Can ChatGPT do my taxes?

The experience is genuinely impressive, which is the problem. You paste in a W-2, describe your year, and get back a clean, organised, plausible answer with the right vocabulary in the right places. Nothing about that output signals which parts are solid and which parts are a guess.

Why the answer reads so well

A language model produces the most likely continuation of your prompt. On a tax question, the most likely continuation is confident, structured, and written in the register of a professional, because that is what tax writing looks like in the material it learned from.

Fluency and accuracy are produced by the same process, so they arrive together and feel identical. A human expert hedges when uncertain, and you can hear it. A model's uncertainty is invisible: the wrong answer is delivered in exactly the tone of the right one.

Two independent measurements exist, and they agree.

Antinozzi and Cooper put common tax questions to GPT-3.5 and GPT-4 through the 2023 and 2024 filing seasons and scored the answers. Their finding, published in the Journal of Emerging Technologies in Accounting in April 2025: the overall share of correct responses ranged from 39 to 47 percent. Upgrading from 3.5 to 4 did not move it. Their conclusion was that ChatGPT is generally not a reliable source of tax guidance for uninformed taxpayers.

Accuracy fell further on exactly the questions people actually ask: the more common ones, the ones with complex answers, the ones needing an evaluation of the taxpayer's own fact pattern, and the ones about law settled after the model's knowledge cutoff. That is a description of an ordinary return.

The Taxpayer Advocate Service, the IRS's internal watchdog, reached a similar place from a different direction. Citing an informal review of sixteen complex tax questions, it reported that two leading tax preparation companies' chatbots gave inaccurate or irrelevant responses up to 50 percent of the time. Those were purpose-built tax chatbots, not general assistants.

Four failures specific to chatbots

The training cutoff. A model knows the tax law that existed when it was trained. Brackets, standard deductions, contribution limits, mileage rates and phase-outs move every January, and Congress moves the rules underneath them irregularly. A model with a cutoff before a change will answer with the old rule, in the same confident voice, and will not tell you that is what happened.

It answers the question you asked. Ask how to deduct a home office and it will explain the home office deduction. It will not ask whether you are an employee rather than self-employed, which is the fact that decides the answer. A preparer's value is often in the question you did not know to ask.

Invented citations. Models generate references that follow the shape of real ones. A plausible section number, a plausible publication title, a plausible page. Every citation from an AI needs checking against IRS.gov before you rely on it, and the ones that look most authoritative are not more trustworthy than the rest.

It agrees with you. Push back on an answer and the model will often revise toward whatever you seemed to want. That is a useful conversational habit and a dangerous research method. If you want the deduction badly enough and ask three times, you will usually get it.

The IRS put this in writing

Artificial intelligence appeared on the IRS Dirty Dozen list of tax scams for the first time in 2026. The warning: taxpayers should not rely on AI-generated responses to complex tax questions, and should verify any calculations or information provided by artificial intelligence.

The agency is not claiming AI is fraudulent. It is saying the output is unverified, and that an unverified figure on a signed return is your problem. The accuracy-related penalty is 20 percent of the underpayment, and no version of "the model said so" has been recognised as reasonable cause.

A better division of labour

If your return is a W-2, a standard deduction and nothing else, IRS Free File does it accurately for nothing, which is both cheaper and safer than a chatbot.

If your return has a business in it, property, equity compensation, a move between states, a divorce, a death, or a year you did not file, the chatbot will still answer. It will answer instantly and in full. That is precisely when its confidence is worth the least.