On October 1, 2025, a rule fourteen years in the making finally took effect. Six federal agencies, the CFPB, FDIC, FHFA, Federal Reserve, NCUA, and OCC, activated quality control standards for automated valuation models, the algorithms that estimate what your home is worth without anyone walking through the front door. It requires lenders to adopt policies ensuring AVMs produce high-confidence estimates, avoid conflicts of interest, and comply with nondiscrimination laws.
That same month, the Urban Institute published a study that tested whether those algorithms actually treat homeowners equally.
They do not.
Researchers Linna Zhu, Judah Axelrod, and Amalie Zinn analyzed 43,024 property transactions in Atlanta and Memphis, using a Bayesian technique called fBISG to infer homeowner race from surnames and addresses. They controlled for lot size, building age, number of bedrooms, exterior property condition (scored by CAPE Analytics computer vision from aerial imagery), neighborhood median income, gentrification status, distressed sales share, and housing turnover rate. They included county fixed effects and ran 100 simulated regressions to account for imputation uncertainty.
After all of that: Black homeowners still faced AVM errors 3.4 percentage points larger than white homeowners with comparable properties in the same neighborhoods. At a median sale price of $420,000, that is $14,960 in additional algorithmic mispricing. Statistically significant at p < 0.01.
The direction matters as much as the magnitude. This was not random noise scattering errors evenly. Black-owned homes were systematically valued lower by the algorithm. Five percent lower, to be precise, after controlling for everything the model could observe about the property and its surroundings. Hispanic homeowners fared slightly worse on error magnitude (3.5 percentage points) and experienced 3.7 percent undervaluation. Asian homeowner sample size was too small for the undervaluation finding to reach statistical significance.
What the Rule Actually Requires
The interagency AVM rule is five requirements long. Lenders must adopt policies designed to ensure a high level of confidence in AVM estimates, protect against data manipulation, avoid conflicts of interest, require random sample testing and reviews, and comply with applicable nondiscrimination laws. It applies to mortgage originators and secondary market issuers, including Fannie Mae and Freddie Mac, when they use AVMs to value a consumer's principal dwelling for a credit decision or securitization.
Notice what the rule does not contain. No specific accuracy threshold. No mandated testing protocol. No required performance metrics disaggregated by race. No mechanism for a homeowner to challenge an AVM estimate the way they can dispute a human appraiser's report. No definition of what "comply with nondiscrimination laws" means in the context of an algorithm that never explicitly uses race as an input but produces racially disparate outputs through historically biased training data. Mondaq's legal analysis of the final rule noted that "the substance is in the defined terms," a polite way of saying the rule is broad enough to mean almost anything and specific enough to require almost nothing.
By design, the rule is principles-based. All six agencies argued that prescriptive requirements would become obsolete as technology evolves, and that institutions should have flexibility to implement quality controls appropriate to their specific AVM usage. This reasoning is defensible in the abstract. In practice, it means the rule has been in effect for nine months, and no institution has been required to demonstrate that its AVM produces equitable outcomes across racial groups, because no one defined what equitable outcomes would look like.
Meanwhile, the Pipeline Is Expanding
While the rule settled into regulatory existence, the government-sponsored enterprises quietly expanded the use of algorithmic valuations. American Enterprise Institute data tracking GSE appraisal waivers shows the share climbing back to 28 percent of all GSE loans by March 2026, up from a post-pandemic trough, driven by higher waiver rates within each loan purpose and a modest shift toward refinance activity where waivers are more common. In the first quarter of 2025, Fannie Mae expanded waiver eligibility to purchase loans with combined loan-to-value ratios between 80 and 90 percent, a population that had previously required a human appraisal. By March 2026, about 9 percent of those newly eligible loans used a waiver, up from 2 percent twelve months earlier.
Freddie Mac introduced its ACE+PDR program in July 2022, combining automated collateral evaluation with property data reports. Fannie Mae followed with Value Acceptance + Property Data in April 2023. Both programs replace the traditional in-person appraisal with some combination of algorithmic valuation and third-party data collection that may or may not include a physical property visit.
Clear story: more homes are being valued by algorithm, fewer by human. Meanwhile, the appraiser workforce is shrinking, an older cohort that is not being replaced at scale, partly because of credentialing barriers and partly because the GSEs are building a future that needs fewer of them. Researchers at the convergence of AI and real estate describe an $11.3 billion appraisal industry undergoing "fundamental market restructuring" as Fannie Mae's Uniform Appraisal Dataset moves from narrative-based reporting to structured XML formatting by 2026.
