What Published IRS Guidance, Tax Court Opinions, and Treasury Proposed Rulemaking Through October 2026 Actually Say About AI-Assisted Tax Legal Work Product — Privilege, Disclosure Obligations, and the Penalty Exposure When AI-Generated Positions Turn Out to Be Wrong
Overview
The intersection of artificial intelligence and tax practice has generated substantial professional commentary but, through October 2026, comparatively sparse formal government output. No IRS Chief Counsel memorandum, Treasury proposed regulation, or Tax Court opinion has yet addressed AI-assisted tax work product as a discrete doctrinal category. What exists instead is a layered architecture of pre-existing rules — attorney-client privilege doctrine, the work-product protection articulated in In re Kellogg Brown & Root, Circular 230, the Section 6662 accuracy-related penalty regime, and the Section 6664 reasonable-cause safe harbor — each of which applies to AI-generated analysis under its own existing terms, without modification. This briefing explains what those existing terms actually require and where the analysis remains genuinely unsettled.
I. Privilege and Work-Product Protection: The Existing Framework Applied to AI Output
Attorney-Client Privilege
The attorney-client privilege as applied in federal tax proceedings is governed by federal common law and, where applicable, by the statutory tax practitioner privilege under IRC Section 7525. Section 7525 extends a privilege analogous to the attorney-client privilege to communications between a taxpayer and a federally authorized tax practitioner — but only with respect to tax advice, and explicitly not with respect to written communications in connection with the promotion of a tax shelter (IRC § 7525(b)).
Neither the privilege under federal common law nor Section 7525 turns on the identity of the instrument that generated the communication's content. The question the privilege asks is whether the communication was (1) between privileged persons, (2) made in confidence, (3) for the purpose of obtaining or providing legal advice. An attorney who reviews and adopts AI-generated tax analysis as the basis for a memorandum to a client does not, solely for that reason, strip the communication of privilege — the attorney remains the privileged person who is providing legal advice. The AI tool is, in this analysis, more analogous to a legal research database than to an independent advisor whose involvement could break the confidential relationship.
What could affect privilege is disclosure. If AI-generated output is shared with a third-party model provider in a manner that allows the provider to access the underlying tax facts, courts may treat that sharing as a voluntary disclosure to a third party outside the privileged relationship. No federal court has yet ruled on this theory in the tax context specifically, and the IRS has not issued guidance on it, but the governing principle — that disclosure to a third party not within the common-interest exception ordinarily waives the privilege — is established doctrine.
Work-Product Protection
The work-product doctrine in federal practice follows the "because of" test articulated in United States v. Adlman, 134 F.3d 1194 (2d Cir. 1998), requiring that the document was prepared because of the prospect of litigation, not merely that it was useful in litigation. Separately, the D.C. Circuit's "one of the significant purposes" formulation in In re Kellogg Brown & Root, Inc., 756 F.3d 754 (D.C. Cir. 2014), informs the attorney-client privilege analysis where a communication serves both business and legal purposes — a related but doctrinally distinct question from the work-product "because of" test.
Applied to AI-generated tax analysis, the doctrine again does not distinguish by authorship. A tax controversy memo drafted with the assistance of an AI tool qualifies for work-product protection to the same extent a manually drafted memo would, provided it was prepared because of anticipated litigation. What practitioners should recognize, however, is that the metadata trail generated by AI platforms — query logs, model prompts, session histories — may not itself be protected as work product if those logs are not embedded in a document prepared in anticipation of litigation. Courts analyzing electronically stored information have not yet addressed AI session logs specifically in the tax context, and the IRS has issued no guidance on the discoverability of AI prompts in examination or Tax Court proceedings.
II. Circular 230 Disclosure Obligations and the Duty of Competence
What Circular 230 Currently Requires
Treasury's Circular 230 (31 C.F.R. Part 10) governs practice before the IRS and imposes duties on covered practitioners — attorneys, CPAs, enrolled agents, and others — through provisions that have never been updated to address AI tools specifically. As of October 2026, no Treasury proposed regulation amending Circular 230 to address AI-generated tax advice has been published in the Federal Register.
The applicable standards are the pre-existing ones. Section 10.37 governs requirements for written advice, requiring that a practitioner base written advice on reasonable factual and legal assumptions, reasonably consider all relevant facts and circumstances the practitioner knows or should know, and not rely on representations of the taxpayer or others if the practitioner knows or should know that those representations are incorrect. Critically, Section 10.37 makes no exception for advice that reflects the output of an automated tool: the practitioner remains responsible for the advice regardless of how it was generated.
