I have enjoyed a relatively easy tax environment for most of my adult life. Most years, I simply verified that my income from my W-2 was correct and clicked through the questions in my preferred tax software. I have no dependents, no side job income, and no assets in my name.
This past year has been a little different. After years of purchasing stock through my company’s employee stock purchase plan, I sold the majority of my stock to raise funds for my upcoming wedding.
There are some relatively tricky rules for selling these stocks, but the gist is that these plans allow employees to buy stock at a discount to the actual stock price. So, determining how much you made (in which case you owe capital gains tax) or how much you lost on the sale of your shares requires some calculations.
So I did what approximately 1 in 5 taxpayers are doing these days, according to recent research from IPX 1031. I asked the AI for help.
I did so with some skepticism. I’ve seen enough stories about AI “hallucinations” (the industry term for chatbots causing problems) that I was half-expecting ChatGPT to mess with my taxes. Moreover, it had only been three years since we tested AI for tax strategy and saw it struggle. It’s also worth noting that OpenAI’s usage policy warns against using its products to automate “high-stakes decisions in sensitive areas without human review.”
Still, once I started chatting using the latest version of OpenAI’s large-scale language model, I felt my hesitation melt away. Not only did he answer my initial question about how ESPP sales were taxed, he broke it down into easy to understand bullet points and asked if I would mind sharing more information.
Since I was using the corporate version of the software, which doesn’t use data to train OpenAI’s models, I uploaded an integrated 1099 form from my brokerage firm.
“This is great.[Your brokerage]actually provided us with everything we needed,” Bott told me. “What’s going on?”
What ChatGPT told me is essentially: Your brokerage uses one number, which is uploaded to your tax software. However, you actually need to use a different number. I had to check some recent W-2s to make sure they had certain items included.
I was ready to press “file”. But first, I turned the conversation to Miklós Ringbauer, a certified public accountant who had previously helped me with tax matters.
He said the information I got was probably correct, but incomplete. He told me that the W-2 check that ChatGPT had assured me wasn’t that big of a deal was actually very important. To make sure it all adds up, either I or a tax professional would have to look at the numbers and compare them to my brokerage’s numbers.
Additionally, some numbers on my 1099 appeared to indicate that I had taken certain taxable actions that I may not have taken, Ringbauer said. ChatGPT didn’t say anything about them. Probably because I didn’t think to ask.
“The question is: Do taxpayers understand what needs to be addressed and do they have the necessary understanding of the documentation to refer to in order to fix it?” Ringbauer said. “My understanding is that many of our customers don’t.”
Overall, the mistakes I encountered were minor. ChatGPT told me that something was “almost certainly” correct when I actually had to do my homework. But the point is that the advice I received seemed so sound and so refreshingly conveyed that I was ready to risk making a mistake and apply.
For people who aren’t tax law experts, that can be a big problem, says JT Egan, a clinical assistant professor of accounting at Purdue University.
“The AI will convince you that the sky is green. That’s very convincing,” he said, citing the time the chatbot incorrectly answered one of the questions Egan gave his students about taxes. “The answer was that the mechanics were perfect, but I had to step back and say, ‘Well, you’re wrong.’
The AI chatbot isn’t trying to trick you. In fact, in recent years, many of the leading models have gotten better at notifying users that they don’t have a complete answer to a particular prompt or are wandering into a gray area, said Jordan Wilson, founder of AI strategy firm Everyday AI.
In fact, ChatGPT started my conversation by saying, “If something is very specific or high-stakes, we’ll flag where we need a CPA’s opinion.”
“But by default, large language models are trained to be helpful assistants,” Wilson says. This means they often sound very confident and “often have hallucinations.”
And we can deliver them with incredible confidence. “This is a very simple return with just one stock plan,” ChatGPT told me. “You don’t need a certified public accountant.”
With that in mind, experts say it’s wise to tread very carefully before using AI to help with your taxes.
“If you make a mistake when using AI to file your taxes, you could get into trouble with the IRS,” Wilson says. “And ‘the AI made it happen’ is not a valid excuse.”
If you plan on using AI as a tool during tax season, here’s what experts say you should keep in mind.
Check your model
No two LLMs are exactly the same, and even if you have a favorite AI company, they typically offer different models with different capabilities, Wilson says.
For complex subjects like taxes, he says, you’ll want to use a “thinking” model. This is a model that breaks down difficult topics by generating step-by-step solutions and shows you the results.
Even if you’re not using the free version of a particular chatbot, “many of these companies are going to stick with a faster model that’s not as smart by default,” Wilson says.
AI companies like Anthropic and OpenAI have tiered pricing, giving you access to different powerful versions of their AI chatbots. Even at higher tiers, the model you’re using may not have been tuned for complex topics to begin with. When I first opened the paid version of my company’s ChatGPT, it defaulted to “Instant” mode for “daily chats.” I clicked on the menu at the top of the chat to switch to “thinking” mode.
Another factor worth checking before uploading your document is whether the AI company is using your information to train its models. Wilson said users of the paid version can usually turn off model training, but users of the free version should think twice before sharing personal information.
Your information will ultimately not be made public, he says. “But if you’re on a free plan, companies can use any data to train their models. And that’s probably not what you want, especially when it comes to sensitive information like bank account information or social security numbers.”
Understand the limits of AI and your limits
Even if AI follows the right process, it can arrive at the wrong answer, experts say. One reason, Wilson says, is that the models may rely on outdated data. The data used to train the model is sometimes months old, he says, and can leave up-to-date tax information missing. This may require the model to obtain the latest information from the Internet.
“Where do we verify that information? Do we find articles from 2024, 2025, or 2026?” Wilson says. “The difference obviously can greatly sway the results and accuracy of anything that uses it.”
In some cases, AI may have the perfect answer to your tax question, but that doesn’t mean it’s the right answer for you. That’s because certain tax rules may or may not apply, given the specific nuances of your situation.
Using a simplified example, consider a taxpayer who wonders if he can deduct his dog. Ask any CPA and they’ll tell you, “It depends,” says Egan. Although you can’t write off your pet, under certain circumstances you can deduct the cost of a guide dog as a medical expense.
When you ask an AI chatbot a question, “it might start with ‘yes,'” Egan says. “The question is, how many people stop reading after they get the answer they want?”
Extrapolating across the entire tax code, it is not difficult to see how users would struggle to obtain accurate returns by relying on AI for answers. Even if an AI answers a question correctly or performs an accurate calculation, it is limited by what it knows about it, Wilson says. Without a holistic approach, he says, these models are set up to fail.
None of this means that the model automatically fails when faced with tax issues. You should ask yourself whether you are prepared to provide the model with all the relevant information needed and ask for all relevant follow-up. If the answer is no, and there’s a lot of money at stake, it’s wise to consult a professional, Wilson and other experts say.
“I think this is one of the biggest mistakes that most people make, even if they’re pretty used to using AI, is thinking, ‘Well, if I input some part of the document and give it the right context, it’s going to work in the output,”’ Wilson says. “But in reality, humans need to do very hands-on work to ensure that the models have the correct context.”
Want to lead with confidence and bring out the best in your team? Take CNBC’s new online course, How to Become an Exceptional Leader. Expert instructors share practical strategies to help you build trust, communicate clearly, and motivate others to do their best work.
