Your tax return is being read by a model somewhere in a federal data center. Not an individual. Not a worker in an Ogden, Utah, beige cubicle with a flickering fluorescent light above their desk. A model. At the same time, it scores, ranks, flags, and discreetly transmits its recommendations to the remaining humans at the IRS so they can be implemented for approximately 150 million other Americans.
Actually, there aren’t many of those humans left. The agency reduced its workforce by roughly 25% between January and May of last year, from 103,000 to about 77,000. Many of the individuals who left were seasoned tax examiners and revenue agents who knew how to track down a six-figure deduction or unravel a convoluted partnership return. A budget cycle does not rebuild that level of expertise. Thus, the organization has done what most organizations do when they lose both personnel and funds simultaneously. It relied more on software.
The IRS had 126 active AI use cases as of last summer. Some of them, like voice bots that respond to inquiries about the status of refunds, are amiable enough that no one finds objectionable. More than 4.8 million calls have reportedly been handled by those bots, and chatbots have answered about 450,000 questions without a human ever responding. Alright. That is the half of the story that is visible. On the enforcement side, models such as the Individual Taxpayer Model, the Large Partnership Compliance Model, and the Discriminant Function crunch through returns six times a year, learning a little more each time. This is the half that most people are unaware of.
According to the official line, honest filers will benefit from this. Fewer compliant individuals will be drawn into “no-change” audits—those in which the IRS knocks, looks around, and finds nothing—if the algorithms perform as the agency hopes. The machine should ideally focus on the real issues, such as partnerships concealing income, self-employed filers deducting a few thousand from their gross receipts, and returns with that oddly round $10,000 charitable deduction that doesn’t exactly match anyone’s bank statement. With a quarter fewer employees, enforcement revenue increased by 12% in fiscal 2026—a figure that is frequently cited favorably during budget hearings.
Beneath, though, is a more subdued tale. Black taxpayers are audited at three to five times the rate of other groups, according to independent research, and the GAO has identified inadvertent algorithmic bias as a contributing factor. A model is taught more than just the law when it is trained on decades’ worth of enforcement data. It picks up the agency’s old habits from you, and habits spread. Robodebt, an automated debt-recovery program that falsely accused thousands of citizens of owing money they didn’t owe, taught Australia this lesson the hard way. Both in court and in public confidence, it fell apart. The similarities are obvious.
Transparency—or the absence of it—is another issue. Because of worries that taxpayers might manipulate the results, the IRS won’t reveal much about how its scoring systems actually operate. Perhaps that makes sense. Perhaps it isn’t. What it leaves behind is a system in which you receive a notice in the mail requesting documentation for a deduction that occurred three years ago, and no one on the other end can adequately explain why your return in particular was selected.
The useful advice has not changed for the majority of filers. Maintain your documentation. Avoid rounding things off. Report the weekend gig’s side income. However, it’s difficult to ignore the fact that something that doesn’t pick up the phone is changing the game’s rules.
Leave a Comment