On my commute home the other day a stranger asked if he could get my opinion on a song he had written. He played it for me on his phone, and it was great! It had a soothing melody and the words (from what I could hear over the loud train) had a nice flow that elicited melancholy. He then excitedly showed me how he used AI to create the song with his lyrics and carefully worded prompts on the envisioned melody and cadence.

What used to take a whole team of producers, singers, musicians and others could be achieved on his device in less than two minutes.

We all know that AI is here to stay and there are a plethora of ways that it can—and has—transformed the accounting profession. As the head of my firm’s tax department, I’ve thought a lot about how we can leverage AI, specifically with respect to tax research.

We tested two popular AI-assisted tax research platforms simultaneously over the summer, and the results were concerning. We fed the same question into both platforms and often got two very confident sounding responses giving us completely opposite answers.

Neither was consistently right or wrong.

In one instance, I asked them to explain the implications of a passthrough entity that wanted to make the California passthrough entity tax (PTET) election in 2025 but had missed the required June 15 estimated payment. One responded—incorrectly—that there would be late payment penalties and interest. It provided citations to support its answer, but when I checked those sources, they either didn’t exist or did not support its response. When I pointed this out, it thanked me for questioning it and then offered additional citations that were likewise fictitious or irrelevant.

The correct answer, by the way, is that you’re out of luck and can’t make the election for the year.

During this testing, I would often get into circular discussions with the platforms. They would give me the wrong answer, agree with the answer I was tentatively considering and then circle back to the original wrong answer.

When I relayed these issues to the sales representatives, both asserted the problem was that I wasn’t asking the questions “correctly” to get the right answer. That response raised an even bigger concern: If seasoned researchers can’t reliably prompt their way to accuracy, how often will less experienced professionals be presented with confidently delivered, but incorrect, results (and blindly accept them)?

At the rate at which AI is developing, those two platforms have likely been significantly improved since the summer. However, my message to tax practitioners using any kind of AI-assisted tax research service is to do so with eyes wide open and with the professional skepticism you would normally use if someone whose tax background you’re not familiar with told you something.

Even in my day to day working with my intelligent colleagues, whom I’ve worked closely with over the years, if we’re talking about a new-to-me tax issue, I’m going to want to see the primary authority to confirm it for myself.

Before the emergence of AI in recent years, I’ve always told our staff not to rely on third-party summaries, even if they come from our tax research software. Primary authority consists of the legally binding sources of tax law, including statutes, regulations and court decisions that the IRS and courts rely on, whereas secondary authority includes explanatory materials such as articles, treatises and research summaries that may aid understanding but cannot stand on their own as legal authority.

Secondary authority is a good starting point and gives you some context, but best practice is to check the citations and trace them to the primary authority to confirm the answer. If best practice is to use tax research software, which has been widely used in the profession for decades, with this approach of trust but verify, why would you not, at a minimum, do the same due diligence when using these newer AI-assisted platforms? Our established tax research platforms now have an AI assistant built in, but the approach doesn’t change: You must keep tracing the citation back to the primary authority.

No matter how confident AI sounds and how much it strokes your ego praising you for asking such great questions, don’t lose sight of the foundational research principle that, if you were to be audited, the IRS is not going to accept a citation to any third-party summary. Primary authority continues to be king. Do your homework. Don’t take citations at face value. Trace the cite to the primary authority and verify.