In February 2026, the Fifth Circuit sanctioned appellate counsel $2,500 for a reply brief containing sixteen fabricated quotations and five misstatements of law and fact, concluding that counsel had used generative artificial intelligence to draft substantially all of it and had verified none of it. The notable feature of Fletcher v. Experian Information Solutions, Inc., No. 25-20086 (5th Cir. Feb. 18, 2026), is not the amount. Rather, it is the court’s explanation that it had already proposed an AI certification rule, published it for comment, and declined to adopt it, because Rule 11 and the Federal Rules of Appellate Procedure already required the diligence that was absent. The absence of an AI-specific rule reflected a determination that no gap existed.
That conclusion should shape how patent practitioners approach these tools. Nearly every reported decision concerns unverified output, indicating that verification is essential. But for a practice holding unpublished applications and pre-filing invention disclosures, the more consequential exposure runs in the opposite direction: not what the tool returns, but what the practitioner supplies to it. Neither exposure is adequately managed by individual judgment exercised case by case. Both are governed by obligations that attach at the firm level including confidentiality, supervision, export compliance, and candor toward the tribunal, and discharging them requires a written acceptable use policy adopted before these tools enter the workflow, not assembled after an incident has revealed its absence.
The Sanctions Record
Existing authority has proven sufficient to reach the unsupervised use of AI in legal filings. The Fifth Circuit sanctioned under Rule 46(c) and its inherent power, observing that attorney-generated fabrications are increasing rather than declining, and its survey of in-circuit decisions spans all four federal districts in Texas. The leading district court decision is Gauthier v. Goodyear Tire & Rubber Co., No. 1:23-cv-281, 2024 WL 4882651, at *3, imposing a $2,000 penalty and requiring completion of a continuing legal education course on generative AI.
Two details bear on the drafting workflow rather than ethics alone. In Gauthier, opposing counsel identified the nonexistent authorities in a reply brief and the responding sur-reply did not address them. In Fletcher, the court stated that lesser sanctions would likely have followed had counsel been forthcoming; instead, counsel attributed the errors to publicly available versions of the cases and disclosed the use of AI only when asked directly a second time.
It is advisable for firms to maintain contemporaneous records of which tool was used on a given task and what was independently verified, so that an account given months later rests on documentation rather than recollection. This is best addressed at the firm level rather than left to individual discretion. Specifically, there should be a written policy identifying approved tools, specifying what each user must record, and assigning responsibility for review gives effect to the supervisory duties that Rule 5.03 and the USPTO guidance place on practitioners answerable for the work of others.
Texas Guidance
The decisions above establish what courts will sanction after the fact, but they do not supply a standard of care to apply beforehand. That standard came in February 2025, when the Professional Ethics Committee for the State Bar of Texas issued Opinion 705, applying to generative AI the framework it had previously applied to metadata in Opinion 665 and cloud computing in Opinion 680. Its principal concern is the input side of the problem, and it supplies much of the substance a firm policy must implement.
Two holdings warrant attention. First, the Committee recognizes that a prompt itself constitutes a disclosure. Because these tools invite conversational exchange, a lawyer who describes the operative facts, the theories under consideration, and the arguments in development may expose privileged mental impressions without transmitting any document. Practitioners who conceive of confidentiality as a question about attachments have misidentified the exposure.
Second, the Committee is unequivocal regarding tools that incorporate user inputs into training data, characterizing the prospect that one client’s confidential information could surface in another user’s session as unacceptable. The opinion carries forward Opinion 680’s four precautions: 1) understanding the technology, 2) reviewing and where necessary renegotiating terms of service, 3) learning the tool’s data-security protections, and 4) training lawyers and staff, and, where satisfied, accommodates AI vendors under Rules 1.05(c) and (d) as it has long accommodated copy services and cloud providers.
Firms serving clients beyond Texas should note that Opinion 705 is comparatively permissive on client consent, requiring only that a lawyer consider informing the client, while expressly observing that ABA Formal Opinion 512 and Florida Opinion 24-1 move closer to requiring informed consent.
Obligations Specific to Patent Practice
The USPTO addressed these tools in Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the United States Patent and Trademark Office, 89 Fed. Reg. 25609 (Apr. 11, 2024). Announcing no new rules, the Office identified the existing obligations that apply: confidentiality under 37 C.F.R. § 11.106, the signature certification attaching to every paper under § 11.18(b), and the duty of disclosure under 37 C.F.R. § 1.56(b), which places materiality determinations with the practitioner rather than a model.
