The short answer: AI assistants can help a lot with your taxes, but you shouldn't let one calculate your return on its own, and none of them can file it for you. In the latest public test, the best model got about one return in five wrong even with web search, and more than half wrong without it.
Here's what the tests found, why AI gets tax returns wrong, and how to use an assistant without putting your refund at risk.
How accurate is AI at tax math?
The best public test is TaxCalcBench, run by the tax software company Column Tax. It gives AI models realistic tax situations, with the same information you'd type into tax software, and checks whether every line of the return comes out exactly right.
The latest version, released in June 2026, uses 50 returns for the 2025 tax year, most with a state return too. The results:
| Returns exactly right | |
|---|---|
| Best model, no web search | 44% |
| Best model with web search | about 78% (38 of 49) |
| GPT-6 Sol with web search | 72% |
| Claude Opus 5.5 with web search | 70% |
That's a big jump from the first version in 2025, when the best model got 32% of simpler returns right. Progress is real and fast.
But a return that's mostly right is still wrong. A test that counts "exactly right" is the one that matters, because the IRS checks the math. A mistake can mean a smaller refund, a bill with interest, or a letter months later.
Why AI gets tax returns wrong
When models get returns wrong, the original TaxCalcBench paper found the same kinds of mistakes again and again.
Using the wrong tax table or rate
Tax brackets, the standard deduction and dozens of limits change every year. A model trained on older material can confidently use last year's numbers. Web search helps, which is why the scores jump when it's turned on, but it doesn't fix everything.
Arithmetic and rounding
A return is a long chain of calculations, and each line feeds the next. One small slip early on throws off every line after it. IRS forms also have specific rounding rules, like "round down to the nearest $1,000" or "multiply by 6%", that models skip or apply in the wrong place.
Getting eligibility wrong
Who counts as a qualifying child, which filing status you can use, whether you qualify for the earned income credit: these follow detailed tests with exceptions. Models often apply a simplified version of the rule, and simplified is wrong for some people.
Missing an interaction
Taxes are full of cross-effects. Your adjusted gross income can shrink a deduction, which changes your taxable income, which changes a credit. Models handle each piece well and lose track of how they connect.
Can an AI assistant file your return?
No. Only companies the IRS has accepted into its e-file program can e-file returns. A chatbot can help you prepare, but the return has to go through tax software or a tax preparer.
The big assistants work with tax companies instead. TurboTax has apps inside ChatGPT and Claude, and you finish the return in TurboTax. OpenAI has said plainly that ChatGPT is not a tax preparer.
What AI is good for
- Explaining. What's a 1099-K? Do I need to report this? Why is my refund smaller this year?
- Organizing. Making a list of the forms you should expect, based on your jobs, accounts and life changes.
- Reading forms. Pulling the numbers off a W-2 or 1099 so you don't have to type them.
- Comparing options. Walking through the trade-offs of filing jointly or separately, or itemizing or taking the standard deduction, before the software confirms the numbers.
- Checking. Asking "does this look right?" about a return a tax program prepared.
What to leave to tax software
- The math. Tax software follows the IRS's rules the same way every time, and shows its work line by line.
- Deciding what you qualify for. Credits like the earned income credit have detailed rules that models often get wrong.
- Filing. Only an IRS e-file provider can submit your return.
A simple test you can run yourself
Ask your assistant the same tax question twice, in two new chats, with slightly different wording. For example:
I'm single, I earned $58,000 in wages in 2026, and I have no other income. What's my federal income tax?
If the two answers differ, that tells you something important: the model is estimating, not calculating. Tax software gives you the same answer every time for the same facts.
How to use AI safely at tax time
- Use it to understand, not to decide. Ask why, not just how much.
- Ask for sources. A good answer points to an IRS form, publication or page. Open it and check.
- Watch for out-of-date numbers. If it doesn't know about 2026 changes like the new deductions for tips and overtime, it's working from old information.
- Keep sensitive numbers out. Don't paste in your Social Security number or bank details. More on safety.
- Let software do the final return. Then use AI as a second reader.
How Navo uses AI
Navo is built on this split. A tested tax engine does every calculation, and AI never decides your numbers. When a return includes something Navo doesn't handle yet, it tells you plainly instead of guessing. You review and sign your return on Navo, and Navo e-files it. Read why Navo never lets AI do the math.
Frequently asked questions
Will AI be good enough to do taxes on its own soon? Scores are rising fast, so maybe. But "right most of the time" isn't the bar for a tax return. Until a model is right every time for every situation it accepts, keep the math in software.
Is it OK to use AI to check a return a preparer did? Yes, as a way to ask good questions. If it flags something, ask your preparer rather than changing the return yourself.
Do the AI tax apps inside chatbots do the math? The apps from tax companies hand off to that company's software, which does the calculation. That's different from the chatbot answering on its own.
Related guides
- How to file your taxes free with AI
- Is it safe to give an AI your tax documents?
- Why Navo tells you when it can't do your return
This post is general information, not tax advice for your situation. Sources: TaxCalcBench leaderboard; TaxCalcBench paper (arXiv:2507.16126); IRS Publication 1345.