If ChatGPT, Gemini, or Claude gets your tax return wrong, the IRS holds you liable, not the AI vendor. You sign the return under penalty of perjury, so a chatbot's mistake becomes your underpayment, your interest, and potentially your 20% accuracy-related penalty.
There is no AI safe harbor in the tax code. A general chatbot is not a licensed preparer, and "the AI told me to" is not the kind of professional reliance that abates a penalty. Below is exactly what you owe if the math goes wrong, when the 20% penalty applies, the stricter 5% line for gig workers, and how to use AI for taxes without making it the filer of record.
General information, not tax advice. This article explains how liability and penalties generally work. It is not legal or tax advice for your situation. Consult a licensed CPA, enrolled agent, or tax attorney before you file.
Who is liable if AI botches your return?
You are, completely. The chatbot signs nothing. When you sign a Form 1040, you sign a declaration under penalty of perjury that the return is true, correct, and complete to the best of your knowledge. The IRS assesses the tax against the name on the return. Not the software. Not the chatbot. Julie Siegel, a former Treasury deputy chief of staff, put it plainly to CBS News: if an AI model misinterprets an IRS form and gets it wrong, "you are the one" holding the bag.
That liability is not new. It is the same standard that applies when a human tax preparer makes an error: the preparer may face their own separate penalties, but the tax, interest, and taxpayer-side penalties still land on you. A public chatbot has no preparer registration, no PTIN, and no professional duty to you, so it carries none of that responsibility. Caroline Bruckner, a tax professor at American University, told CBS News that "AI on its own is not capable of preparing an accurate tax return," partly because these tools can surface deductions from prior years that no longer apply.
What does the IRS charge you for a wrong return?
Get it wrong and you can owe three separate things: the tax you should have paid, interest on that amount, and an accuracy-related penalty. The penalty under Internal Revenue Code section 6662 is 20% of the underpayment attributable to the error. That is on top of paying the tax itself, and interest runs until the balance is cleared.
The 20% penalty is not automatic on every small mistake. Two of the most common triggers for individual filers are negligence and a substantial understatement of income tax. Negligence means you did not make a reasonable attempt to follow the tax laws, such as failing to report income shown on a Form 1099 or claiming a deduction that looks too good to check. A substantial understatement is a size test measured in dollars and percentages, not intent.
When is an understatement "substantial"?
For an individual, IRC section 6662(d) makes an understatement substantial when it exceeds the greater of 10% of the tax required to be shown on the return, or $5,000. That threshold still stands as of July 2026. Here is the line for your own return. If your correct tax was $40,000 and you reported far less, the 10% line sits at $4,000, and the $5,000 floor governs. If your correct tax was $80,000, the 10% line ($8,000) governs instead. Cross that threshold and the 20% penalty can attach to the understated portion.
Here is what most AI-tax coverage skips: if you are a freelancer or gig worker, one rule hits you harder. Claim the Section 199A qualified business income deduction and the substantial-understatement threshold drops from 10% to 5% of the tax required to be shown. Half the buffer, same 20% penalty. AI tools routinely misapply the QBI deduction because the rules shift with income phase-outs and business type. That makes it a real risk for the exact gig workers and freelancers most likely to lean on a chatbot.
Does "the AI told me to do it" get you off the hook?
No. The tax code does offer relief through the reasonable cause and good faith standard, and the IRS can remove or reduce a penalty if you acted in good faith and can show a genuine reason you could not meet your obligation. But relying on a general AI chatbot is very unlikely to clear that bar. The classic form of reasonable cause is reliance on professional advice, and a public chatbot is not a professional advisor. Rely on a qualified CPA and their error can support reasonable cause. Rely on a chatbot and it cannot.
Reliance on a tax professional generally has to meet real conditions: the advisor must be competent and qualified, you must give them all the relevant facts, and you must actually and reasonably rely in good faith on their judgment. A chatbot that confidently invents a citation, mixes up tax years, or cannot see your full situation does not satisfy that framework. There is no separate "AI safe harbor" written into the code, so the burden stays on you to show your position was correct or that your reliance was reasonable.
If you use AI to explore a tax question, treat its answer as a lead to verify, never as authority. Confirm every figure against the actual IRS form instructions, a current IRS publication, or a licensed preparer before it touches your return.
AI chatbot vs. licensed preparer: who carries the risk?
| Dimension | General AI chatbot | Licensed preparer (CPA / EA) |
|---|---|---|
| Who signs and is liable for the return | You alone; the vendor assumes no tax liability | You sign, but the preparer signs too and has professional duties |
| Preparer registration (PTIN) | None; not a registered tax return preparer | Yes; regulated, with a PTIN and standards of conduct |
| Counts as reasonable-cause reliance | Very unlikely; not professional advice | Can support reasonable cause if you gave full, accurate facts |
| Sees your complete, verified situation | Only what you paste; may hallucinate or use stale rules | Reviews your full documents and current-year law |
| Accountable if it is wrong | No recourse against the tool for your tax | May carry professional liability and can help fix errors |
The table is not an argument that AI is useless for taxes. It is a reminder of where the risk sits. A chatbot can be a fast study aid for understanding a concept. It cannot be the party that stands behind your numbers, because nothing about a public model makes it responsible for what you file.