Section 10.22 imposes a due diligence standard, requiring practitioners to exercise due diligence in preparing returns and in determining the correctness of representations made to the IRS. Section 10.34 addresses standards for tax returns and advice on positions taken on returns, prohibiting a practitioner from recommending a position that the practitioner knows or should know lacks a reasonable basis. These standards all vest liability in the practitioner, not in the tool.
The Competence Obligation
Section 10.35 requires that a practitioner possess the legal knowledge, skill, thoroughness, and preparation reasonably necessary for the representation. The ABA's first formal guidance on generative AI and professional obligations — ABA Formal Opinion 512, issued July 29, 2024 — covers competence, confidentiality, communication, candor, supervision, and fee duties, and explicitly calls for independent verification of AI-generated output. While ABA formal opinions do not bind tax practitioners who are not attorneys, the competence principle is substantively mirrored in Section 10.35. A practitioner who transmits AI-generated tax analysis to a client without independent verification is at risk of a Section 10.35 violation if that analysis turns out to be legally incorrect.
The USPTO's guidance published in the Federal Register on April 11, 2024 (89 FR 25609) — requiring practitioners to review AI-generated content before submitting it — reflects the same underlying norm across federal practice contexts: AI output generates personal professional liability for the practitioner who adopts it.
III. Section 6662, Section 6664, and the AI Accuracy Question
The Accuracy-Related Penalty
IRC Section 6662 imposes a 20 percent accuracy-related penalty on the portion of an underpayment attributable to, among other things, negligence, disregard of rules or regulations, or a substantial understatement of income tax. A substantial understatement exists if the understatement exceeds the greater of 10 percent of the tax required to be shown on the return or $5,000. The penalty rises to 40 percent for a gross valuation misstatement.
Nothing in Section 6662's text or in Treasury Regulation Section 1.6662 varies the standard based on whether a tax position was developed with AI assistance. The analysis the IRS and Tax Court conduct is functional: was there substantial authority for the position, and was adequate disclosure made?
The Reasonable Cause Defense Under Section 6664
Section 6664(c)(1) provides that no accuracy-related penalty is imposed with respect to any portion of an underpayment if it is shown that there was reasonable cause for the taxpayer's position and the taxpayer acted in good faith. Treasury Regulation Section 1.6664-4 provides that the most important factor is the extent of the taxpayer's effort to assess the proper tax liability, and that reliance on the advice of a competent tax advisor can support reasonable cause — provided the advisor had sufficient expertise, was provided all relevant facts, and the taxpayer reasonably relied on the advice.
Here is where AI use introduces genuine doctrinal risk. If a practitioner relies on AI-generated analysis without independent verification, and that analysis contains a hallucinated or incorrect legal proposition, the practitioner may not qualify as a "competent tax advisor" whose advice supports the taxpayer's reasonable cause defense. The Tax Court has not yet addressed this directly, but the existing framework in Treas. Reg. § 1.6664-4 requires that the advisor actually analyze the relevant facts and applicable law. AI-generated analysis that is adopted wholesale — without the practitioner's independent legal judgment being applied — may fail that test.
IV. The Absent Guidance and Its Practical Consequences
The core doctrinal reality through October 2026 is that neither the IRS nor Treasury has issued AI-specific guidance on any of these three questions. Practitioners are operating under general frameworks that apply by analogy. The privilege analysis tracks existing third-party disclosure doctrine. The Circular 230 analysis tracks the existing written-advice and competence standards. The penalty analysis tracks the existing reasonable cause framework — where AI-generated positions are most vulnerable when the practitioner cannot demonstrate independent legal analysis underlying the advice.
The practical implication is straightforward: in every context examined here, the existing rules impose liability on the practitioner, not the tool. AI output that is verified, integrated into a practitioner's independent analysis, and communicated through proper channels is treated the same as manually generated analysis. AI output that is transmitted without verification or independent judgment exposes both the practitioner and the client under standards that have not needed amendment to reach that result.
The Legal Stack is an independent research publication. This briefing is analytical commentary, not legal advice. All doctrinal claims are attributed to publicly available sources; no proprietary data or survey findings are cited.
Filed under Compliance & Regulation → · The Legal Stack accepts no vendor funding for its research.
More Research
View all →10 min
10 min
10 min