A patent lawyer in Delaware ran his reply brief through a legal AI tool and filed it. Every case in it was real. The brief argued for a claim construction his client had never taken.
THE LEAD PLAY
Every Citation Was Real
Judge Jennifer Choe-Groves issued a memorandum on August 20 in Disruptive Resources, LLC v. Ballistic Barrier Products Inc., No. 1:24-cv-00321-JCG (D. Del.). No fabricated cases. No sanctions. The problem sat upstream of both.
Plaintiff's counsel told the court he used "Strongsuit" to generate an initial outline of his reply brief on claim construction and to pull statements from a deposition. StrongSuit is a legal-specific product, sold to litigators for exactly this work. Portions of the brief that came out the other end "included an erroneously proposed construction that deviated from the construction of the terms Plaintiff had advanced through the claim construction process."
In a patent case that is not a drafting slip. Claim construction is the case. The brief asked the court to read the client's own patent terms in a way the client had spent months of the process arguing against.
Two sections carried incorrect constructions. A third came out entirely, for errors the court called more pervasive. Counsel moved to strike the problem section of his own reply brief, and the parties filed an amended joint brief with footnotes explaining what had happened.
How it survived review is the part worth reading twice. Counsel did review the draft. From the opinion: when editing, he "jumped from section to section" and mistakenly "believed [he] had edited the entire brief, when in reality, [he] had missed the 1-page section."
The court declined to sanction him, because "counsel took full responsibility for his mistakes, was not evasive, and prioritized candor to the Court and opposing counsel by disclosing the errors quickly." Then, plainly: "The Court warns counsel that any future incidents involving AI mistakes in this case may result in sanctions."
Now run that brief through whatever verification workflow your firm built this year. Every citation resolves. Every quotation appears where it is said to appear. Every proposition is supported by the case cited for it. The brief passes clean, and the brief is still a disaster, because no step in a cite-check asks whether the document takes your client's position.
We have anchored five plays in this newsletter on verifying citations. This is the first one where verifying citations would have changed nothing.
The Play this week: Take the two or three most position-heavy filings you have in flight. Claim construction briefs, Markman submissions, summary judgment oppositions, anything where what you ask for has to line up with what you have already told the court.
For each one, have somebody who did not draft it run a position check instead of a cite check. Three columns. What does this draft ask for. What have we already asked for on this record, in prior filings, contentions, the joint chart. Do they match. That is the entire exercise, it takes about twenty minutes on a brief, and it catches the class of error a cite-check is structurally blind to.
Then fix the upstream half. Ask your litigators one question: at what point in drafting does the tool touch the work? If the answer is at the beginning, as it was here, your review protocol is pointed at the wrong stage. An outline error propagates into every section built on top of it and leaves no artifact for a verification pass to find. The check has to happen where the tool enters, not where the draft ends.
SECOND CHAIR
The appendix your expert signed
Two days before the Delaware order, Judge Tiffany Cartwright ended LeDoux v. Outliers, Inc., No. 3:24-cv-05808-TMC (W.D. Wash.). Summary judgment for defendants. Remaining claims dismissed with prejudice, no leave to amend.
The mechanism was an appendix. Quoting its own earlier sanctions order, the August 18 opinion recounts that plaintiff's counsel used Claude or ChatGPT to "generate a formatted citation table" for academic articles, and "provided the same AI-generated citation table to both experts as an appendix, and neither expert caught the errors in the citation data before signing their reports."
That table was the list of materials the experts represented they had relied on. Dr. Gabriel Holguin's hallucinated citations appeared only in portions of his report that were later withdrawn, and it did not save him. The court found they "shatter[] his credibility with this Court," then excluded him:
"This level of involvement from counsel in drafting (with AI) the list of materials supposedly relied upon by Dr. Holguin, and Dr. Holguin's lack of diligence in verifying the sources that he represented were the basis for his opinion, prevent Plaintiff from meeting her burden to show that Dr. Holguin's opinion is 'based on sufficient facts or data.'"
A second expert, Evelyn Cadman, went out on unrelated grounds. Four others had already been stricken at partial summary judgment. With no admissible expert evidence left on medical causation, the case was over.
The read: Your AI policy governs people who work for you. The document that ended this case was signed by somebody who does not. Retention letters almost never mention generative AI, and the reliance appendix is the softest target in an expert report because it reads as clerical. Under Rule 702 it is the foundation of the opinion. If anyone at your firm drafts any part of an expert's report, the source list included, the expert has to verify it before signing, and you need to be able to show that happened.
