries. Before the study ran, the vendors had made specific claims. LexisNexis told customers its product delivered “100% hallucination-free linked legal citations.” Thomson Reuters and Casetext stated that CoCounsel “does not make up facts, or ‘hallucinate,’” and that its retrieval architecture was “eliminating” or “avoiding” hallucinations entirely. These were not aspirational marketing generalities. They were affirmative, falsifiable representations made to practitioners deciding whether to trust the output in court filings. Stanford’s results were precise and unfavorable. Lexis+ AI, the best-performing system tested, answered 65% of queries accurately and hallucinated on more than 17% of queries, roughly one in six. Westlaw’s AI-Assisted Research performed significantly worse: accurate on 42% of queries and hallucinating on approximately 33%, nearly double the rate of its competitor. Ask Practical Law AI, also a Thomson Reuters product, provided incomplete or refused responses on more than 60% of queries. Stanford tested general-purpose tools for comparison. GPT-4 hallucinated between 69 and 88% of the time on legal queries. The purpose-built legal tools outperform them. The problem is the distance between 33% and “hallucination-free.” That distance is measurable. Stanford’s team identified a second failure mode more dangerous than fabricating nonexistent cases: the system cites a real case that does not support the proposition for which it is cited, or attributes to that case a holding it never reached. Both types appeared in *Farris*. A fabricated citation to a nonexistent case is catchable with a basic database search. A fabricated quotation attached to a real citation requires reading the actual opinion. CoCounsel’s output in *Farris* cited real Sixth Circuit precedent. The problem was what the output claimed those opinions said. An attorney who ran a citation check would have found the cases existed. Only an attorney who read the underlying opinions would have caught the error. The product made the second kind of hallucination easy to miss. LexisNexis’s Chief Product Officer disputed the methodology publicly. The company has not released its internal accuracy data. One hundred percent hallucination-free, per the marketing. Seventeen percent error rate, per Stanford. The gap between those two numbers is an evidentiary record, not a rounding dispute. **The Design Behind the Numbers** One technical fact clarifies why this matters. A large language model does not retrieve legal authority from a verified database. It predicts statistically probable text sequences. When CoCounsel generated the appellate brief in *Farris*, the system was not locating *United States v. Washington* and pulling the correct holding. It was producing text that, given the case context and its training data, constituted a plausible appellate argument. The quotations it invented were syntactically correct, professionally formatted, and wrong. This is not a malfunction. It is the architecture. Probabilistic text generation cannot guarantee citation accuracy because the system never incorporated verification into its design. A zero-fabrication architecture requires the system to generate a citation only when it can confirm that citation against a locked, verified corpus of primary legal sources. That architecture exists. Stanford’s researchers documented it. The vendors deployed a probabilistic system and marketed it as though it carried the properties of a deterministic one. Howe did not know that. His subscription materials did not explain it. The marketing said otherwise. *[See When Attorneys Stop Checking AI’s Work, The Technology Blind Spot (2026).]* **Rule 1.1 Has a Ph.D. Problem** Model Rule 1.1, Comment 8 requires attorneys to keep up with “the benefits and risks associated with relevant technology.” Courts cite this provision in virtually every AI sanctions case. It is the doctrinal hook on which attorney liability for AI hallucinations has been hung. The comment is correct as a principle. Its application in the legal AI context has an unexamined assumption buried inside it. The assumption: the attorney can access the information necessary to evaluate the risk. Westlaw’s CoCounsel is a closed system. The model weights are proprietary. The training data composition is undisclosed. The RAG corpus architecture is a trade secret. The specific error distributions across query types are not published. An attorney subscribing to CoCounsel cannot inspect any of these things. What she can read is the marketing material. LexisNexis told her the product was 100% hallucination-free. Thomson Reuters told her CoCounsel does not hallucinate. She made a professional judgment based on the information available to her. That information was false. A J.D. teaches statutory interpretation, civil procedure, legal reasoning, and professional responsibility. It does not teach retrieval-augmented generation architecture, stochastic output distributions, or calibrated confidence intervals in large language models. It was