A $300,000 house in Atlanta, owned by a woman who pays her property taxes for fifteen years, writes the checks on time, and never thinks much about the rate because nobody sends you a letter explaining that the rate is unfair. Over that decade and a half, she overpays by $24,750 compared to the owner of a $900,000 house two miles north, who benefits from a lower effective tax rate on every dollar of assessed value. Not lower taxes in absolute terms, but a lower rate, applied silently, compounding every year, hidden inside an envelope she never thought to challenge.
That $24,750 buys a new roof, or a kitchen renovation in a starter home, or the margin between building equity and treading water. It happens in 97.7% of U.S. counties, according to University of Chicago researchers who studied assessment data nationwide, which makes property tax assessment America's largest stealth tax: over $600 billion collected annually, systematically regressive for as long as anyone has bothered to measure it.
How "Data Rot" Eats Assessments From the Inside
Peter Gariepy, a CPA who consults on property tax policy, coined a term for what happens inside legacy assessment systems: "data rot." Legacy computer-assisted mass appraisal (CAMA) systems use smoothing algorithms that pull every property toward the local average, which means a $30,000 home gets assessed at $60,000 while a $500,000 home gets assessed at $420,000. Both errors look small in percentage terms, but one of them doubles your tax burden.
Wealthy homeowners appeal, hiring attorneys and commissioning private appraisals to successfully reduce their assessed values at rates between 40% and 60%, according to the National Taxpayers Union. Those reduced values become comparable sales in the system's training data, so next cycle the algorithm learns that similar homes sell for less and assessments drop accordingly. Meanwhile, the owner of the $30,000 home, who cannot afford an attorney and may not even understand the appeals process, never challenges the inflated figure. Her over-assessment becomes permanent, compounding year after year in a vicious cycle that Gariepy calls data rot.
The Philadelphia Federal Reserve measured the result: owners of inexpensive houses pay effective tax rates nearly 50% higher than owners of expensive houses in the same jurisdiction. In Atlanta, 69% of the lowest-value properties are over-assessed, compared to 32% of the highest-value properties.
AI Enters the Assessor's Office
Riverside County, California, signed a five-year contract with C3 AI to overhaul its appraisal pipeline, and the results are concrete: thirty separate legacy models collapsed into four unified AI models, a 40% speed improvement in the appraisal cycle, and tasks that previously consumed hours finishing in minutes. New York City's Department of Finance launched a parallel pilot using C3 AI to revalue condominium properties with the stated goal of creating "a more fair and transparent way to assess properties."
Cook County, Illinois, the most studied assessment jurisdiction in the country, offers the most instructive case because it shows both sides of the reform ledger. After a ProPublica investigation exposed deep unfairness under the prior assessor, Cook County Assessor Fritz Kaegi reformed the system, and a University of Chicago study found the reforms saved homeowners $1.9 billion. Researchers also built PTAXSIM, a publicly available R package that simulates property tax impacts using 2006 through 2020 assessment data for every parcel in the county, which anyone can download and run against their own property. That transparency is the point.
But Cook County also demonstrates why transparency alone isn't sufficient, because commercial property owners responded to the reformed residential assessments by filing appeals that shifted approximately $2 billion in tax burden back onto homeowners over three years. In 2024, roughly 240,000 Cook County homeowners saw property tax spikes of 25% or more, totaling an estimated $500 million in increased bills. The reform improved the residential assessment methodology while commercial appeals clawed the savings back through a different door entirely.
Showing the Math
Consider a homeowner purchasing a $300,000 property in a jurisdiction with the national average effective property tax rate of approximately 1.1%, which yields a fair annual tax bill of $3,300. Now apply the Philadelphia Fed's finding that lower-value properties face effective rates up to 50% higher, pushing the effective rate to 1.65% and the annual bill to $4,950, an overpayment of $1,650 every single year.
| Metric | Fair Assessment | Regressive Assessment |
|---|---|---|
| Property Value | $300,000 | $300,000 |
| Effective Tax Rate | 1.10% | 1.65% |
| Annual Tax Bill | $3,300 | $4,950 |
| Annual Overpayment | $0 | $1,650 |
| 15-Year Cumulative Overpayment | $0 | $24,750 |
Inputs: national average effective tax rate of 1.1% per the Tax Foundation, 50% regressivity penalty per the Philadelphia Federal Reserve (top-decile vs. bottom-decile comparison), 15-year holding period per NAR median tenure data. No compounding or rate increases applied. The $24,750 figure is a floor estimate.
