The IRS isn’t exactly known for its use of cutting-edge technology. The federal agency still accepts paper tax returns and can take years to send delayed tax refunds.
Yet the IRS has enough artificial intelligence (AI) projects that they couldn’t initially provide an accurate number to the Treasury Inspector General for Tax Administration (TIGTA) — an independent federal agency that oversees the IRS.
Not only that, but employees on the IRS’ network were able to access external generative AI tools such as ChatGPT — even though they’re prohibited — according to a recent TIGTA report.
The IRS is using its own AI tools primarily to improve its overall operations — presumably a good thing. But IRS employees’ use of external AI tools is a cause for concern for TIGTA, which writes in its report:
“When employees send queries to generative AI online tools, they might inadvertently reveal sensitive or non-public information to the generative AI tool.”
In other words, when IRS employees use AI tools that are not controlled by the federal government, they could end up leaking sensitive information to the companies behind or other users of those AI tools.
Here are some of TIGTA’s findings based on a review of traffic logs:
- 531 attempts were made by IRS employees to access ChatGPT from Oct. 31, 2023, to Jan. 29, 2024.
- IRS employees successfully accessed one particular generative AI website 191 times over the same period. (TIGTA’s report does not name that particular website.)
- One employee was able to bypass IRS security controls to successfully access a prohibited generative AI website. (The report doesn’t name that site, either.)
Based on its findings, TIGTA made two formal recommendations to the IRS:
- Track and regularly review employee traffic to AI platforms, and update rules and permissions accordingly to prevent access to prohibited AI tools.
- Speed up the process of creating new governance that ensures the safety, security and trustworthiness of AI developed for or used by the IRS.
The IRS agreed with both of TIGTA’s recommendations and had already started working on the latter. The IRS’ chief data and analytics officer also said the agency will hire additional employees to assist with such efforts.
Now, what about all those AI projects the IRS initially struggled to take inventory of?
The IRS reported to TIGTA in mid-2023 that they had 33 AI projects. TIGTA’s review at the time determined that 16 of those 33 projects were not actually AI.
Several months later, the IRS provided an updated inventory, saying the agency had 68 AI projects that were in development or already active as of Dec. 30, 2023. TIGTA’s report details the purpose of these 68 projects:
- 12 projects are for improving customer service. This includes the Wage and Investment Refunds Chatbot Project, for example, which answers frequently asked taxpayer questions about topics like tax refunds with predetermined answers.
- 27 projects are dedicated to improving enforcement. One example is the National Research Program Redesign Project, which uses AI to help with tax gap estimation and improper payment reporting, among other tasks. (The tax gap refers to the difference between the amount of money that taxpayers owe the IRS and the amount that the IRS has actually collected.)
- 29 projects are dedicated to improving overall IRS operations. This includes the Appeals Case Memorandum Project, which uses AI to analyze documents and give insight into improving future IRS audits. There’s also the Machine Translation Project, which uses AI to translate text into Spanish.
Add a Comment