Can I put client tax data into ChatGPT?
Not into a personal ChatGPT account without each client’s consent.
Federal tax law penalizes a preparer who shares a client’s return information, or uses it for anything but the return. The sharing happens when the data leaves your firm, whatever the AI company does with it later. A personal account, paid Plus and Pro plans included, needs each client’s consent first. A business account may fit an exception for outside help, but only if everyone who receives or views the data is in the United States. Until the vendor commits to that, get each client’s consent, and have IT record where the data is stored and where the AI runs. The same test applies to Claude, Copilot or any other AI tool. It turns on the account type, the contract, and where the people who receive or view the data are.
Not into a personal account, without each client’s consent. 26 U.S.C. 7216(a) makes it a misdemeanor for a return preparer to knowingly or recklessly disclose information furnished for the preparation of a return, or to use it for any purpose other than preparing the return, punishable on conviction by a fine of up to $1,000, or imprisonment of up to one year, or both, and by a fine of up to $100,000 where section 6713(b) applies.
26 U.S.C. 6713(a) adds a civil penalty of $250 for each disclosure or use, capped at $10,000 in a calendar year, and it requires no showing of knowledge or recklessness.
A contracted account changes the analysis. 26 CFR 301.7216-2(d)(1) permits disclosure to another return preparer located in the United States for preparation or for auxiliary services in connection with preparation, without consent, and 26 CFR 301.7216-1(b)(2)(i)(B) treats a provider of those auxiliary services as a return preparer. A consumer account with no contract does not fit that provision. Located means where the people who receive or view the information are, not where the provider is headquartered.
Does it matter whether the AI trains on what we type?
Firms often ask whether the AI provider trains its models on what staff type in. It’s a fair question, but it belongs to a different rule. The training setting matters under the FTC Safeguards Rule, at 16 CFR 314.4(f). It also bears on your firm’s duty to limit how long it keeps data, at 16 CFR 314.4(c)(6).
The setting doesn’t decide the §7216 question, because 26 CFR 301.7216-1(b)(5) defines disclosure as making tax return information known to any person in any manner. The disclosure happens when the information leaves the preparer, whatever the provider does with it later.
The two conditions for sharing without consent
The exception for auxiliary services, meaning outside help with preparing the return, has two conditions.
- The first is that the provider is located in the United States. That’s because the rules treat the provider, and its employees who assist, as preparers, and 26 CFR 301.7216-2(d)(1) covers only a preparer located there. Located means where the people who receive or view the information are, not where the provider is headquartered. Before tax return information goes to a preparer outside the United States, consent under 26 CFR 301.7216-3 is required.
- The second condition is that the service helps with preparation and doesn’t make substantive determinations or give advice affecting the tax liability reported. An AI that drafts a workpaper summary for a preparer to check is the case a firm would argue fits. A firm wouldn’t argue the same for an AI that decides a position, which the preparer then signs without review. Neither case is settled, and this page doesn’t decide either one.
Where the data is stored and where the AI runs are two separate settings, and they’re set workspace by workspace. 26 CFR 301.7216-2(d)(1) reaches only a recipient located in the United States, meaning where the people who receive or view the information are. For a firm relying on it, those two settings are the facts counsel will ask for, along with who at the vendor can view the data.
Whether automated processing on a server abroad, with no person viewing it, is a disclosure has not been decided. Claremont does not build on the argument that it is not.
At OpenAI, data residency stores a workspace’s content in a region you choose. Inference residency is a separate setting that keeps the AI itself running in that region. Data residency covers storage only, so on its own it leaves the AI running wherever OpenAI puts it.
Both settings are opt-in and limited to the business tiers, so a consumer account has neither. For ChatGPT Enterprise and Edu, the setting applies only to new workspaces.
What your firm should have on file
- The vendor’s enterprise terms, and the workspace’s own residency settings: where content is stored, and whether the AI is also set to run only in the United States. They’re set up separately.
- Your firm’s own written explanation of why the exception applies, kept with your risk assessment.
- A security evaluation of the tool itself, as software your firm uses. 16 CFR 314.4(c)(4)
- An AI acceptable-use policy that names the approved tools and the kinds of data each one may receive.
- Staff training that covers the tool, tied to the risk assessment. 16 CFR 314.4(e)
Related questions
Is a paid ChatGPT Plus subscription a business account?+
No. Plus and Pro are consumer tiers on a personal account. A business account is run by the firm and shows the firm’s workspace, and it’s the firm that accepted its data terms.
What if the data is de-identified first?+
Removing the identifiers before the text leaves the firm changes what is disclosed. The question then becomes whether what’s left is still tax return information, which depends on what was removed and on what could identify the client again.
Does a consent in the engagement letter cover it?+
Not if it’s a general clause. 26 CFR 301.7216-3 requires consent that is knowing and voluntary, signed and dated, and obtained before the disclosure. It must also identify the intended purpose and the specific recipients. A general clause in an engagement letter doesn’t meet that standard.
How many violations is one upload?+
One file holding thirty clients’ information can count as thirty disclosures, and 26 U.S.C. 6713(a) sets the civil penalty per disclosure or use.
The obligations on this page are three of the many that bind a firm holding client financial data. The map shows the rest, and which control satisfies each one.
Get a one-page estimateRun it on your own toolsStrip identifiers firstSee the obligation mapGeneral reference, not legal or tax advice. Every figure and deadline on this page was read against the primary source. Claremont Security performs assessments and writes documentation. Formal sign-offs belong to licensed firms.