not designed to. The technology did not exist when the professional responsibility framework was built. Comment 8 was added to require basic technology literacy, not a graduate degree in machine learning. But the technology currently being deployed in legal practice requires something closer to the latter to critically evaluate. Butler Snow had written AI policies. Matthew Reeves held the title of assistant practice group leader. He had read the firm’s guidance. He knew the policy required written approval before using ChatGPT for legal research. He used it anyway, without verification, and without approval. Judge Manasco found his conduct “tantamount to bad faith.” The policy could not reach the moment when a fluent, authoritative screen of AI-generated text suppressed the verification instinct of an experienced partner who had decided the policy did not apply to him. As currently applied, the doctrine produces this result: an attorney violates Rule 1.1 by believing a vendor’s published claim about its own product, the vendor faces no consequence for making a false claim, and the attorney’s license is the price of the error. That is the doctrine. It may be the correct doctrine, given the current state of case law. It is not, however, a defensible endpoint. A competence standard that requires attorneys to independently verify technical performance claims that vendors actively conceal and affirmatively misrepresent is a standard that has structurally outrun the information attorneys can access. **The Disclaimer at the Bottom of the Agreement** Open any major AI legal research platform’s subscriber agreement. Navigate to the liability section. There, without exception, you will find a disclaimer of warranties, a limitation of liability capped at subscription fees paid, and, in many agreements, a provision requiring the subscribing firm to indemnify the vendor for consequences arising from tool use. The marketing department says trusted, reliable, and hallucination-free. The legal department writes: no warranties, express or implied. Attorneys read the marketing. Attorneys sign the terms of service. The terms of service disclaim the marketing. Whether that structure holds under commercial law is a question courts have not yet resolved for AI legal research products. Under the Uniform Commercial Code, a general warranty disclaimer does not automatically defeat an express warranty created by an affirmative statement of fact about a product’s performance, where the two provisions conflict and the affirmative statement was material to the transaction. “100% hallucination-free linked legal citations” is an affirmative statement of fact about product performance. “No warranties, express or implied” is a general disclaimer. When they conflict, the disclaimer does not prevail by default. That is a description of a legal question that no one has yet argued in a courtroom. Not a litigation prediction. A gap. When it gets argued, the evidentiary foundation will be the vendor’s own marketing materials, the subscriber agreement, and a peer-reviewed study from Stanford showing the distance between the claim and the performance. That is a straightforward record to assemble. **The Verification Duty Is Real. It Is Not the Only Duty.** Verification is mandatory. ABA Formal Opinion 512, issued July 29, 2024, states the standard plainly: attorneys must independently verify the accuracy of AI-generated output before relying on or submitting it to a tribunal. Howe violated this obligation. The Sixth Circuit’s response was proportionate and legally sound. The verification duty is a professional obligation, and no argument about vendor liability diminishes it. The strongest version of the opposing argument runs like this: AI tools are sold with known limitations, bar guidance imposes a verification duty, and an attorney who fails to verify assumes the risk. The Sixth Circuit said as much. The ABA said as much. If you verify the output and the brief is clean, no one is harmed. This piece does not dispute any of that. Two independent obligations can coexist. The attorney’s duty to verify and the vendor’s obligation to perform as marketed are not alternatives. Products liability law routinely allocates concurrent duties across multiple parties. A pharmaceutical manufacturer that misstates a drug’s safety profile cannot escape liability because a physician should have caught the error. The physician’s verification duty runs in parallel with the manufacturer’s disclosure obligation. The fact that courts have not applied that same allocation to AI legal research vendors reflects where the doctrine stands today. It does not reflect where it must end. An attorney who verifies output and catches no errors has no claim against the vendor regardless of architecture. An attorney who files fabricated content without verification faces sanctions regardless of what the marketing materials said. A vendor that marketed a zero-hallucination product and delivered a one-in-three hallucination product has, eventually, a question