Glass Box or Black Box
Iowa passed a law in 2026 requiring assessors to explain any valuation spike exceeding 10% over two years, which shifts the burden of proof from the homeowner to the assessor's office for the first time in any state. Researchers at UChicago Booth School have published a K-Segment valuation model that formally combines accuracy and fairness metrics, embedding equity constraints directly into the optimization function rather than bolting them on after the fact.
Gariepy frames the choice starkly: "glass box" versus "black box." A glass box system publishes the comparables it selected, the weights it assigned, and the adjustments it made, so a homeowner can check whether the algorithm compared her 1,200-square-foot ranch to another 1,200-square-foot ranch down the street or to a 2,400-square-foot colonial in a different school district. A black box system outputs a number and asks you to trust it.
C3 AI's promotional materials emphasize speed and consolidation: Riverside County's thirty models became four, and tasks that took hours now take minutes. What neither C3 AI nor Riverside County has published is the model architecture, the feature weights, or a regressivity audit of the outputs. Fast and opaque is an improvement over slow and opaque, but it is not the reform that the Philadelphia Fed's data demands.
The Case Against Transparency
Disclosing how the model works could make regressivity worse, not better, because if a glass box system reveals exactly which comparables and adjustment factors drive each assessment, homeowners with resources gain a precise roadmap for gaming the appeals process. They can select counter-comparables, challenge specific weights, and argue adjustment factors with surgical precision that an unrepresented homeowner cannot replicate, which means the appeals gap, already running at 40 to 60% success rates among those who file, could widen rather than narrow.
There is a deeper problem with AI-driven assessment reform. Systems trained on historically regressive assessment data will reproduce the bias in the training set unless fairness constraints are explicitly engineered into the model's architecture. The Urban Institute has documented that automated valuation models disproportionately understate home values in majority-Black neighborhoods, where lower sales prices driven by systemic factors feed back into the algorithm as "evidence" of lower value, creating a feedback loop that glass box visibility can expose but cannot, by itself, break.
What This Analysis Didn't Prove
The $24,750 overpayment figure applies the Philadelphia Fed's top-decile versus bottom-decile regressivity finding to a national average tax rate, and regressivity varies enormously by jurisdiction, with some counties performing worse and a handful performing better. The median disparity across all homeowners is lower than the 50% figure, which represents the extreme tails of the distribution rather than the typical experience.
Cook County's $1.9 billion in residential savings is real, but the $2 billion commercial clawback demonstrates that assessment reform operates inside a zero-sum tax system where shifting burdens off one class of property inevitably shifts them onto another. C3 AI's deployments in Riverside County and New York City are recent enough that independent, longitudinal outcome data does not yet exist, so we are relying on vendor-reported speed metrics rather than third-party fairness audits.
The Urban Institute's AVM bias findings are strongest in neighborhoods with thin transaction volumes, and in high-liquidity markets with robust comparable sales data, AI valuation models perform significantly better, raising the possibility that the regressivity problem may be driven more by human appeals behavior than by model architecture in well-traded markets.
What You Should Do
Look up your assessment, because every county publishes assessment records online. Compare your assessed value to recent comparable sales within a half-mile radius, and if your assessed value exceeds your home's likely market value by more than 10%, you are almost certainly overpaying.
File an appeal when the numbers don't add up. The National Taxpayers Union reports that 40 to 60% of appeals succeed, with the average successful appeal reducing assessments by 10 to 15% and saving $500 or more per year, and your county assessor's website lists the appeal window, the required forms, and the comparable sales evidence format. Most residential appeals do not require an attorney.
Ask whether your county uses AI in assessments, and if it does, ask whether the model's methodology is publicly documented. If it is not, you are inside the black box, and you should plan to appeal regularly rather than trust the output. Iowa's new law may spread to other states. Until it does, the burden of proof sits with you, and the $24,750 question is whether you're willing to carry it.