AI Esquire
Menu
Plans from $497/monthBuy Intake AI
Professional responsibility · AI governance · Rule 11

The AI Policy Failed to Stop the Brief. It Still Saved the Firm.

A law firm's artificial intelligence policy did not stop a partner from filing a brief containing fabricated quotations, inaccurate citations, and mischaracterized authority. The policy still mattered. In a November 13 letter opinion, Delaware Vice Chancellor Lori W. Will sanctioned the signing lawyer $2,500 under Court of Chancery Rule 11, payable personally, but declined to sanction Richards, Layton & Finger. The split result is more useful than another lecture about hallucinations. It shows that governance has at least two jobs. Controls should prevent mistakes, but they also create evidence about whether the institution acted reasonably when a person bypasses them. The court found exceptional circumstances relieving the firm of the joint responsibility Rule 11 ordinarily assigns. It found no similar escape for the lawyer whose signature certified the brief. That distinction should change how firm leaders evaluate AI policies. A document in a handbook is not the control. The control is the system that the firm can prove existed before the failure.

The first correction did not correct the legal problem

The underlying sequence matters. On January 22, 2026, the plaintiffs filed an answering brief in an advancement dispute. Five days later, opposing counsel identified fabricated quotations, inaccurate citations, and hallucinated legal propositions. Lead counsel acknowledged that a generative AI tool had been used to revise the brief and filed a corrected version. That filing removed quotation marks around nonexistent quotations but retained inaccurate descriptions of the cases.

The response then became part of the problem. Counsel initially attributed the failure to an oversight during a paralegal's cite-check. After opposing counsel identified the remaining mischaracterizations, he objected that they had not met and conferred first and maintained that the decisions were real and broadly supported the argument. The court later emphasized that there is nothing to negotiate when false citations or quotations are presented to a tribunal. Opposing counsel was justified in going directly to the court.

Vice Chancellor Will ultimately ruled for the plaintiffs on the merits of the advancement dispute. She retained jurisdiction over the filing failures and ordered the signing lawyer and his firm to show cause. That separation is important. A legal position can prevail while the way counsel supported it violates Rule 11. Accuracy is not optional because the conclusion might have been defensible with different authorities.

The episode also exposes why a superficial correction can increase risk. Removing quotation marks addresses appearance, not substance. Once a court or adversary identifies AI-generated errors, the responsible lawyer must return to the primary authorities and recheck the entire affected work product. Editing around a visible defect while leaving the fabricated synthesis intact is not remediation. It is continued advocacy of the same unsupported proposition.

Rule 11 divided the institutional question from the personal one

Court of Chancery Rule 11(b)(2) provides that a lawyer presenting a filing certifies, after an inquiry reasonable under the circumstances, that its legal contentions are warranted by existing law or a nonfrivolous argument to change it. Rule 11(c)(1) ordinarily reaches the institution too. Absent exceptional circumstances, a law firm must be held jointly responsible for a violation committed by a partner, associate, or employee.

The opinion did not treat a law firm policy as an automatic safe harbor. It used professional-responsibility standards to ask whether the firm had made reasonable efforts to give reasonable assurance of compliance. Delaware Rule of Professional Conduct 5.1 requires managerial lawyers to make those efforts. ABA Formal Opinion 512 similarly tells managers to establish clear policies for permissible generative AI use and supervisors to take reasonable steps to ensure that lawyers and nonlawyers meet their obligations.

Richards, Layton & Finger had evidence, not merely an aspiration. The firm barred unapproved generative AI tools in March 2023. It formed a working group in 2024 to evaluate tools and ethics guidance. Its board adopted a comprehensive policy in July 2025 requiring experienced-lawyer review for accuracy and completeness and prohibiting AI as a substitute for independent judgment. More than 200 employees completed an accredited ethics program addressing hallucinations, candor, competence, confidentiality, meritorious claims, and supervision by January 2026.

Those safeguards did not prevent this brief. They nevertheless answered a different legal question: had the firm acted reasonably before the violation? The court concluded that it had and found exceptional circumstances relieving it of joint responsibility. That conclusion is fact-specific. Another court applying another rule could weigh the same controls differently. But the opinion gives firm leaders a concrete picture of what a court may regard as a substantial, multi-layered effort rather than a paper policy written after the incident.

The person closest to the policy can still be the control failure

The individual facts are uncomfortable precisely because the lawyer was not an uninformed outsider. According to the opinion's account of his affidavit, he chaired the firm's AI subcommittee and was leading a pilot evaluation of Perplexity Enterprise, which had not been approved for general firm use. He used it to help draft and condense arguments, then delegated citation review to a junior lawyer and a paralegal. He filed without personally checking the authorities, citations, and quotations after that review.

