Measuring Tax Enforcement with Generative AI

Abstract

I use a generative AI model to create new measures of tax enforcement based on 10-K filings. The model can identify active, ongoing IRS audits at a 96% accuracy rate compared to a tax researcher manually labeling the same disclosure. I find considerable time-series variation in disclosures of IRS audits in a manner consistent with changes in the IRS budget. I further find that the disclosed audit measures are positively correlated with existing measures of tax enforcement including audit probabilities published by the IRS, 10-K downloads by the IRS, a publicly available predicted audit risk measure developed on proprietary IRS data, and changes in unrecognized tax benefits from settlements with tax authorities. These new disclosed audit measures can be used by researchers to further study the determinants and consequences of tax enforcement.

Keywords: Generative AI, tax enforcement, tax audit, IRS audit