"No" was the short answer to the question posed by a session titled, “Will Artificial Intelligence Take Over Tax Practice?” at the Foundation for Accounting Education’s Taxation of Financial Instruments Conference on Jan. 9. 

Panelists Travis Thompson, a criminal tax associate at law firm Sideman and Bancroft, LLP, and  Caitlin Tharpe, a federal tax controversy and litigation associate at law firm Steptoe LLP, considered many of the facts and perceptions concerning generative AI tools such as ChatGPT, while discussing the practical implications of using the technology in everyday practice.

“You could just read the news and hear about how ChatGPT and generative AI are all going to take over our lives and change the practice of tax forever,” said Thompson, who said that he has been talking about AI and how the IRS has been using it since 2018.

He framed the discussion with a simple definition of artificial intelligence—“machines acting in ways that seem intelligent”—before mentioning three types of Narrow AI, which Technopedia defines as "a learning algorithm  designed to perform a single task, and any knowledge gained from performing that task will not automatically be applied to other tasks." The three types of Narrow AI that he delineated were machine learning: the use of predictive computer programs that automate learning through experience using large data sets; natural language processing: technology that enables machines to understand and generate language in its written and spoken form, such as Microsoft Spell Check and Google Translate; and robotics: automated devices that perform physical tasks in the real world.

Thompson jocularly referred to the movies The Terminator and 2001: A Space Odyssey as people’s concepts of AI and what it does. But, he said, the specter of self-aware machines such as HAL will never happen, according to practitioners and AI experts.

AI is a data accumulator, he said, comparing it to a smartphone. “The phone is just something [that was] slapped on at the end,” he said. “The companies want your data, right?"

He said that the IRS and other taxing agencies are using AI programs "in order to enforce the tax code.”

Tharpe then continued that theme by noting that AI is not good for certain things, such as research, objectivity or accuracy, but it is useful for tasks such as creating, improving, and summarizing and generating and validating data.

Tax professionals can use generative AI for producing basic letters and memoranda, such as drafting a letter to the IRS regarding the receipt of an audit letter, such as a CP 106 notice. she said. She warned that the technology cannot be relied upon for tax advice, illustrating her point with a slide that showed Chat GPT’s response to her prompt. “As a language model, ChatGPT is not a licensed tax professional and cannot provide specific tax advice,” it stated, in part.

“Simple basic questions ChatGPT can answer, but … if you were asking a question, you'd want to ensure that you validate the result that you're getting because of ChatGPT's potential to hallucinate,” she said. "Hallucinate" is the term used to describe an AI model’s generation of incorrect information that is presented as fact, she explained. She also cautioned that ChatGPT “is not aware of events beyond 2021, because that is the information on which it was trained and so it's not aware of recent events.”

The duo then spent some time discussing IRS Circular 230, which defines “practice,” spells out who may practice before the IRS, and describes their duties and obligations.

“The practitioner needs to have the necessary competence,” said Tharpe. “And there are interesting questions presented like … you can't just be taking answers from ChatGPT whole cloth when you don't know whether they're right, and that won't make you a competent practitioner. But there's also the need to keep up with technology and become competent in the use of technology.”

Specifically, Circular 230  states that the practitioner needs to ensure the accuracy of statements that are made to the IRS, she said, and to exercise due diligence in supervising others and relying on the work product of another person. The supervisor must ensure that “people that that you supervise as a tax practitioner, younger people perhaps, ... are... using generative AI ... properly.”

The panelists also considered how the IRS is using AI. Thompson pointed out that the agency has been using the technology and algorithms to select taxpayers for audit and to identify and determine issues.

But history has recently demonstrated how that reliance can result in bias, Thompson said .He illustrated that issue by bringing up the widely reported discovery by Stanford University researchers that the IRS audits Black taxpayers at 2.9 to 4.7 times the rate of non-Black taxpayers.

The study attributed the main source of the disparity between Black and non-Black taxpayers audit rates to "differing audit rates by race among taxpayers claiming the Earned Income Tax Credit (EITC)." This was mainly due to automated algorithms that flag discrepancies in claims for tax credits, it found.

The researchers “didn't know how the IRS was using the AI. They were going to test for racial bias in the tax system, based on the use of artificial intelligence,” said Thompson. “So, Stanford had to bring AI engineers and computer engineers to develop their own set of AI independently because the IRS is clearly not going to share what they're using.”

But the IRS is seeking to continue to increase its AI capabilities, Thompson said. In a July 2023 performance work statement, the IRS Office of Research, Applied Analytics and Statistics explained that the IRS uses AI due to “fewer resources in the field” and in order to focus on “specific non-compliant issues” and to “reduce the no-change rate.”

The panelists concluded the session with a warning for tax professionals.

“As technology in this space develops in the tax world, the economics of how we practice is changing,” said Thompson. “The days of $400 and $500 an hour tax research projects are slowly coming to a close, right? Because if I can get to the answer in 10 seconds using generative AI products in a closed system, then that means everybody can get to that answer in 10 seconds."

“So, how can we justifiably charge our clients that much?” he asked. “So, keep that in mind as well for the future of tax practice.”