A client had a gray-area question. You ran it through an AI tax assistant, got a clean answer with citations attached, and it looked like a finished position. It's already in the file, maybe already on a return.
The question now isn't whether AI is good enough to use. It's whether what you already relied on would survive an IRS examiner, a malpractice claim, or your own partner asking you to defend it line by line.
That question has a precise legal answer, and it turns on one distinction most practitioners haven't had to think through yet: the applicable standard attaches to the authority you relied on, not the tool that found it for you. An authority an AI tool surfaces is exactly as good as that authority always was, checked and applied correctly. An authority an AI tool fabricated isn't weak, it's nothing, and it protects you exactly as much. This article covers the standards ladder that decides which one you're holding, the preparer penalty that runs on it, the malpractice exposure already attaching to this exact scenario, and the verification workflow that tells you which side of the line you're actually on.
Bizora's research traces every answer to the IRC, Treasury Regulations, IRS rulings, and case law, with a View Steps trail that lets you check that distinction in minutes instead of guessing.
IRC § 6662 imposes a 20% accuracy-related penalty on an underpayment attributable to negligence or a substantial understatement of tax, rising to 40% for gross valuation misstatements and certain other categories. For an undisclosed position, that substantial understatement penalty is avoided only if there was substantial authority for the position under § 6662(d)(2)(B)(i).
Treas. Reg. § 1.6662-4(d)(2) defines substantial authority as "an objective standard involving an analysis of the law and application of the law to relevant facts," explicitly "less stringent than the more likely than not standard… but more stringent than the reasonable basis standard." Practitioners commonly place it around a 40% likelihood of success, above reasonable basis (roughly 20%) and below more likely than not (over 50%). Reasonable basis alone doesn't avoid the penalty; it only works paired with adequate disclosure, filed on Form 8275 for a position that doesn't contradict a regulation, or Form 8275-R for one that does. If the AI-surfaced authority behind a position only clears the reasonable basis bar, disclosure isn't optional, it's the entire penalty defense.
The test is a weighing exercise, not a citation count. Under § 1.6662-4(d)(3)(i), substantial authority exists "only if the weight of the authorities supporting the treatment is substantial in relation to the weight of authorities supporting contrary treatment," and every relevant authority, including contrary ones, gets weighed. Because it's objective, what the taxpayer or preparer personally believed is irrelevant.
The regulation then lists exactly what counts, and the list is closed: the Code and other statutes, regulations, tax treaties and their official explanations, revenue rulings and procedures, court cases, committee reports, private letter rulings and technical advice memoranda issued after October 31, 1976, general counsel memoranda issued after March 12, 1981, and similar IRS pronouncements published in the Internal Revenue Bulletin. It draws the line that matters here in plain language: "Conclusions reached in treatises, legal periodicals, legal opinions or opinions rendered by tax professionals are not authority."
That sentence answers the article's title, but not the way it first sounds. The AI tool's own conclusion was never going to count regardless, the same as a named professional's opinion doesn't count. What decides the question is the primary authority underneath the AI's answer: the actual Code section, regulation, ruling, or case it points to.
Verify that authority exists, says what the tool claims, and applies to the facts, and it counts exactly as much as if a human had found it with a law library. The standard attaches to the authority, never to the research method. A hallucinated citation isn't a weaker version of that authority; there's no authority there at all, and the position built on it holds up exactly as well as a position built on nothing, which is to say it doesn't.
IRC § 6694(a) penalizes a return preparer, the greater of $1,000 or 50% of the income derived from the return, for an unreasonable position the preparer knew or should have known about. Treas. Reg. § 1.6694-2 sets three paths to a reasonable position: substantial authority for an undisclosed, non-shelter position; reasonable basis with adequate disclosure; or, for a tax shelter or reportable transaction, a reasonable belief the position would more likely than not be sustained. Willful or reckless conduct raises the penalty to the greater of $5,000 or 75% of income derived under § 6694(b).
An AI tax assistant satisfies none of these paths on its own, because it isn't authority. It cannot make a shelter position more likely than not, and a preparer who signs off on a hallucinated citation has, by definition, no authority for that position at all. The reasonable-cause relief in § 1.6694-2(e) turns on the preparer's own documented diligence, not on the fact that a tool was involved.
For taxpayers, the parallel reasonable-cause defense under IRC § 6664(c) and Treas. Reg. § 1.6664-4 is explicit that reliance on a professional advisor "does not necessarily demonstrate reasonable cause and good faith" unless that reliance was itself reasonable. "The AI told me so" protects nobody in that chain.
On June 24, 2026, the IRS Office of Professional Responsibility issued Alert Issue No. 2026-19, its first formal guidance on AI use in federal tax practice. It creates no new rules; it applies existing Circular 230 duties, and flags confidentiality exposure under IRC §§ 6713 and 7216, to a new fact pattern.
Under § 10.22 (due diligence), practitioners "must thoroughly review all AI-created documents and language" before delivery, and that review "cannot be delegated to an algorithmic process." Under § 10.35 (competence), technological literacy about how the AI tool actually works is now part of competent practice. Under § 10.37 (written advice), a practitioner "cannot rely on GAI projections or representations without verification," and where a tool's reasoning is opaque, relying on it anyway "may be unreasonable." Firm leaders carry § 10.36 obligations to train staff and vet AI tools; § 10.27 bars billing for AI-compressed work as if it still took the old amount of time.
