A man in the Midwest received a letter from the IRS last spring informing him that he owed over $90,000 in unpaid taxes. He possessed a W-2. He had never made over $30,000 in a single year. An agent eventually acknowledged that the agency had erred when he called to resolve the issue. He never found out what set off the alert. No one informed him if an algorithm had simply determined that he appeared suspicious or if a human had examined his return.
A disturbing aspect of the state of American tax enforcement in 2026 is captured in that story, which was posted on a Reddit thread that received hundreds of comments in recent weeks. Just two years ago, the IRS had 54 active AI applications; today, it has over 129. These systems handle millions of taxpayer calls without a human ever picking up the phone, flag discrepancies across years of filing history, score tax returns for audit potential, and highlight “highest-value” cases for criminal investigation. Additionally, the individuals whose financial lives are being assessed by these instruments have very little insight into how any of it operates.
It’s difficult to ignore how subtly this change occurred. Since the middle of the 20th century, the IRS has used some kind of algorithmic scoring, such as the former Discriminant Information Function, or DIF score, which was applied to each return. However, the scope and goals of what is currently taking place are completely different. Millions of returns are now analyzed concurrently by machine learning models, which identify patterns that differ from a filer’s personal history, anomalous deduction ratios, and variations in income from year to year. Recent years have seen the introduction of the Large Partnership Compliance Model, which focuses on hedge funds and private equity structures that were previously too complicated to analyze. It chose 82 high-risk returns in 2021; prior to AI, comparable models were choosing single digits.
The Selection and Analytic Platform, or SNAP, is a tool that the IRS paid Palantir $1.8 million to create in order to identify fraud and audit cases from what the agency called a “fragmented landscape” of more than a hundred internal systems. According to documents WIRED was able to obtain, the pilot has been analyzing gift tax returns, clean energy credits, and disaster zone claims, using unstructured data hidden in supporting documents. Scholars who research the gift tax code note that SNAP is operating in a field where even seasoned lawyers cannot agree on methodology if it is examining balance sheets and appraisal disclosures on private business transfers. It’s still unclear how a machine weighs those documents.
The timeline associated with the opacity contributes to its significance. Taxpayers who receive AI-generated audit notices now have as little as ten days to reply before the IRS moves to levy bank accounts, according to CPA Diane Kennedy. If the call doesn’t drop at the connection point, which seems to happen frequently enough that it’s a known grievance, ten days is not much time when you don’t know what set off the flag and when wait times to speak with a live IRS agent can exceed two hours.
In a report released in March, the Government Accountability Office identified significant shortcomings in the IRS’s documentation and oversight of its own use of AI. The agency had not even documented the intended benefits of the AI in over 25% of use cases. A number of the instruments used to construct criminal cases were completely absent from the official inventory. Independent research indicates that Black taxpayers are being audited at rates three to five times higher than comparable filers, and the GAO discovered evidence of inadvertent algorithmic bias. AI tends to replicate discriminatory patterns quickly and on a large scale when it is trained on historical data that already contains those patterns.
The IRS has used law enforcement exemptions to defend the secrecy of its algorithms, claiming that disclosing details would enable tax evaders to manipulate the system. On the surface, that is not an irrational worry. However, it puts regular filers in a unique situation where they are subject to a financial judgment that they are unable to see, fully comprehend, or appeal in any technical way. Experts frequently refer to Australia’s Robodebt scandal, in which an automated welfare debt system wrongly pursued hundreds of thousands of people before being declared illegal. Although the systems differ, the structural issue remains the same. A demand was made by an algorithm. Clearly, no one was in charge of making sure it was correct.
Whether the IRS has the human ability to identify errors made by its own machines is still up for debate. In 2025 alone, the organization lost about 25,000 workers, including 63 members of the analytics team in charge of AI supervision. In many cases, the people who are most likely to notice when the algorithm goes wrong are no longer there.