Expertise in a system does not eliminate the need to follow its controls. It can create the opposite risk. A pilot leader may feel unusually capable of judging an unapproved tool, unusually familiar with its output, or unusually entitled to treat the exception as part of the experiment. Governance that binds ordinary users but leaves product champions to informal judgment contains an exception large enough to swallow the policy.

The court's answer was straightforward. The signing lawyer's duty of candor and reasonable inquiry was personal and nondelegable. Lawyers routinely and appropriately rely on juniors and paralegals for cite-checking. But when a lawyer knowingly introduces generative AI text into a draft, the signer must ensure that the underlying authorities were reviewed and support the propositions for which they are cited.

That is not a demand that partners repeat every mechanical step performed by a team. It is an allocation of accountability. Delegation can distribute work, but it cannot transfer the certification attached to a signature. The sanctioned lawyer later accepted full responsibility, expressed sincere contrition, and had a distinguished record before the court. Those facts affected the remedy. They did not erase the violation.

A policy should create an auditable control system

Many firms still evaluate AI governance by asking whether they have a policy and whether everyone attended training. Leiske v. Kidd suggests a more demanding audit. Could the firm prove when the rules took effect, which tools were approved, who could authorize an exception, what review the policy required, what subjects the training covered, who completed it, and whether the safeguards existed before the disputed filing? Those records helped the court distinguish institutional effort from individual departure.

Prevention is only the first layer. A mature system also needs detection. High-risk work should carry a visible review path, especially when an experimental tool is involved. A matter team should be able to identify which portions of a filing came from an AI-assisted process, who read the primary sources, and who performed the final verification. This does not require preserving every prompt forever. It requires evidence proportionate to the professional risk.

Response is the third layer. When a defect surfaces, the firm needs an escalation path that moves the work out of ordinary editing and into incident review. The scope of the error should be mapped. The underlying authorities should be re-read. Other filings using the same workflow may need review. The court and client may need a prompt, accurate explanation. Blame should not be shifted to staff, and an adversary's decision to notify the tribunal should not be treated as a procedural offense.

Finally, exceptions need controls stronger than the default, not weaker. Pilots are where uncertainty is highest. The policy should name the sponsor, approved users, permitted matters, confidentiality limits, verification protocol, monitoring, and stop conditions. A firm that cannot show how an experimental tool moved from evaluation to a court filing does not have a pilot. It has an undocumented production system.

The opinion rewards serious governance, not perfect prevention

The headline lesson is not that a strong policy saves a firm from every sanction. Rule 11's exceptional-circumstances language, the Delaware professional rules, the specific record, and the court's assessment of reasonableness drove this result. A firm that adopts similar words without similar implementation should not expect the same outcome. Nor does the opinion resolve possible duties to clients, insurers, regulators, or other courts.

Its narrower lesson is still important. A control can be effective even when a person evades it. Safety systems, conflicts procedures, trust-account controls, and cybersecurity programs are not judged solely by whether a failure ever occurs. They are also judged by whether the institution identified the risk, designed reasonable barriers, trained the people involved, documented compliance, detected departures, and responded honestly.

That is a better standard for AI governance than the fantasy of zero error. Generative tools will change, lawyers will make mistakes, and exceptional users will sometimes believe the rule was written for somebody else. The firm's obligation is to build a system that makes the right conduct normal, the exception visible, and the response credible. The lawyer's obligation at filing remains simpler. Read the authorities. Check the propositions. Own the signature.

Sources and further reading

Primary and industry sources used to support this page. External guidance should be reviewed in context and for your jurisdiction.

  1. Court of Chancery letter opinion, Leiske v. KiddThe official November 13, 2026 opinion resolving the Rule 11 show-cause proceeding, detailing the filing errors, the firm's safeguards, and the $2,500 personal sanction.
  2. Delaware Court of Chancery RulesThe official rules, including Rule 11's certifications, institutional responsibility, and sanctions provisions.
  3. Delaware Lawyers' Rules of Professional ConductThe official rules referenced by the court, including managerial responsibility, candor, and supervision.
  4. ABA Formal Opinion 512The ABA ethics opinion on generative AI competence, confidentiality, candor, supervision, fees, and managerial policies.
  5. Bloomberg Law, Delaware Chancery Court sanctions attorneyNovember 13 reporting on the sanction and the court's treatment of the law firm's governance efforts.
Put the framework to work

See Intake AI handle your firm's real workflow.

Bring the intake questions, routing rules, or coverage gap you want to improve. We will demonstrate the system against them.

Call Intake AI now (941) 941-6967Book a 30-minute working session