Across the country, courts and bar regulators continue to address artificial intelligence in law practice. The general consensus appears to be that the existing rules of professional conduct already govern its use, so no new category of rule has been needed. Where courts and regulators have acted, they have done so by applying or amending the existing rules, not by creating a new category of their own. The Arkansas Supreme Court, for instance, has issued Administrative Order No. 25 and amended its Rules of Professional Conduct in 2025. What follows is an overview of where the rules, and the reported enforcement, appear to stand as of mid-2026.
The starting point: the rules are applied to AI use
The recurring theme across the guidance is that AI is treated as a tool, and the lawyer remains responsible for the work. The American Bar Association’s first comprehensive guidance on generative AI, Formal Opinion 512 (July 29, 2024), reads the Model Rules onto AI use across six familiar duties, and the state authorities that have followed tend to track the same set:
- Competence. Rule 1.1, and its technology-competence comment, are generally read to require a functional understanding of any AI tool a lawyer uses, including its known failure modes. Several authorities describe this as a continuing obligation rather than a one-time matter, given how fast the tools change.
- Confidentiality. Rule 1.6 is the issue most opinions spend the most time on. The common thread is that a lawyer should understand a tool’s data-handling, retention, and training practices before entering client information, and that consumer-grade tools that train on inputs raise different concerns than commercial or enterprise tools with contractual no-training terms.
- Candor and verification. AI output, especially citations, must be checked by a lawyer before it is filed or relied upon. AI does not supply an excuse for a false statement to a court.
- Supervision. Rules 5.1 and 5.3 are generally read to reach AI the way they reach a nonlawyer assistant: the supervising lawyer is expected to have policies and to review the work.
- Communication and fees. Authorities differ on the details (below), but the general expectation is that a lawyer bills for time actually worked and discloses material AI use where it bears on the representation.
None of these is a new rule. Each is an existing duty applied to a new tool.
Where does the guidance still differ?
The application of professional conduct rules to AI use leaves several questions unsettled, and the published guidance points in different directions:
- Disclosure to the client. Some authorities find no per se duty to tell a client that AI was used; others recommend disclosure, or advance consent, when client information goes into a consumer tool. The trend appears to be a risk-based approach rather than a blanket rule, but it varies by jurisdiction.
- Billing for AI efficiency. Most guidance reads the reasonable-fee rule to mean that a client should not pay for hours that AI kept a lawyer from having to work. How that principle applies to value-based or flat-fee arrangements, where the bill is not tied to hours, is less settled, and the guidance we have reviewed does not resolve it uniformly.
- Disclosure to the court. No bar ethics rule we have seen requires disclosing AI use in a filing. That question has instead been answered, variably, by individual judges through standing orders. At least one state high court has gone the other way at the policy level: the Illinois Supreme Court’s policy states that “Disclosure of AI use should not be required in a pleading” (eff. Jan. 1, 2025).
- Agentic AI. Most guidance was written with lawyer-operated chatbots in mind. As tools begin to take actions on their own, a few authorities have started to address agentic AI, but it is early, and the rules here are still forming.
The practical takeaway from the divergence is narrow: on the unsettled questions, the governing answer is local, and it is worth checking the specific jurisdiction and judge rather than assuming a national rule.
What does the enforcement record show?
Reported decisions suggest a trend toward increasingly serious consequences for filing fabricated, AI-generated authority.
The issue appears to have entered public view in 2023, when a federal court in the Southern District of New York addressed a brief that cited cases which did not exist, generated by ChatGPT, and sanctioned the lawyers involved. In the period since, the reported orders have escalated penalties from monetary sanctions toward consequences for a lawyer’s license and standing.
Two recent decisions illustrate the current end of that range. In the Ninth Circuit, two lawyers who filed a brief with fabricated citations were fined and suspended from practice before that court for six months, with a referral to the state bar and a two-year requirement to certify AI use and verification in future filings. The court also stressed that a lawyer who learns of an error in a filing, including an AI hallucination, should immediately alert the court and opposing counsel and disclose its source. In the Sixth Circuit, a panel imposed double costs under Federal Rule of Appellate Procedure 38, made the lawyers jointly liable for the other side’s appellate fees, and added a punitive sanction of $15,000 against each lawyer under the court’s inherent authority, for briefs containing fake citations and misrepresentations of the record. State disciplinary bodies have moved in the same direction; in one matter, the Nebraska Supreme Court temporarily suspended a lawyer pending a disciplinary hearing after an appellate brief was found to contain dozens of defective citations.
The common thread is straightforward, and the courts have stated it plainly: a citation must be one a lawyer has personally read and verified, whatever its source. These are decisions from other circuits and states, persuasive rather than binding in Arkansas, but the direction is consistent across them.
