Expert: Sam Taube, Lead writer, investing and taxes team, NerdWallet The NerdWallet Data Studies team tested three AI chatbots -- ChatGPT, Gemini and Perplexity -- by asking each of them the same seven tax questions, using three testers for 63 graded transcripts and more than 50,000 words of chat. The tools nearly aced the three questions taken from the IRS practice quiz for enrolled agents, and general guidance such as how to approach a first filing was broadly in line with NerdWallet's own writers. The failures clustered where tax questions stop being trivia and start depending on a person's facts. Errors in the transcripts, all presented in the same confident register: 1. An incorrect standard deduction amount, off by nearly $3,000. 2. A claim that an electric vehicle credit was still available for a fictional new Tesla -- it is not. 3. A claim that the tester's $10,000 in overtime pay is not deductible -- it is. 4. Repeated instructions to file taxes in the wrong state, with the chatbots leaning on federal rules and underweighting state-specific information. 5. Inconsistency on open-ended questions: asked to estimate a tax bill, one service replied '~$32,267' while another said 'tens of thousands'; asked for a tax software recommendation from the same prompt and filing profile, one chatbot ranked FreeTaxUSA first for one tester and TurboTax first for another. 6. Assumptions the testers explicitly told the chatbots not to make -- every chat opened with an instruction to forget previous chats and assume nothing about their background. One tool pulled in biographical details from earlier questions, and another apparently used the tester's VPN location when telling them where to file; one mixed up the state where a tester said they went to college with the state where they said they worked a summer job. Sam Taube, lead writer for NerdWallet's investing and taxes team: 'A couple of years ago, even the cutting-edge AI models couldn't reliably do basic arithmetic. Recent updates have improved their math skills a lot, but there's still a risk of errors. And other accuracy problems -- like the tendency to cite nonexistent, hallucinated cases in response to legal questions -- still come up in 2026.' Taube adds: 'Taxes involve both of those subjects -- math and law. Even though AI has gotten better at the former, and may get better at the latter in the future, it's not a reliable source of truth yet.' The confidence is the defect, not a side effect. NerdWallet's write-up puts it plainly: 'AI chatbots seem stuck on authoritative cruise control. As a result, wrong answers can sound right.' That matters because the wrong answers here are not exotic hallucinations -- they are wrong numbers, a credit that has expired, a missed deduction and the wrong state, delivered in the same calm tone as the correct answers around them. The team's own guidance is to use the chatbot as a starting point and verify against the IRS or a tax professional before acting; Taube's example is advising a prior-year IRA contribution (sound) alongside harvesting investment losses before year-end when only losses on investments already sold can be deducted (faulty). Verdict: incorrect. A tool that nearly aces a professional quiz can still hand a taxpayer a standard deduction figure that is nearly $3,000 off, an expired credit as still available and a filing state that is simply wrong -- and it does so without flagging any uncertainty. Source: https://www.nerdwallet.com/taxes/studies/doing-taxes-with-ai