The author is one of many taxpayers who prepare their own returns. For years he had a rather simple return, because his income was reflected on a W-2, and he had nothing else happening in his life that would have created any complexity on the return. But in 2025 he purchases some employer stock through an employee stock purchase plan, and then sold most of the stock to generate cash for his wedding expenses.
Knowing that the tax rules for the sale of employer stock aren't very simple, he turned to ChatGPT. He reports that his first question brought an answer as to how the sales are treated for tax purposes, broken down into "digestible bullet points." He uploaded his Form 1099 from the brokerage firm that processed his stock sales. ChatGPT then told him that he needed to use a number different from what the brokerage reported, though it isn't clear from the article whether that number was gross sales, net sales, basis, or something else. He was instructed to examine his "last few W-2s to see that they included a certain line item." He was also told by ChatGPT, "“This is a very simple return with one stock plan wrinkle. You do NOT need a CPA.”
Fortunately, at that point he contacted a CPA that he knew. The CPA, after being brought up to speed, concluded that the author "had gotten possibly correct but also incomplete information." The analysis of the W-2 forms was "quite important" in determining if the brokerage indeed was using the correct numbers. It turned out that, according to the CPA, some of the numbers in the Form 1099 seemed to indicate that the author had engaged in transactions that he "may not have made." The CPA suggested that ChatGPT has not provided any information about this issue probably because the author had not asked. A high quality tax professional would know to provide information that is relevant even if the client fails to ask, because most clients don't know that they need to ask or what to ask.
The author concludes that the advice he obtained from ChatGPT "seemed so sound and was so breezily delivered that [he] was ready to file and risk having made a mistake." The author spoke with an accounting professor who shared, "AI will convince you that the sky is green. It is so convincing. The professor cited a time when a chatbot incorrectly answered one of the tax questions he gives his students. He continued, “It gave me this response that the mechanics were perfect, but I had to take a step back and say, ‘Well, you’re wrong.’”
An AI strategy company founder informed the author, “But by default, large language models are trained to be helpful assistants. Oftentimes you’re going to run into hallucinations.” Hallucinations is a fancy word for wrong answers.
The author's conclusion? He sums up advice from experts: "[Y]ou’d be wise to tread very carefully before using AI to help with your taxes." He quotes the AI strategy company founder, “If you make a mistake while using AI to do your taxes, it could get you in trouble with the IRS. And a valid excuse isn’t, ‘The AI made me do it.’” So true.
The author then shares what "pros say to keep in mind" when using AI as a tool to prepare tax returns. I'll let you read the article because my advice would simply be, "Don't." My experience with "AI" is that it's not ready for prime time. I asked one AI site "who is James Edward Maule?" and the answer was so amazingly incorrect I wondered if the "I" in "AI" should be changed from intelligence to ignorance. Why is AI still insufficient? AI engines use data scraped from the internet. It doesn't do a good job distinguishing between factually correct and factually incorrect data. It doesn't do a good job distinguishing between, for example, the text of a Code section published on the internet in 2015 that is still good law and the text of a Code section published on the internet in 2015 that is no longer good law because it has been repealed, amended, or otherwise set aside. It doesn't do a good job deciding whether a word in one clump of data should be matched with the same word in another clump of data. It reminds me of first-year law students who focus on the words of a case at face value rather than considering the context in which those words are being used.
Perhaps someday "AI" will function with the critical thinking skills, wisdom, judgment, experience, and perspective that expert humans bring to the table. But that day isn't today. That day may never arrive. In the meantime, "AI" is probably useful with simple tasks that computers do well, such as computation or mere data compilation without analysis or judgment. But "AI" surely isn't ready to answer questions that can be answered properly only with the skills that "AI" does not possess.