What about privilege and work product?
A separate line of cases has begun to address what happens to AI conversations in discovery, and early results cut in two directions.
On one side, a federal court has held that a person’s conversations with a consumer AI tool were not protected by the attorney-client privilege or the work-product doctrine: the tool is not an attorney, the consumer terms of service undercut any expectation of confidentiality, and counsel had not directed the use. (United States v. Heppner, No. 25 Cr. 503 (JSR) (S.D.N.Y. Feb. 17, 2026).) On the other, a court has held that a self-represented litigant’s AI-assisted drafting remained protected work product, reasoning that AI programs are “tools, not persons” and that using one does not, by itself, waive the protection. (Warner v. Gilbarco, Inc., No. 2:24-cv-12333 (E.D. Mich. Feb. 10, 2026).)
Read together, the two address different questions: one is whether AI use can create a protection that was never there, and the other is whether ordinary work-product protection survives the use of a drafting tool. In both, the factor that appears to matter is the same one the confidentiality guidance emphasizes, namely the terms under which the tool operates and whether a lawyer is involved.
What has Arkansas done?
Arkansas has formally addressed AI in two respects:
Administrative Order No. 25 (adopted Dec. 11, 2025) is the Arkansas Supreme Court’s first order directed at AI in the courts. It is an administrative measure that applies existing duties in the AI context rather than a new rule of professional conduct. As we read it, it has two parts: an awareness section, applicable broadly to those participating in the court system, cautioning that some AI tools retain and train on the data entered into them and that careless disclosure of confidential or sealed information can implicate the conduct rules and the orders limiting access to court records; and a prohibition section, directed at court personnel and others with access to the courts’ internal systems, against exposing internal court data to a public-model AI tool absent committee approval. The order distinguishes between publicly hosted models and private, organization-controlled deployments. Notably, it does not impose a blanket requirement that attorneys disclose AI use in their filings.
Alongside the order, the Arkansas Supreme Court amended its Rules of Professional Conduct in 2025, in a per curiam addressing Rule 5.3 (responsibilities regarding nonlawyer assistance) and the Preamble. (In re Amendments to the Arkansas Rules of Professional Conduct, 2025 Ark. 115.) As we read the amendment, it places AI output within the supervising lawyer’s existing responsibility, which forecloses the argument that a tool, rather than the lawyer, was at fault.
Two practical notes follow for lawyers handling Arkansas matters. First, Arkansas does not yet have a freestanding formal ethics opinion on AI; the conduct rules and AO 25 are the current sources, and the Supreme Court’s automation committee has signaled that further work is underway. Second, the federal trial courts in Arkansas do not appear to have a district-wide AI standing order, but individual judges may set their own requirements, so it is worth checking the assigned judge’s practices before filing.
What comes next?
A few developments are worth watching, although none appear settled at the time of publication:
California has advanced proposed amendments to six of its Rules of Professional Conduct (1.1, 1.4, 1.6, 3.3, 5.1, and 5.3) that would, among other things, expressly require verification of AI-assisted citations. As of this writing the comment period has closed and the proposal remains pending; it is not adopted, and it would still require action by the California Supreme Court. If adopted, the amendments would carry disciplinary force, rather than the advisory weight most current AI guidance has.
On the evidence side, a proposed Federal Rule of Evidence 707 would subject certain machine-generated evidence to a reliability gate similar to the one that governs expert testimony. It remains in the rulemaking pipeline; the public comment period has closed, the proposal is under committee review, and the earliest it could take effect is late 2027.
Finally, the reliability research that underlies much of the verification guidance continues to develop. One widely cited study reported that leading legal-research AI tools still returned a meaningful share of incorrect or fabricated answers, on the order of 17 percent for one tool and 33 percent for another in the tested set. (Magesh et al., 22 J. Empirical Legal Stud. 216 (2025).) That figure is one data point, and the tools change, but it helps explain why the verification duty has been the most consistent theme in the guidance.
Where things appear to stand
Pulling the threads together, the guidance we have reviewed is consistent on a small set of points and unsettled on the rest. It is consistent that the existing rules govern, that a lawyer must understand a tool before using it, that AI output must be verified before it is filed, that client confidentiality drives tool selection, and that the supervising lawyer remains responsible for the result. It is unsettled, and jurisdiction-specific, on whether and when AI use must be disclosed to a client or a court, on how AI efficiency affects fees, and on how the rules apply to tools that act on their own.
The articles on this site are published by Dobson & Vines, PLLC for general informational purposes only, carry no guarantees of accuracy, and are not legal advice. Reading them, or contacting the firm, does not create an attorney-client relationship. The law changes over time and turns on the specific facts of each matter; do not act on anything here without advice from a qualified attorney.