AI in Tax Preparation

General-purpose models are unreliable for tax answers

A study posted on SSRN entered common tax questions into ChatGPT (models 3.5 and 4) during the 2023 and 2024 tax seasons and found the overall share of correct responses ranged from 39% to 47%. The authors concluded the chatbot is "generally not a reliable source of tax guidance for uninformed taxpayers," and that accuracy was worse for common questions, complex answers, fact-pattern evaluation, and post-cutoff tax changes.

In Tax Notes, researchers tested GPT-4 on the SARA dataset of statutory tax cases. GPT-4 answered 186 of 276 true/false cases correctly (67%), and calculated the exact tax liability correctly only about one-third of the time, miscalculating by more than 10% in nearly a quarter of cases. Notably, the errors were not mathematical—they involved misreading the statutes. That finding is important: the model failed at legal interpretation, not arithmetic.

Accuracy depends heavily on task, domain, and retrieval

A Stanford Law white paper testing LLMs on multiple-choice tax law exams found that GPT-4—especially when combined with chain-of-thought and few-shot prompting and supplied with the correct "gold truth" legal text—could reach high accuracy, but "not yet at expert tax lawyer levels." Studies of Austrian and EU VAT law reported accuracy ranging widely with configuration; even under an idealized "perfect retrieval" setup, accuracy for determining where tax was owed reached only about 73%, and the authors concluded that a 20%–30% error rate is "unacceptable" for real practice and that an LLM-only assistant is "insufficient for practice."

Study-bound figures: Reported LLM accuracy in legal decision-making has ranged from roughly 19% to 98% depending on the domain, task type, and whether the model was purpose-built and supplied with authoritative text. A single headline accuracy number is meaningless without its task and conditions.

Hallucinated citations are a documented, sanctioned problem

The risk is not hypothetical. In the 2023 case Mata v. Avianca, attorneys relied on ChatGPT-generated research and submitted fictitious case law, resulting in judicial sanctions. A comparative analysis in The International Tax Journal reports that aggregated datasets recorded roughly 800 documented cases of AI-related citation errors across at least 25 jurisdictions by late 2025, with a sharp increase in 2025. A separate experimental study of tax-law reasoning found a strong hallucination effect affecting about 50% of analyzed court-ruling cases.

Reconciling research with product design

These findings do not mean AI is useless in tax prep—they explain why well-designed products constrain it. Established tax software performs the tax calculations through programmed logic, uses AI mainly for document extraction and workflow, and keeps a human expert as the signing preparer for complex returns. The empirical weakness of generative models on open-ended legal reasoning is precisely why the reliable uses are narrow and supervised, and why the IRS insists on human validation.

Preparer due diligence

Treasury Regulation section 1.6695-2 sets out four due-diligence requirements for paid preparers claiming the EITC, CTC/ACTC/ODC, AOTC, or head-of-household status: completing and submitting Form 8867, computing credits with worksheets, not knowing (or having reason to know) that information is incorrect, and keeping specified records. The IRS is explicit that tax preparation software "helps, but it cannot replace your professional judgment or responsibility under the law." The preparer must still ask reasonable questions, evaluate whether information makes sense, and follow up on anything inconsistent, incomplete, or suspicious—and decline to prepare the return if not satisfied.

Preparers must also make additional reasonable inquiries when a knowledgeable preparer would conclude that information appears incorrect, inconsistent, or incomplete, and must document those inquiries and the client's answers at the time of the interview. These records must be kept for three years.

Circular 230 and AI

The IRS's Office of Professional Responsibility has directly addressed AI. Its Circular 230 guidance lists the standards implicated by AI use—including section 10.22 (due diligence), section 10.34 (standards on returns and documents), section 10.35 (competence), and section 10.37 (written advice)—and states that due diligence is required in assessing the reliability of AI-created results. It warns that AI chatbots fabricate facts, that there is often no transparency about the source of an AI response, and that practitioners "can't accept the responses from AI tools" but must validate them against actual court cases, citations, and tax law.

Data security is mandatory

Tax data is highly sensitive, and protecting it is a legal requirement, not an option. Federal law—under the Gramm-Leach-Bliley Act and the FTC Safeguards Rule—requires tax return preparers to create and maintain a written information security plan to protect client data, regardless of firm size. IRS Publication 4557 provides the framework, including implementing multi-factor authentication, encrypting sensitive files and emails, limiting data access to those who need it, maintaining audit logs of who did what and when, and vetting service providers so that contracts require them to maintain appropriate safeguards.

This last point is critical for AI: before entering taxpayer data into any AI tool, a preparer must know whether the vendor is a service provider under the security plan, what happens to the data, whether it is used to train shared models, and how it is protected. Unauthorized disclosure of taxpayer information can trigger civil and criminal penalties under IRC sections 6713 and 7216, which the Circular 230 guidance also flags in connection with AI use.

Hallucinated or incorrect tax conclusions

A generative model can produce a confident, well-written answer that misstates the law or cites nonexistent authority. Because the errors are often interpretive rather than mathematical, they can be hard to spot.

Overreliance and automation bias

A polished interface or a "100% accuracy" label can discourage substantive review. The IRS is explicit that software cannot replace professional judgment.

Stale or out-of-scope knowledge

Research shows models perform worse on tax changes occurring after their knowledge cutoff and on questions requiring current-year specifics. Tax law changes annually.

Data security and unauthorized disclosure

Entering taxpayer data into an unvetted AI tool can violate the firm's information security plan and potentially IRC sections 6713 and 7216.

Weak documentation and audit trail

Due-diligence records must show what information was used, how it was obtained, and what inquiries were made. AI-generated content does not satisfy this on its own.

Question 1: Which tax-prep task is generally the most reliable AI use?

Question 2: What did research find about general-purpose models answering tax questions?

Question 3: Under IRS guidance, what must a practitioner do with an AI-generated tax answer?

Question 4: What is legally required before a paid preparer enters client data into an AI tool?

Question 5: Which statement best describes responsible AI-assisted tax preparation?