Can I use ChatGPT for tax research if I remove the client's name?

Removing the name is a start, not the finish. Figures, dates and circumstances can still identify a client or amount to tax return information under section 7216. Frame the question as a genuine hypothetical with altered numbers, or use a vetted business-tier tool with training switched off. On a free consumer tier, keep it to general how-does-this-rule-work research.

Does Internal Revenue Code section 7216 apply to AI tools specifically?

The statute predates modern AI, but it is written broadly enough to cover it. Section 7216 penalises the disclosure or use of tax return information without proper consent, regardless of the channel. Sending that information to a third-party AI provider is a disclosure. The tool being new does not create an exception, so the consent and use rules apply as they would to any other outside party.

Is an enterprise AI plan enough to satisfy my confidentiality duty?

It helps a great deal, but it is not automatic. A business or enterprise tier with a signed data processing agreement, no training on your data and short retention closes most of the technical gap. Your professional duty still requires you to assess the specific tool, keep the client informed where needed and confirm the settings actually match the contract. Read the agreement rather than trusting a marketing page.

What are the penalties if a preparer discloses tax return information improperly?

Under section 7216 an unlawful disclosure or use is a criminal misdemeanour, punishable by up to one year of imprisonment, a fine of up to 1,000 dollars, or both, together with the costs of prosecution. The civil companion, section 6713, adds 250 dollars per disclosure up to 10,000 dollars a year. Beyond the statute, an AICPA confidentiality breach can trigger disciplinary action and state board consequences.

Can I put a client's financial statements into AI to write a summary?

Not into a general consumer tool. A financial statement tied to a named entity is confidential client information, and for a tax engagement it is likely tax return information as well. If you need AI help summarising it, use a vetted enterprise tool with no training and a data agreement, and consider whether the client's consent is required first. When in doubt, anonymise heavily or keep it out.

Do I need client consent to use AI on their work?

Often yes, and the form matters. Section 7216 requires specific written consent, obtained before disclosure, for tax return information going to an outside party. The ABA's Formal Opinion 512 says boilerplate consent in an engagement letter is not adequate for law firms. Even where consent is not strictly required, telling clients how you use these tools is good practice. Confirm the exact requirement with your own regulator.

Explainer

What can accountants safely put into AI, and what must stay out?

September 22, 2026 · 8 minute read

The short answer

Keep anything that identifies a client out of a general consumer AI tool: names, Social Security numbers, financial statements and tax return information. That data is protected by the AICPA confidentiality rule and Internal Revenue Code section 7216, which carries criminal penalties. Anonymised questions, general research and first-draft copy are fine. Confirm the specifics against your own regulator or state board.

The two duties that decide what stays out

Before you weigh any tool's settings, start with the duties you already carry. For a CPA in public practice, the AICPA Confidential Client Information Rule, section 1.700.001 of the Code of Professional Conduct, is blunt: a member shall not disclose confidential client information without the client's specific consent. Pasting a client's figures into a tool that stores them, or uses them to improve a model, is a disclosure. The rule does not care that the recipient is software rather than a person.

The second duty is narrower and sharper. Internal Revenue Code section 7216 is a criminal statute aimed at tax return preparers. Knowingly or recklessly disclosing or using tax return information for any purpose other than preparing the return, without the taxpayer's consent, is a misdemeanour. The maximum is up to one year of imprisonment, a fine of up to 1,000 dollars, or both, together with the costs of prosecution. Its civil companion, section 6713, adds a penalty of 250 dollars per disclosure, capped at 10,000 dollars a year. The phrase tax return information is broad. It covers essentially anything the client gives you in connection with the return, not just the finished 1040.

Law firms sit under a parallel regime. Model Rule 1.6 governs confidentiality, and in July 2024 the ABA issued Formal Opinion 512, its first formal guidance on generative AI. The opinion flags self-learning tools as a specific disclosure risk and says boilerplate consent buried in an engagement letter is not adequate. The underlying principle holds across both professions. This article is general information, not advice, so confirm the exact requirements against your own state board, your bar and your own adviser.

What must never go into a general consumer tool

The test is simple. If a piece of information identifies a client, or reached you because you prepare their return, treat it as radioactive in a consumer chatbot. That includes anything that could be reassembled into an identity even when the obvious name is stripped out.

Keep these out of any tool you have not vetted:

Names, addresses, dates of birth and any client contact details
Social Security numbers, EINs, ITINs and bank or account numbers
Full financial statements, trial balances and general ledgers tied to a named entity
Anything drawn from a tax return: income figures, deductions, schedules, prior-year data
Payroll records, K-1s, W-2s and 1099s
Legal matter details, case facts and privileged communications for law firms
Uploaded source documents: PDFs of returns, engagement letters, bank statements

What is generally safe to put in

A confidentiality duty is not a ban on the technology. It is a ban on the data. Strip the identifying detail and most of the day-to-day value survives. The tool does not need to know whose numbers these are to help you reason about them.

