The IRS estimates that hundreds of billions of dollars in owed taxes go unpaid every year. This “tax gap” is only growing––in the most recent IRS projections, it was estimated that in the year 2022 $696 billion went unpaid, and the IRS will likely recover only $90 billion of that sum. In the last few years, the IRS has turned to AI to reduce this gap. A Treasury Inspector General For Tax Administration report released on May 19, 2025 gives some insight into these developments.
Although AI has the potential to reduce the tax gap and decrease the number of unnecessary audits, it also means some taxpayers may face heightened scrutiny. Taxpayers should be aware of these developments to understand how they can affect their risk of being audited.
What Do We Know About the IRS’s AI Models?
Three groups the IRS is targeting are large partnerships, self-employed taxpayers, and small business owners.
Two AI models target self-employed taxpayers and small business owners. Self-employed underreporting constitutes the largest portion of underreported individual tax liability. This may be in part because their income is less visible to the IRS because they are not employees who have employers to file information returns and withhold taxes. For decades, the IRS has used the Discriminant Index Function (“DIF”) to screen returns for audits. The DIF is not AI, but a computer program that searches returns for values that are unusual in light of historical data. If the DIF marks a return as unusual, that return is more likely to be audited. Recently, the IRS’s Small Business/Self-Employed Division began using an AI model to analyze DIF-selected returns and identify issues for auditors. For each line item, the model compares the reported amount to what the AI expects. Line items with greater deviations from the AI’s expectation are marked high-risk. The IRS runs the model about six times per tax year and the model learns with each run. This automation allows the IRS to reassign some “classifiers”––IRS employees who highlight issues for audit in DIF-selected returns––into auditing. IRS management estimates that the resulting increase in auditor staffing and simplification of the training return classification process could allow it to recover between $31.7 and $34.2 million more annually.
The Small Business/Self-Employed Division also possesses a return selection model that would apply a risk rating to each self-employed taxpayer’s individual income tax return. Tests revealed that the model would recover more dollars per hour than the previous method but also that the model would cause a higher proportion of unnecessary audits. In light of these results, the Division has not yet decided to implement this second model.
The IRS’s Large Business and International Division has also been experimenting with AI. One model targets large partnerships. The IRS has had trouble in the past auditing large partnerships. This is because large partnerships can have complex nested structures that make it difficult to track income and losses. The AI “Large Partnership Compliance Return Selection Model” screens each large partnership’s tax documents to determine the risk that it is reporting incorrect information. As of April 2024, the Large Partnership Compliance Return Selection Model selected 82 returns for audit for the 2021 tax year. But there is not yet any data showing whether the selections were effective.
What Do We Not Know?
The IRS is notoriously secretive regarding the development of its AI programs. On the one hand, if too much information gets out, taxpayers can find ways to cheat the AI. On the other hand, excessive secrecy leaves the IRS unaccountable for mistakes or problems with the models. So what don’t we know about these AI models?
We know that the Small Business/Self-Employed Division’s programs are trained on the tax returns from the present year, but we also know that each model is trained on other, unidentified, information.
We know that the Large Business and International Division’s Large Partnership Compliance Return Selection Model performs two functions to identify the risk that each Large Partnership reported incorrect information on a return. The first function is based in data science––the model looks through returns to spot abnormal entries. The model’s second function is undisclosed.
Data evaluating the models’ effectiveness is limited. Part of this is because the models are so new. But it is also because the IRS has not developed adequate processes to evaluate the models’ performance once they are deployed.
Why is secrecy a problem? AI models trained on faulty data can become inaccurate or biased. Without disclosure, the public has no way to know of this or any other issues with the AI models. The IRS has been pressed to release more information but has declined, claiming transparency would hamper its ability to effectively administer the tax system.
Some takeaways. The IRS is using AI to free resources from the process of selecting returns for audit and is moving them to the audit process. This means the IRS will have the ability to conduct more audits for groups covered by divisions using AI models, like small business owners, self-employed people, and large partnerships. Recent and potential staffing cuts at the IRS may reduce or neutralize this increased efficiency. And it is too early to say whether audits on AI-selected returns will be better targeted than those predating the models. As the IRS utilizes AI to streamline its selection processes, it is more important than ever for taxpayers to ensure they are complying with the tax code. Seeking professional advice can help taxpayers prevent audits and proactively plan to minimize tax liabilities in the future.
For any questions, please reach out to the Tax, Trusts and Estates practice group at Mandelbaum Barrett PC.