What AI Gets Wrong About Your Taxes
Artificial intelligence is changing how people approach tax prep. We get some version of the same question almost every week. Can ChatGPT do my taxes. Should I be using AI for bookkeeping. Is my CPA going to be replaced by a chatbot. Fair questions. Over the past year we have actually run these tools against real returns inside our firm. Not as a marketing exercise but as a stress test. Here is what we have seen.
Where AI is genuinely helpful
When AI is pointed at a clean return with good source documents, it can be useful. It gives you a second pass. We have seen it catch balance sheet items that did not tie out, prior year fixed asset disposals that were not reflected on the current year return, and inconsistencies between the workpaper write up and the actual numbers. Those are real catches. The kind of thing a tired human might miss late at night in the middle of filing season. It is also good at summarizing long IRS notices, drafting client emails, and answering basic conceptual questions. That part is real. We are not anti AI.
Where AI fails
This is the part that does not show up in the marketing. It hallucinates current law. On a recent return, an AI tool flagged a $31,500 MFJ standard deduction as incorrect and insisted the correct 2025 number was $30,000. The return was right. The AI was using outdated information. This is one of the most dangerous failure modes. Confident, specific, and wrong. If we had accepted the flag, we would have made the return worse trying to fix something that was not broken. It misreads what it is looking at. On another return, the AI flagged that SDI tax was not flowing through to Schedule A. It was not SDI. It was Medicare tax, which does not flow through to Schedule A and never has. The issue was not the return. It was the model not understanding the line item. It flags timing it does not understand. We saw a return where the AI flagged a missing Form 8606 tied to a backdoor Roth conversion that had not happened yet. It belongs on next year’s return. The AI could not follow the timeline. It generates noise that looks like analysis. A typical AI review of a correct return produces a long list of flags that sound technical but do not hold up. You will see something like Line 15 may incorrectly show 4,074 when it should be 0, followed a few sentences later by the model confirming the number is in fact 0. It looks thorough. It adds nothing. Filtering through that takes longer than just reviewing the return properly. It cannot make judgment calls. Should you take a more aggressive position or stay conservative. Is something defensible or likely to draw attention. Should you elect into PTE this year or wait. Those are not lookup questions. They require experience and accountability. AI does not have either.
What this means for you
If you are using AI to understand a concept, draft questions, or organize your records, that makes sense. If you are using it to prepare a return or take a position, you are relying on a tool that does not reliably know which year’s tax law it is applying. We sign every return we prepare. That signature means the work has been reviewed by people who know when something is off and understand why. AI does not sign returns. That matters.
Our take
AI is a force multiplier for trained professionals. It is not a replacement.
The model that works is simple. Humans make the calls. AI supports the parts that do not require judgment. Anyone selling a fully autonomous AI tax solution is selling the version that exists in a demo, not what shows up in real returns. If you have a question about a return, something an AI tool told you, or a situation that feels unclear, reach out. It is easier to get it right now than fix it later.