to answer that no court has yet asked. **Before Thursday** Three concrete actions belong on this week’s calendar. Pull the subscriber agreement for every AI legal research tool your firm uses. Find the liability limitation clause and the indemnification provision. Know whether you have agreed to hold the vendor harmless for consequences arising from tool failures. Know what the damages cap is. Know whether the agreement permits the vendor to modify its product without notice in ways that affect accuracy. This is the document that will govern the first proceeding in which the vendor liability question gets litigated. Know what it says before that proceeding arrives. Call your malpractice carrier and ask whether your current AI tool usage and verification protocols fall within your E&O coverage. Ask specifically whether your policy addresses AI-generated errors in court filings. Carriers are beginning to update renewal questionnaires to ask about AI tool use and verification procedures. The claims pipeline is building. Getting ahead of this conversation costs a phone call and may reveal a gap in your coverage before a sanctions order does. Build the Stanford hallucination rates into your firm’s AI governance framework as a calibrated number, not background reading. One in three is the rate at which Westlaw’s AI-Assisted Research product, in a peer-reviewed study, produced incorrect legal conclusions. That number belongs in your associates’ training, your procurement criteria, and your workflow verification requirements. A policy that says “verify AI output” without specifying the error frequency your attorneys are verifying against is a policy missing half its information. **The Record Being Built** Farris is a published Sixth Circuit decision. It names the product. It documents the mechanism of failure. It describes the consequence to a real client who waited longer for resolution of his criminal appeal because a premium AI tool generated false content and his attorney filed it. That opinion now sits in the evidentiary record that will support the first products liability claim against a legal AI vendor. That case is building. When it arrives, the plaintiff’s counsel will have the Stanford study, the verbatim marketing claims, the subscriber agreement’s disclaimer-warranty conflict, a documented chain of causation from product failure to client harm, and a growing body of sanctions orders from courts across the country that have consistently examined the attorney and never examined the product. There is a precedent pattern worth noting. Courts that developed products liability doctrine in other high-stakes domains did so after years of recorded failures that established the evidentiary predicate. The pharmaceutical industry faced strict liability after documented gaps between marketing claims and clinical outcomes became systematic. The pattern in legal AI is following a similar trajectory: documented failures, peer-reviewed performance data, marketing claims in evidence, and a growing sanctions record that identifies the product by name in published opinions. Farris added another published entry. The trajectory is not ambiguous. John C. Farris is still waiting for his appeal. The tool that generated the brief is still available on subscription. No court has yet asked whether the vendor had an obligation that matters. Someone will. **About the Author** JD Morris is Co-Founder and COO of LexAxiom, an Agentic AI platform for the business of law. Over a 25-year career, he has built and scaled enterprise technology products across Dell, EMC, VMware, and Cisco, including the first exabyte eDiscovery platform. He holds dual MBAs from Columbia Business School (Finance) and UC Berkeley Haas (Marketing), a Master of Legal Studies in Cybersecurity Law from Texas A&M, and a Master of Engineering from George Washington University. He writes The Technology Blind Spot on the intersection of emerging technology and law. Connect with him on LinkedIn at http://www.linkedin.com/in/jdavidmorris, on X at @JDMorris_LTech, or on Bluesky at @JDMorris-ltech.bsky.social. **References** 1. United States v. Farris, No. 25-5623 (6th Cir. Apr. 3, 2026) (per curiam) (Clay, Gibbons, and Hermandorfer, JJ.). 2. Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning & Daniel E. Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216 (2025). 3. ABA Standing Comm. on Ethics & Pro. Resp., Formal Op. 512 (July 29, 2024): Generative Artificial Intelligence Tools. 4. Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023). 5. ByoPlanet Int’l, LLC v. Johansson, 792 F. Supp. 3d 1341 (S.D. Fla. July 17, 2025). 6. Johnson v. Dunn, No. 2:21-cv-1701-AMM (N.D. Ala. July 23, 2025). 7. Noland v. Land of the Free, L.P., No. B331918 (Cal. App. 2d Dist. Sept. 12, 2025). 8. United States v. Washington, 715 F.3d 975 (6th Cir. 2013). 9. United States v. Anthony, 280 F.3d 694 (6th Cir. 2002). 10. Model Rules of Pro. Conduct r. 1.1 cmt. 8 (Am. Bar Ass’n 2024).

Originally published on LinkedIn Newsletter: The Technology Blind Spot