Will AI get you audited? What the IRS actually uses it for
Using ChatGPT to understand a deduction does not flag your return. The IRS has no way to know which tool helped you. The IRS does use machine learning to help select some audits, but the flagship program aims at the largest and wealthiest taxpayers, not typical 1040 filers. Its Large Partnership Compliance effort applies machine learning to spot where large partnerships may be shielding income.
When the IRS announced the expanded effort, then-Commissioner Danny Werfel described it as focusing new capacity on 75 large partnerships averaging more than $10 billion in assets and roughly 1,600 millionaires who each owed at least $250,000. He added that "middle- and low-income filers will continue to see no change" from the historically low audit rates of prior years. So the audit worry for an ordinary DIY filer is not an AI dragnet. It is a wrong number that trips a routine mismatch or math check.
The other risk: what you paste into the chatbot
Accuracy is one exposure. Privacy is the other. Pasting a full W-2, a 1099, or your Social Security number into a public chatbot puts identity-level data into a system you do not control. Andy Sambandam, founder and CEO of a privacy technology firm, told WJLA that people should not upload a tax return or an entire bank statement to AI, and should instead treat these tools as a "guiding tool" for general questions.
Once financial documents go into an AI system, you cannot reliably pull that data back, and skillful prompting could surface pieces of it. A Social Security number cannot be reissued the way a password can be reset, and tax data is exactly what refund-fraud and identity-theft schemes want. Keep the identifying documents out of any general chatbot, whether you are asking it a tax question or anything else.
How to use AI for taxes without becoming the cautionary tale
Use AI as a study aid, not the party that signs the return. It can explain what a term means, walk you through a concept, or help you build a checklist of questions to ask a professional. It should never be the last word on a dollar figure that lands on your return, and it should never hold your raw identifying documents.
- Ask AI to explain, not to file: use it to understand a rule, then verify against the current IRS form instructions or a licensed preparer.
- Never paste identifiers: keep your SSN, full W-2, 1099s, and bank details out of any general chatbot.
- Check the tax year: AI often cites deductions or limits from an old year, so confirm the number applies to the year you are filing.
- Watch the QBI trap: if you claim the Section 199A deduction, the substantial-understatement threshold is stricter (5%), so double-check that math.
- Keep your work: save the IRS source you relied on for each position, which supports a good-faith record far better than a chat transcript.
- Escalate anything complex: for business income, capital gains, rental property, or multi-state issues, get a CPA or enrolled agent to sign off.
Where a private memory layer fits
The safe way to get AI help with your own documents is to keep those documents in a space you own, rather than pasting them into a public chatbot. MemX is an external, user-owned memory layer built to be private by architecture: per-user isolation and encryption at rest. You can organize and privately search your own tax paperwork, W-2s, 1099s, receipts, and prior returns, instead of dumping them into a general model's chat window.
MemX does not give tax advice and does not file your return. It helps you keep your records organized and searchable on your own terms, so when you do sit down with a preparer or software, your documents are ready. The liability for the return is still yours, which is exactly why controlling where your data lives matters.
The bottom line
AI can help you learn about taxes. It cannot take responsibility for them. If a chatbot botches your return, the IRS assesses the tax, interest, and a possible 20% accuracy-related penalty against you, and "the AI told me to" will not clear the reasonable-cause bar. Use AI to understand, verify everything against IRS sources or a licensed professional, and keep your identifying documents out of any public model. "The AI told me to" is not a defense the IRS accepts.
01If ChatGPT makes a mistake on my taxes, who is liable to the IRS?
You are. You sign your return under penalty of perjury, and the IRS assesses tax, interest, and any penalty against you, not the AI vendor. A public chatbot is not a registered preparer and carries no tax liability for your return.
02What is the penalty if my tax return is wrong?
The accuracy-related penalty under IRC 6662 is 20% of the underpayment tied to the error, on top of the tax and interest you owe. Common triggers are negligence and a substantial understatement of income tax.
03What counts as a substantial understatement for an individual?
For most individuals, it is when your understatement exceeds the greater of 10% of the tax required to be shown, or $5,000. If you claim the Section 199A QBI deduction, the threshold tightens to 5%.
04Can I avoid the penalty by saying I relied on AI?
Very unlikely. Reasonable cause usually means good-faith reliance on a qualified professional, and a general chatbot is not one. There is no AI safe harbor in the tax code, so the burden to justify your position stays with you.
05Is it safe to upload my W-2 or SSN to ChatGPT?
No. Security experts warn against uploading tax returns or full financial documents to general AI tools, because that data is hard to remove and can feed identity theft or refund fraud. Keep identifiers out of public chatbots.
