Can ChatGPT Analyze Your Tax Return? Here's the Catch

Can ChatGPT actually analyze a tax return?

Yes. ChatGPT can open a PDF, pull numbers off a W-2 or 1099, and produce a readable summary of your income, withholding, and possible deductions. OpenAI markets this directly: upload a document and "get instant answers, summaries, and insights through natural conversation."¹ The file-upload and code-interpreter features genuinely work for this. A CNBC reporter uploaded a consolidated 1099 from her brokerage and got back a confident, specific answer: "This is a very simple return with one stock plan wrinkle. You do NOT need a CPA."²

That sentence is the whole problem in miniature. ChatGPT can read the document. Whether it reads it correctly, and whether it should be trusted with it at all, are separate questions.

How accurate is ChatGPT on real tax calculations?

Not very. TaxCalcBench, an open benchmark built from 51 real federal tax returns, tested leading general-purpose models on full return calculations. The results were rough: GPT-5 with web search topped out at 41.67% strict accuracy, Gemini 2.5 Pro hit 32.35%, and Claude Opus 4 landed at 27.45%.³ That's a coin flip at best, on clean, correctly formatted data, not a messy scan of somebody's actual paperwork.

A peer-reviewed study in the Journal of Emerging Technologies in Accounting reached a similar conclusion: ChatGPT is generally not a reliable source of accurate tax information, and its accuracy gets worse on exactly the questions that matter most, the common, complex ones that depend on a taxpayer's specific facts.⁴

Back to that CNBC reporter: her CPA later caught reconciliation items the AI never flagged. The advice wasn't wrong exactly. It was "possibly correct but also incomplete," presented with the confidence of a full answer when it was a partial one.² Even a purpose-built AI tax platform, Filed, which runs a multi-agent system designed specifically for document-aware tax calculation, only reached 72.5% strict accuracy on the same benchmark and still needed human review on roughly a quarter of returns.³ If a tool built for exactly this job needs a human backstop, a general chatbot needs one more.

OpenAI itself doesn't pretend otherwise. Its usage policies explicitly restrict "automation of high-stakes decisions in sensitive areas without human review," naming financial activities and credit among the restricted categories.⁵ ChatGPT was never sanctioned by its own maker as an unsupervised tax tool.

What happens to your data after you hit upload

This is where the real story is, and it's not about math errors.

When you upload a return to free, consumer ChatGPT, your conversation and files can be used to train future models unless you've opted out. Business, Enterprise, and API tiers work differently: OpenAI states it does not train on that data by default, and enterprise customers control retention.⁶ Most people uploading their own 1040 are on the free or Plus tier, not the enterprise one.

Even the "your chats get deleted" promise has a caveat that matters. In 2025, a federal court order in the New York Times v. OpenAI litigation forced OpenAI to preserve every ChatGPT consumer and API output log indefinitely, including chats users had already deleted, overriding OpenAI's own 30-day deletion policy.⁷ OpenAI's standard practice is to auto-delete conversations within 30 days once they're not under legal hold,⁸ and that broader preservation order was lifted on September 26, 2025.⁸ But a slice of user data from April through September 2025 remains under legal hold regardless of what anyone clicked delete on.⁸⁷ The lesson isn't about that specific lawsuit. It's that a company's legal exposure, not your privacy settings, can decide how long your SSN and income history sit on someone else's servers. For a broader look at how AI tools reach into your files and chats on demand, see what AI connectors actually do with your data.

The software ownership problem: you don't control the infrastructure holding your SSN

This is the part consumers skip past. When you rent your software, you inherit that vendor's fate. An outage, a breach, a subpoena, a policy change, a lawsuit, it all lands on whatever data you handed over, and you have no vote in any of it.

A tax return is not a spreadsheet. It's your SSN, your income, your dependents, your bank routing numbers, sometimes your address history. Handing that to a general-purpose chat product is fundamentally a rented-infrastructure decision: you're trusting a third party's retention policy, security posture, and litigation exposure with a document that can be used to steal your identity. It's worth asking, before you upload anything sensitive, whether the convenience is worth the cost of not knowing exactly where that copy lives or who can subpoena it.

The professional's version of this problem

Here's what makes the consumer's casualness stand out: licensed tax preparers are legally barred from doing what you're doing to your own return.

Under IRC Section 7216, a preparer who knowingly or recklessly discloses or uses a client's tax return information for anything other than preparing that return can face a fine of up to $1,000, up to a year in prison, or both.⁹ The IRS's last formal guidance on this rule predates the AI era by more than a decade, dating to 2013, leaving real ambiguity about whether pasting client data into ChatGPT counts as a violation.¹⁰ The AICPA's response has been to tell preparers to be cautious anyway and get signed client disclosure before using AI on their tax data, even as 60% of tax professionals now use AI for tax research weekly, up from 33% a year earlier.¹⁰

Preparers also carry a second, separate obligation. The FTC Safeguards Rule, under the Gramm-Leach-Bliley Act, classifies tax preparers as financial institutions and legally requires them to maintain a written information security program to protect client data.¹¹ There is no consumer equivalent. No regulator requires you to have a security plan before you upload your own W-2 to a chatbot. The professional bar exists precisely because this data is dangerous in the wrong hands. Consumers operate under no such bar at all.

Real leak and scam vectors to know about

This isn't theoretical. In mid-2025, thousands of ChatGPT conversations shared via the "Share" link feature, some containing sensitive personal and financial details, were indexed by Google search before OpenAI removed the discoverability toggle.¹² Users who thought they were sending a private link to a friend or accountant had, in some cases, made that conversation publicly searchable.

The IRS's 2026 "Dirty Dozen" scam list flags a related and growing threat: AI-powered tax scams, including voice-cloned IRS impersonation calls and fake AI-generated IRS portals designed to harvest taxpayer data during filing season.¹³ That's a different attack surface than uploading to a legitimate AI product, but it points to the same trend: the more comfortable people get feeding financial documents into AI interfaces, the more attractive that habit becomes to exploit.

How to use AI for deduction-hunting without the exposure

AI can still be useful here, just not as the last stop before you file. A few guardrails:

1. Redact identifiers before you upload. Black out your SSN, account numbers, and address on any document before it goes into a chat window. The deduction categories don't need your identity attached.
2. Use a paid tier with retention controls, not the free consumer product. Business and Enterprise tiers don't train on your data by default and let you control how long it's kept.⁶
3. Treat the output as a checklist, not a filing. Use ChatGPT to brainstorm deduction categories you might have missed, then verify each one against IRS guidance or validated tax software.
4. Get a second opinion from a human or a dedicated tool. Even purpose-built tax AI needed human review on a quarter of returns.³ A general chatbot's answer should never be the final word on your own.
5. Never share a chat link containing financial details. The Google-indexing incident shows "private" sharing isn't always private.¹²

The bottom line

About 1 in 5 Americans say they plan to use AI to help file their 2026 taxes, and nearly half say they trust AI to give accurate tax guidance.¹⁴ The accuracy numbers don't support that level of trust yet, and the data-handling numbers should concern anyone who assumes "private chat" means private.³⁴⁷¹² Convenience now, exposure later, is the actual trade you're making. A licensed preparer wouldn't be allowed to make that trade casually with your data. You're making it every time you drag a PDF into a chat window and hit enter.