The IRS’s inspector general is warning the IRS over its embrace of artificial intelligence, saying the tax agency must develop guardrails to prevent AI from being used in discriminatory ways.
The IRS is using AI to improve customer service, translate documents into other languages and help pick which tax returns to audit and what level of examiner to assign.
But the inspector general said without careful guidelines, AI could reinforce existing issues with discriminatory approaches to tax enforcement.
“Although the use of AI can improve IRS operations and overall tax administration, the IRS must be diligent in its use to ensure that taxpayer rights are adequately protected,” the IG said. “Biases arise from the fact that AI systems are created using data that may reflect preexisting biases or social inequities.”
The inspector general said it was “imperative that the IRS accelerate implementation of governance and oversight structures” to blunt those issues.
Investigators said the IRS is making a serious effort to develop AI rules, but investigators said some holes have already emerged.
For one thing, the tax agency sometimes seems confused about what AI is. The inspector general said the IRS initially provided a list of 33 AI projects last year, but 16 of them weren’t even AI.
The IRS then provided a final list of 68 projects.
It said 30 of them are already in use and 38 are in development.
Many of the projects are chatbots. They try to answer taxpayers’ questions automatically, searching for key words in the questions and spitting out responses tailored to those prompts.
Still others are part of enforcement. The IRS is using one AI tool to identify organizations that bogusly claim tax-exempt status.
And the IRS hopes to use AI to combat discrimination with two pilot programs aimed at reducing race disparities in tax returns selected for audits.
Investigators also issued a warning about IRS employees seeking out third-party AI websites, particularly generative AI, which takes user prompts and spits out tailored content.
“When employees send queries to generative AI online tools, they might inadvertently reveal sensitive or nonpublic information to the generative AI tool entities, which might be available to access by other users,” the warning said.
The IRS tries to block access on its network to third-party generative AI options. More than 500 attempts to access OpenAI’s ChatGPT were blocked from Oct. 31, 2023, to Jan. 29, 2024.
The inspector general said employees have found workarounds, with one worker using an external training platform to get to a generative AI site. The IRS moved to fix the loophole.
The IRS, in an official response to the inspector general, acknowledged investigators’ cautions and said the agency takes potential pitfalls seriously.
“As you describe in the report, we have created robust AI governance framework and processes,” said Barry Johnson, chief data and analytics officer.
He agreed with both of the inspector general’s recommendations for improvements.
• Stephen Dinan can be reached at sdinan@washingtontimes.com.
Please read our comment policy before commenting.