The Internal Revenue Service’s use of artificial intelligence in selecting tax returns for National Research Program audits that are used to estimate the tax gap needs more documentation and transparency, the U.S. Government Accountability Office stated.

In a report issued June 5, 2024, the federal government watchdog noted that while the agency uses AI to improve the efficiency and selection of audit cases to help identify noncompliance, “IRS has not completed its documentation of several elements of its AI sample selection models, such as key components and technical specifications.”

GAO noted that the IRS began using AI in a pilot in tax year 2019 for sampling tax returns for NRP audits. The current plan is to use AI to create a sample size of 4,000 returns to measure compliance and help inform tax gap estimates, although GAO expressed concerns about the accuracy of the estimates with that sample size.

“For example, NRP historically included more than 2,500 returns that claimed the Earned Income Tax Credit, but the redesigned sample has included less than 500 of these returns annually,” the report stated.

IRS told GAO that it “is exploring ways to combine operational audit data with NRP audit data when developing its tax gap estimates. IRS officials also told us that if IRS can reliably combine these data for taxgap analysis, IRS might be better positioned to identify emerging trends in noncompliance and reduce the uncertainty of the estimates due to the small sample size.”

The report also highlighted the fact that the agency “has multiple documents that collectively provide technical details and justifications for the design of the AI models. However, no set of documents contains complete information and IRS analyst could use to run or update the models, and several key documents are in draft form.”

“Completing documentation would help IRS retain organizational knowledge, ensure the models are implemented consistently, and make the process more transparent to future users,” the report stated.

By Gregory Twachtman, Washington News Editor