The alert's own framing is direct: "Technology serves as a powerful tool, not a substitute for professional judgment." In practice, that means an examiner or an OPR inquiry can now reasonably ask how an AI-assisted position was verified, and a firm without a documented answer is exposed regardless of whether the underlying research was actually correct.
This isn't hypothetical. In Mata v. Avianca (S.D.N.Y. 2023), attorneys were sanctioned $5,000 after submitting a brief citing six nonexistent AI-generated cases. Since then, a public tracker of AI hallucination sanctions (the AI Hallucination Cases database) has logged well over a thousand cases, with penalties climbing sharply: a Wyoming court revoked an attorney's pro hac vice admission over fabricated citations from an in-house AI platform, and an Oregon federal court imposed a six-figure sanction against counsel who submitted fifteen nonexistent cases and eight fabricated quotations.
The exposure has reached tax practice directly. In Clinco v. Commissioner, T.C. Memo. 2026-16, Judge Mark Holmes identified that three of four cases cited in a filed brief appeared to be AI hallucinations, each citation pointing to a real reporter page occupied by an entirely different case. He imposed no sanction that time, noting a bit of embarrassment "is enough for now," then flagged separately that other courts have begun sanctioning lawyers for AI shortcuts and that the Tax Court has not done so, "Yet."
Tax practice hasn't had its own Mata moment yet, a case where an AI-hallucinated citation actually draws a monetary sanction. Judge Holmes's footnote is a direct signal that it's coming, and the standard applied will be the identical one already used against attorneys outside tax: the citation either exists and says what you claimed, or the sanction follows.
In Thomas v. Commissioner, the Tax Court struck a pretrial memorandum that relied on fabricated cases outright. Even an IRS attorney's own filing has triggered a hearing over a nonexistent citation, a reminder that the government's side carries the identical exposure.
Tax research software failures don't stay confined to the IRS relationship; they follow the same path as any other accounting malpractice claim. CPA malpractice insurer CAMICO has published AI-specific risk guidance distinguishing supervised AI use from autonomous AI that generates advice without human review, and recommends firms formally document which uses are authorized. CNA, which underwrites the AICPA's professional liability program, warns that overreliance on AI output paired with a failure to disclose it can support a claim that the CPA "lacked subject matter competence," and cites ABA Formal Opinion 512, which says attorneys should not rely on AI outputs "without independent verification or review," a standard the opinion ties to the duty of competent representation.
Malpractice insurers including CAMICO and CNA have signaled AI use is now a renewal-underwriting concern, and firms without a documented AI policy can expect tougher questions, and the trajectory looks like cyber coverage did a decade ago: disclosure questionnaires first, governance-contingent terms next. Accountant liability for an AI-sourced error is not a novel legal theory; accountant negligence claims already cover exactly this fact pattern, a professional who relied on a source without adequate verification.
Most practitioners have not actually read their own policy on this point. Before the question comes up in a claim, pull your professional liability policy and check whether it addresses AI-assisted work at all, whether disclosure is a condition of coverage, and whether an exclusion is already sitting in the fine print. Almost nobody has checked yet, which is exactly why it's worth ten minutes now.
Six steps, drawn directly from OPR Alert 2026-19, the AICPA's guidance, and how malpractice insurers describe adequate review:
The degree of scrutiny should scale with the stakes. A quick research question for internal orientation needs less verification than an answer going into a written position or a filed return. What doesn't scale down, ever, is the requirement to actually check.
An AI tax assistant is a research accelerant, not a source of authority, and treating it as the latter is the exact failure mode showing up in sanctions orders and malpractice guidance right now. Used correctly, it compresses the time between a question and the primary authority that answers it; the verification step is still the professional's, and it's the step that actually makes the position hold up.
Verification is the whole job either way, on Bizora or off it. What a clickable primary-source trail changes is the cost: instead of a fresh research project to confirm what the AI told you, checking the citation behind an answer becomes a minute of work. For more on what the IRS now expects from AI-assisted tax work, see our breakdown of IRS Issue 2026-19. Research your next gray-area position at bizora.ai and see the citation trail behind the answer before you rely on it.
The following authorities and guidance inform this article:
No, but the primary authority it points to can. The AI tool's own conclusion is not authority, the same way a tax professional's opinion isn't under Treas. Reg. § 1.6662-4(d)(3)(iii). What counts is the underlying Code section, regulation, ruling, or case, verified independently, and it counts exactly as much as if a human had found it.
Substantial authority is a higher, objective standard, roughly a 40% likelihood of success, that avoids the accuracy-related penalty on an undisclosed position without more. Reasonable basis, around 20%, is lower and only avoids the penalty if the position is adequately disclosed on Form 8275 or, for a position contrary to a regulation, Form 8275-R.
Yes. CPA malpractice insurers, including CAMICO and CNA, treat uncritical reliance on AI output as a governance and competence risk, and courts have already sanctioned attorneys for filing AI-hallucinated citations. The exposure sits with the professional who relied on the output, not the AI vendor.
It applies existing Circular 230 duties to AI use: due diligence review of AI output before it reaches a client or the IRS, documented verification of facts and citations, and firm-level procedures for AI oversight. It doesn't create new rules, it makes clear that existing duties already cover AI.
The Tax Court has flagged the problem directly. In Clinco v. Commissioner, the court identified likely AI-generated fake citations in a filed brief and warned that forbearance on sanctions won't last. Federal courts outside the tax context have already imposed six-figure penalties for the same failure.
Bizora AI turns hours of manual research into seconds, with every answer backed by primary source citations. Start your 7-day free trial. No credit card required.
Start Free TrialRelated articles