In practice the safe lane is research, structure and drafting where no real client data appears.

These are low risk in a general tool:

General tax and accounting research: how a rule works, where to look in the code
Anonymised or fully hypothetical scenarios that cannot be traced to a client
First drafts of client emails, proposals and engagement copy with no figures
Explaining a concept to a client in plainer language, written generically
Formatting, formulas and spreadsheet logic with dummy numbers
Marketing, blog posts and social content for your own practice
Summarising a public document, such as published IRS guidance

Consumer, no-training and enterprise tiers are not the same product

The single biggest control you have is which version of a tool you use. The free consumer tier and the enterprise tier of the same brand can treat your data very differently, and that gap is the whole point.

On many consumer tiers your conversations may be used to improve the provider's models unless you turn that setting off. On business, enterprise and education tiers, the major providers state that your data is not used for training by default. OpenAI's enterprise privacy documentation says this plainly, with one honest caveat: if an organisation explicitly opts in to share data, it can still be used. So the tier matters, and the settings within the tier matter too.

Before you trust a tool with anything sensitive, confirm in writing:

Training: is your input used to train or improve the model, and can you switch it off
Tier: are you on a business or enterprise plan with a data processing agreement, not the free consumer app
Retention: how long is your data stored, and can you set it to zero retention
Access: who at the provider can see prompts, and is human review of your content possible
Sub-processors: where is the data hosted and which third parties touch it
Contract: is there a signed agreement, because a duty to a client is not satisfied by a webpage promise

The tool will invent a tax position with total confidence

Confidentiality is the compliance risk. Accuracy is the malpractice risk, and it is easy to underrate. A general model will produce a fluent, authoritative answer to a tax question and cite a code section, a revenue procedure or a case that does not exist or does not say what the model claims. It has no obligation to be right and no awareness when it is wrong.

There is precedent from the neighbouring profession. In 2023 a federal court in the Southern District of New York sanctioned lawyers who filed a brief containing AI-generated case citations that turned out to be fabricated. The parallel for a tax practice is direct. A hallucinated section number in a client memo, or a made-up treatment relied on in a return, is your professional exposure, not the vendor's.

Treat every substantive output as an unverified draft from a confident junior. Check the citation against the primary source. Confirm the rule still stands. The tool speeds up the first stretch of the thinking. The part that carries your name and your licence stays with you.

Common questions

Can I use ChatGPT for tax research if I remove the client's name?: Removing the name is a start, not the finish. Figures, dates and circumstances can still identify a client or amount to tax return information under section 7216. Frame the question as a genuine hypothetical with altered numbers, or use a vetted business-tier tool with training switched off. On a free consumer tier, keep it to general how-does-this-rule-work research.
Does Internal Revenue Code section 7216 apply to AI tools specifically?: The statute predates modern AI, but it is written broadly enough to cover it. Section 7216 penalises the disclosure or use of tax return information without proper consent, regardless of the channel. Sending that information to a third-party AI provider is a disclosure. The tool being new does not create an exception, so the consent and use rules apply as they would to any other outside party.
Is an enterprise AI plan enough to satisfy my confidentiality duty?: It helps a great deal, but it is not automatic. A business or enterprise tier with a signed data processing agreement, no training on your data and short retention closes most of the technical gap. Your professional duty still requires you to assess the specific tool, keep the client informed where needed and confirm the settings actually match the contract. Read the agreement rather than trusting a marketing page.
What are the penalties if a preparer discloses tax return information improperly?: Under section 7216 an unlawful disclosure or use is a criminal misdemeanour, punishable by up to one year of imprisonment, a fine of up to 1,000 dollars, or both, together with the costs of prosecution. The civil companion, section 6713, adds 250 dollars per disclosure up to 10,000 dollars a year. Beyond the statute, an AICPA confidentiality breach can trigger disciplinary action and state board consequences.
Can I put a client's financial statements into AI to write a summary?: Not into a general consumer tool. A financial statement tied to a named entity is confidential client information, and for a tax engagement it is likely tax return information as well. If you need AI help summarising it, use a vetted enterprise tool with no training and a data agreement, and consider whether the client's consent is required first. When in doubt, anonymise heavily or keep it out.
Do I need client consent to use AI on their work?: Often yes, and the form matters. Section 7216 requires specific written consent, obtained before disclosure, for tax return information going to an outside party. The ABA's Formal Opinion 512 says boilerplate consent in an engagement letter is not adequate for law firms. Even where consent is not strictly required, telling clients how you use these tools is good practice. Confirm the exact requirement with your own regulator.