When to (not) trust an AI tax research answer
AI and the CPA
Carol Lin Vieira | August/September 2026 Footnote
Editor's note: Updated July 27, 2026
If you’ve started using AI for tax research, you’ve probably had this experience: You ask a question, you get a confident, well-structured answer and something feels slightly off. You can’t immediately say why, you just know you wouldn’t sign your name to it.
That instinct is worth trusting and it’s worth knowing what to look for.
Foundation models — the general-purpose AI behind tools like ChatGPT and Copilot — are built to produce fluent, authoritative-sounding text. The problem is that fluency and accuracy are not the same thing and, in tax research, the difference is important.
Here are four signs that your AI output might not be trustworthy.
- No cited authority. If a tool gives you a conclusion without pointing to a specific code section, regulation or ruling, you have no way to verify it. Ask: Which code section is that based on? The quality of that answer might help you decide whether the first one was worth keeping.
- A jurisdiction-specific question gets a general answer. Tax law varies materially by jurisdiction. If your tool returns a single paragraph covering California and Texas software sales tax without separating the analysis, it has collapsed distinct issues into one. This is a reliability problem, not a formatting issue.
- A multi-issue fact pattern compressed into one answer. A rigorous research process identifies and sequences each issue separately. One paragraph that addresses everything at once is summarizing, not reasoning.
- No visible reasoning trail. You should be able to see why the tool reached its conclusion. If you can’t audit it, you can’t defend it.
Try this question today: “A business sells SaaS and downloadable software to customers in California and Texas and has one remote employee in each state. What sales tax and income/franchise tax issues arise?” How does response quality vary across your research tools?
