AI GOVERNANCE
An experiment becomes a live system the moment its output reaches a customer, whatever the project plan still calls it.
The Write-Off: The policy passed. The brief didn’t.
THE WRITE-OFF · ISSUE 002
A Delaware judge has fined a senior partner $2,500 for filing an AI-assisted brief containing fabricated quotations, and spared his firm because its AI governance was genuinely good. The awkward twist is that the partner chaired the firm’s AI subcommittee and was trialling an unapproved tool at the time. Sound policy protected the institution; it could not protect the deliverable from the person who stepped around it.
TODAY’S STORY
Rose Krebs, Law360
·
16 November 2026
On 22 January 2026, Richards, Layton & Finger, one of Delaware’s best-known firms, filed an answering brief in Leiske v. Kidd, an advancement dispute in the Court of Chancery. Five days later, the other side pointed out that the brief contained fabricated quotations, inaccurate citations and what the court later called “hallucinated legal propositions”.
The response made things worse. The signing partner, Richard Rollo, acknowledged that a generative AI tool had been used to revise the brief and attributed the miss to a paralegal’s cite-check. The corrected version removed the quotation marks around the invented quotes but kept the inaccurate descriptions of the cases. The following day he criticised opposing counsel for not conferring before alerting the court, and insisted the authorities were real.
In an opinion dated 13 November, Vice Chancellor Lori Will fined him $2,500 under Rule 11, payable personally within 14 days. His sworn affidavit explained that, as chair of the firm’s AI Subcommittee, he had been leading a pilot of Perplexity Enterprise, a tool not approved for general use, and had used it to draft and condense arguments before handing citation checks to a junior lawyer and a paralegal. The firm escaped sanction entirely. It had banned unapproved AI tools in March 2023 and adopted a board-approved policy in July 2025 requiring an experienced lawyer to review AI output, backed by accredited training completed by more than 200 staff. “In short, RLF did the right thing,” the court wrote. The partner accepted full responsibility, and the court noted his years of distinguished practice.
THE VERDICT
Richards Layton deserves credit, and the court gave it. The firm’s escape was bought years earlier, with a 2023 ban on unapproved tools and a board-approved policy backed by real training, all in place before the brief was filed. That is about the cheapest insurance a board will ever buy. The failure happened in the gap the policy did not close: a sanctioned experiment, run by the firm’s own AI lead, producing work that went straight to a judge. The $2,500 is small change. Nine months of show-cause proceedings, a sworn affidavit from the firm’s president and a public opinion naming the AI subcommittee chair are rather less small.
Pilots exist so an organisation can learn safely. This one was used on a real filing in a real case. The moment an unapproved tool’s output entered a court submission, the evaluation became production, with none of production’s controls.
The partner drafted with AI, then passed verification down the chain. The court was blunt that this duty is personal and nondelegable. In business terms, the person accountable for the output was not the person who checked it, and a citation check confirms that a case exists, which is a different question from whether it says what the brief claims.
Removing quotation marks fixed the visible error and left the underlying fabrication in place. “It is not enough to remove quotation marks and leave a fabricated legal synthesis in place,” the court said. Any operations leader will recognise the pattern: close the defect ticket, keep the process that produced it.
The first instinct was to point at a paralegal and then at opposing counsel. That instinct spent credibility the firm later had to rebuild under oath. When an AI error surfaces, the quickest route to a contained incident is to assume the output is wrong until the sources prove otherwise.
The principle we apply with clients is simple: deterministic by default, AI by advantage. Let the model draft and condense if it genuinely saves time. The verification step should be deterministic, owned and evidenced, and it should sit with whoever puts their name to the result.
Ring-fence pilot tools. Trials run on sandbox material, or their output carries a label and full production review before it leaves the building.
Make the signatory verify personally. Whoever signs a board pack or a regulatory filing confirms the sources themselves, with AI-touched passages flagged in the draft.
Record AI use on every external deliverable. A one-line declaration of which tool touched what makes review targeted, and makes any later affidavit much shorter.
Re-verify everything after the first error. One fabricated quote means every authority in the document is suspect until checked again.
Treat outside challenge as free quality control. When a customer, auditor or counterparty spots an error, thank them and check. Arguing about process first is the expensive option.
Richards Layton shows governance paying for itself in hard cash and reputation. The uncomfortable corollary for boards is that policies need to bind the enthusiasts as tightly as the sceptics, because the enthusiasts are the ones running the pilots.
THE QUESTION FOR MONDAY’S LEADERSHIP MEETING
If the answer includes the phrase “just testing”, the pilot has already gone live.
NEWTHISTLE NUGGET
AI GOVERNANCE
An experiment becomes a live system the moment its output reaches a customer, whatever the project plan still calls it.
The Write-Off: The policy passed. The brief didn’t.
SOURCES
Richards Layton Atty Fined For AI Gaffe, Firm Not Sanctioned, Rose Krebs, Law360, 16 November 2026
Leiske v. Kidd, C.A. No. 2025-0426-CDW (LWW), letter opinion, Delaware Court of Chancery, 13 November 2026
Legal AI Weekly: AI Policy Helps Law Firm Avoid Sanctions, Oliver Roberts, Legal AI Weekly, 16 November 2026
The Weekly Inference: Courts Are Sanctioning Lawyers Who Use AI Carelessly, Steptoe, July 2026