Twenty-five dollars a client
On August 20, in the middle of ILTACON week, Clio put a public price on an AI outcome. Grow AI is a set of intake agents working phone, email, and web chat around the clock: answering questions, screening leads against firm criteria, booking consults, and going back after prospects who have stopped responding. It costs $25 per converted lead.
Clio's own description of what triggers the charge: "Firms pay when a client is hired who was captured through the voice agent, or when Grow AI revives a prospect who has stopped responding and they go on to sign their engagement agreement. If Grow AI had no role in the hiring, there's no charge." Existing Grow customers get their first three conversions free. The product launched globally, though the voice agent is US only.
Set that against the rest of the week. Not one of the ILTACON launches published a price. Harvey shipped a full platform rewrite on August 18 and a post-trained model on August 20 with no pricing attached to either, and Law360 Pulse reported on August 19 that it is holding to seat-based pricing while the rest of the market argues about consumption.
The read: Put $25 next to what a signed matter actually costs you through paid channels. Nobody at your firm is going to argue about that number. They are going to argue about what counts as a conversion. Clio's line is that you pay when Grow AI had "a role in the hiring," and nothing in the announcement says who decides whether it did. Anyone who has run intake has had that argument about the lead who called twice, got a follow-up text, and signed nine days later. Ask what the attribution window is, and whether you can pull the record behind any given charge.
STILL WATCHING
California SB 574 was read a third time and amended on the Assembly floor on August 21, the second amendment since it came off the suspense file on August 13. Both houses have to pass bills by August 31, and the Senate still has to concur in the Assembly amendments. Seven days.
California AB 1651 was approved by the Governor and chaptered on August 22 as Chapter 116, Statutes of 2026. It adds Business and Professions Code section 6060.15 and requires the State Bar to disclose where AI was used in developing or administering the bar exam and its study materials. The State Bar now carries a disclosure duty about its own AI use.
The California rules amendments to RPC 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 have not moved. Public comment closed May 4. That is 112 days without a Board of Trustees vote.
Thomson Reuters v. ROSS Intelligence, No. 25-2153, argued in the Third Circuit on June 11. Seventy-four days, no opinion.
The Copilot for Word prompt injection. 172 days since Hakon Maloy reported it, 28 since he published it. Microsoft's next Patch Tuesday is September 8.
QUICK HITS
A Florida appellate court sent the bill to the lawyer. In Capital Standard, LLC v. U.S. Bank N.A., No. 2D2024-1392 (Fla. 2d DCA, August 21), the court catalogued at least 31 instances in the amended initial brief of real cases cited for propositions they do not support, along with fabricated quotations, then 15 more in the reply, including a case that does not exist. Attorney Kenneth H. Keefe ignored the show cause order for over a month. Both briefs stricken, $500 for the late response, $1,000 for the AI misuse, fees to U.S. Bank to be set on remand, referral to The Florida Bar. And the sentence that lands: "Attorney Keefe is solely responsible for paying the fee award and fine and may not charge his clients for those amounts."
The SRA put supervisors on notice. The Solicitors Regulation Authority published a warning notice on misuse of AI on August 17. The line to forward to your practice group leaders is that those who supervise "may also be found to have breached regulatory requirements and professional duties if false citations are put before the court without adequate review and/or supervision." It adds that client information "should only be entered into AI systems where appropriate contractual, technical and organisational safeguards are in place to protect confidentiality," and tells in-house solicitors to be particularly mindful of which tools are available to them. It is a UK regulator, and the supervision duty it describes maps onto Model Rule 5.1 without much translation.
A vendor published a benchmark, and the number is doing work for the vendor. NetDocuments released a Legal Context Engineering Benchmark on August 18: the same agent answering 300 questions across 10 real matters and 874 documents, with and without structured context from its own context graph. Claimed result, 48 percent lower cost per correct answer at equal quality, 52 percent fewer tokens. It is vendor-run, vendor-published, and not independently audited, and the extrapolated million dollars a year for a 2,000-lawyer firm assumes that firm asks four million questions a year. The direction is credible. The dollar figure is marketing.
Harvey published a result that did not work. Harvey released Tenet on August 20, an open-weight legal model post-trained on a Kimi K3 base. Its own Legal Agent Bench figures put the diligence criteria pass rate at 60.1 percent against 43.8 for the base model. Harvey also disclosed that on the hard subset of Scale AI's PRBench, the improvement was not statistically significant. In a week when most of the field shipped launches carrying no numbers at all, one of them published a null result.
Somebody at your firm started a brief from an AI outline this morning. The cite-check at the end of it is looking at the wrong thing.
See you in the next one.