IRS Rules That Bite When AI Helps Prepare a Return

The IRS has not published a rule about artificial intelligence in return preparation. It does not need one. Four existing provisions already decide most of the question, and the one that catches firms first is not the accuracy rule — it is the consent rule for disclosing return information to a third party.

There is no AI-specific IRS guidance

As at the time of writing, the Internal Revenue Service has issued no revenue ruling, revenue procedure, notice or Circular 230 amendment that addresses the use of generative AI by a paid preparer as such. The proposed amendments to Circular 230 published in the Federal Register on 20 December 2024 (REG-116610-20) modernise several practice rules but are technology-neutral; they do not create an AI category. That silence is itself the answer: the obligations are drafted around the preparer, not around the tools the preparer uses, so a model that drafts, extracts or reviews changes nothing about who signs and who is liable.

Circular 230 is the practice standard

Circular 230 is Treasury Department Circular No. 230, codified at 31 C.F.R. Part 10, and it governs practice before the IRS by attorneys, CPAs, enrolled agents and enrolled actuaries. Four sections do the work here. Section 10.22 imposes a duty of diligence as to accuracy in preparing and filing returns and in any written representation to the Service. Section 10.34(d) permits reliance in good faith on information furnished by the client, but not where the information appears incorrect, inconsistent or incomplete — a standard that a model’s confident extraction of a figure from a document does not satisfy on its own. Section 10.35 requires competence, meaning the knowledge, skill, thoroughness and preparation appropriate for the matter. Section 10.36 puts responsibility on those with principal authority for a firm’s practice to have adequate procedures to ensure compliance, which is the provision that turns “a junior used a chatbot” into a firm problem. The text is on the eCFR.

Read together, these say that the preparer’s duty is unchanged in substance and heavier in practice. A tool that produces plausible output quickly increases the volume of material a preparer must diligently check, and Circular 230 has no de minimis exception for output somebody else’s software generated.

Section 6695(g) diligence and Form 8867

For returns claiming the earned income credit, the child tax credit and related credits, the American opportunity credit, or head-of-household filing status, IRC § 6695(g) and Treasury Regulation § 1.6695-2 impose a specific paid-preparer due-diligence regime with a per-failure penalty, indexed annually. The regulation requires four things: completing Form 8867 and submitting it with the return; computing the credit using a worksheet or equivalent; knowing the applicable law and not knowing or having reason to know that information used is incorrect, inconsistent or incomplete; and keeping the specified records, including a record of how and when information was obtained, for three years.

The third element is where AI-assisted workflows fail. “Reason to know” is an objective standard, and the regulation requires the preparer to make reasonable inquiries and to contemporaneously document them. A model that summarises a client’s intake documents and proposes a filing status has produced neither an inquiry nor a record of one. The fourth element compounds it: the retention requirement covers the questions asked and the client’s answers, so if the questions were generated and the answers were parsed by a tool that does not persist them, the record does not exist.

Section 7216: sending return data to a model

This is the provision most likely to be violated on day one, and it is usually the last one anybody reads. IRC § 7216 makes it a criminal misdemeanour for a return preparer to knowingly or recklessly disclose or use tax return information other than as permitted, and IRC § 6713 imposes a parallel civil penalty per disclosure. The implementing rules in Treasury Regulation §§ 301.7216-1 through 301.7216-3 define “disclosure” and “use” broadly and require the taxpayer’s prior written consent, in a prescribed form and with prescribed content, for disclosures that are not on the permitted list.

Pasting a client’s W-2 figures, K-1 or narrative facts into a hosted model is a disclosure of tax return information to a third party. The auxiliary-services provisions permit some disclosures to contractors who assist in preparing the return, but they carry conditions — including that the recipient be bound not to make further disclosure or use — and a consumer chatbot accepted under standard terms is unlikely to meet them. The regulations are on the eCFR at 26 C.F.R. Part 301.

Safeguards, the WISP and the Security Six

Paid preparers are “financial institutions” for the purposes of the FTC Safeguards Rule, 16 C.F.R. Part 314, whose amended requirements have been fully effective since June 2023. The Rule requires a written information security program with a named qualified individual, a risk assessment, access controls, encryption of customer information in transit and at rest, secure development practices for applications used to transmit customer information, multi-factor authentication, and oversight of service providers by contract. IRS Publication 4557 and Publication 5708 give the Service’s own restatement of this, including the template written information security plan most small firms use.

Adding an AI vendor is, under this Rule, adding a service provider with access to customer information: it needs to appear in the risk assessment, in the vendor oversight process, and in the incident response plan. A breach at that vendor is a breach you must handle. This is the same structural obligation that PCI DSS imposes when a model touches payment data and that New York DFS spells out for its covered entities, and the evidence you would gather to satisfy one largely satisfies the others.

Where the guidance genuinely is silent: whether a signature on a return substantially drafted by a model raises any distinct issue under the preparer signature rules, whether AI-assisted preparation must be disclosed to the client, and how the Service will treat reliance on a tool in abating a § 6694 understatement penalty. Those are open. Do not let anybody tell you otherwise.