The guidance then addresses a consideration largely absent from general commentary on legal AI. The Office warns that AI tools may operate on servers located outside the United States and that technical data entered into such tools may thereby be exported, implicating the Export Administration Regulations and the deemed export rule at 15 C.F.R. § 734.13(b). Practitioners contemplating export of subject matter abroad for preparation of U.S. applications are directed to seek clearance from the Bureau of Industry and Security. The Office further forecloses the apparent workaround: a foreign filing license does not authorize export of subject matter abroad for preparation of an application to be filed in the United States.
Ordinary procurement makes the significance clear. A practitioner subscribes to a drafting tool, accepts the terms of service, and enters a draft specification. Nothing in that sequence discloses where inference occurs. For a general commercial practice, the question is confidentiality. For a practice handling unpublished applications in controlled technologies, it may also be export compliance, which cannot be cured by disclosure to the client after the fact. Notably, no published decision appears to address privilege waiver arising from what a lawyer entered into a tool. Sanctions decisions are a lagging indicator, and input-side failures are both harder to detect and less correctable.
Elements of a Defensible Workflow
The obligations described above converge on at least five controls, each addressing a different point of failure. The first two govern what enters the tool and are directed at the confidentiality and export exposures identified in Opinion 705 and the USPTO guidance. The third governs what leaves it and is directed at the conduct that has produced every reported sanction. The fourth preserves the firm’s ability to account for its own practices afterward. The fifth governs the client relationship. A program addressing only the third, as most firm policies presently do, leaves the larger exposures unmanaged.
First, contract terms precede capability. Opinion 705’s precautions concerning terms of service and data-security protections are threshold inquiries, not periodic ones. Whether inputs are used for training, where the model executes, what retention applies, and what security commitments the vendor has made determine whether a tool may receive confidential subject matter at all. Server location warrants separate confirmation, as enterprise licensing does not by itself establish domestic hosting.
Second, task boundaries are established in advance. Asking a model to identify inconsistencies within a draft differs materially from asking it to determine which references are material under § 1.56. The first is analysis; the second is judgment the rules assign to a practitioner, as is the certification under § 11.18(b) that accompanies every paper filed.
Third, a practitioner verifies the output against the original sources. Reading a draft for plausibility is not verification, and neither is submitting the output to a second AI tool. Each citation, quotation, and characterization of authority must be confirmed against the document itself. This is the control absent in every reported sanction. Neither Fletcher nor Gauthier turned on whether AI had been used; each turned on whether anyone confirmed that the cited material existed.
Fourth, use is documented contemporaneously under a defined recordkeeping practice rather than at each user’s discretion. No general obligation to disclose the use of AI to the USPTO exists, and documentation serves a different purpose. A firm policy should specify what the record must capture such as the tool used, the task it was used for, and the verification performed, along with where records are retained and for how long, so that the firm can answer accurately and immediately rather than reconstructing its practices under a show-cause order.
Fifth, billing reflects actual effort. Opinion 705 permits charging for time spent using and reviewing the tool and permits passing per-use costs through at cost, but does not permit billing for the time the tool saved. Addressing this in the engagement letter, together with whether the client’s jurisdiction requires consent, keeps the efficiency discussion ahead of the invoice.
What these authorities require is not restraint but structure. Neither Opinion 705 nor the USPTO guidance announced new rules, and the Fifth Circuit expressly declined to adopt one. Each concluded that existing obligations were adequate to address the problem. The gap is therefore not regulatory but institutional. What is needed is a written policy, adopted at the firm level and put in place before tools are deployed rather than after an incident, that identifies approved tools and the contract terms they must satisfy, defines the tasks for which AI may and may not be used, assigns verification to a practitioner, prescribes what each use must record, and states how AI-assisted work is billed and what the client is told.
Firms that establish those controls first are not the ones constrained by them; they are the ones positioned to apply these tools to the matters where the efficiency is worth having. Opinion 705 is explicit that lawyers should not unnecessarily retreat from technology that saves clients significant time and money.
For law firms and companies incorporating generative AI into patent practice, a structured approach to confidentiality, export compliance, verification, and oversight can help manage risk while preserving the efficiencies these tools can provide. Contact Conley Rose to discuss how your organization can develop practical policies and procedures for the responsible use of AI in patent practice.