Where the Bias Actually Lives
The Urban Institute study identifies two sources of the racial gap, and the distinction matters for anyone trying to fix it.
First, biased training data. AVMs learn from historical transactions, comparable sales, and property assessments that reflect decades of redlining, steering, and discriminatory appraisal practices. When an AVM in Memphis pulls comparable sales from a majority-Black neighborhood where home values were systematically suppressed for generations, it is not introducing bias. It is faithfully reproducing it. No algorithm knows it is doing this, any more than a calculator knows the numbers you feed it were wrong. Garbage in, precision-formatted garbage out. Nobody audited the inputs.
Second, optimization bias. AVMs are trained to minimize prediction error across the entire dataset. In any training set, the majority of transactions involve white homeowners, because white households own approximately 72 percent of owner-occupied homes in the United States. An algorithm optimized for overall accuracy will, without explicit correction, perform best for the population it sees most and worst for the populations it sees least. As the Urban Institute researchers put it: "Without built-in fairness constraints, predictive accuracy tends to come at the cost of equity."
Both mechanisms produce the same result, neither uses race as an input, and neither is addressed by a rule that says "comply with nondiscrimination laws" without specifying what compliance looks like when the discrimination is structural rather than intentional, baked into training data rather than written into code, invisible to the model and often invisible to the humans who deploy it.
Strongest Case Against This Finding
Not everyone agrees that AVMs are the villain here. Veros Real Estate Solutions studied 50 Chicago ZIP codes in 2021 and found no evidence of racial bias in their proprietary AVM, concluding that the proportion of undervalued properties was not correlated with neighborhood racial composition. Freddie Mac's own researchers and the Urban Institute's earlier work have argued that AVMs could reduce racial bias relative to human appraisers, who bring their own conscious and unconscious prejudices into every inspection. Freddie Mac data showed 12.5 percent of properties in Black neighborhoods received human appraisals below contract price, compared with 7.4 percent in white neighborhoods, an uncomfortable number for anyone defending the status quo of in-person valuations.
This is a legitimate argument, and it deserves a precise response. But the real question is not whether AVMs are worse than human appraisers. It is whether AVMs are fair. A tool that produces 5 percent systematic undervaluation for Black homeowners is not exonerated by the fact that the tool it replaced was also biased. Veros examined neighborhoods, not individuals, and used a single proprietary AVM in a single city. By contrast, the Urban Institute used individual-level data with race imputation validated against HMDA records, controlled for property condition through computer vision, and found the gap persists within the same neighborhoods, which is a fundamentally different level of rigor answering a fundamentally different question about whether individual homeowners are treated equitably regardless of race.
What This Means for Your Home
If you are buying, selling, or refinancing a home in 2026, there is roughly a one-in-four chance your home's value will be determined partly or entirely by an AVM, and that probability is rising every quarter as the GSEs expand waiver eligibility, the appraiser workforce contracts, and the economics of algorithmic valuation overwhelm the economics of sending a human being to your front door. It will never see your kitchen renovation, never notice the water stain on the basement ceiling, never register that the neighbors' house sold below market because of a divorce. It will pull comparable sales from a database shaped by a century of housing policy, run them through a model optimized for aggregate accuracy, and produce a number. If you are a Black homeowner, the Urban Institute's data suggests that number will be, on average, 5 percent lower than it would be for a white homeowner with a comparable property in the same neighborhood.
On a $420,000 home, that is $21,000 in missing equity, and over a thirty-year mortgage, it compounds into constrained refinancing options, reduced borrowing capacity, and a measurable drag on intergenerational wealth that the homeowner never sees because the algorithm's output arrives as a single number with no explanation, no margin of error, and no appeal process.
Your federal AVM rule gives you a right that sounds meaningful: your AVM must comply with nondiscrimination laws. But unlike a human appraisal, where you can request a reconsideration of value by identifying specific comparable sales the appraiser overlooked, there is no established process for challenging an algorithmic valuation. You cannot ask the model to show its work, because the model's work is a matrix of coefficients trained on millions of transactions that no human being fully understands. Lenders are not required to tell you that an AVM was used, not required to disclose which AVM, and not required to publish the AVM developer's performance metrics by race.
Nine months after the rule took effect, the enforcement infrastructure is the same as it was before: a homeowner who believes their AVM-generated valuation was discriminatory can file a complaint with the CFPB or pursue a Fair Housing Act claim, a process that requires proving discrimination without access to the algorithm, its training data, or its performance metrics.
What Would Actually Fix This
The Urban Institute recommends several interventions, disaggregated by the source of bias. For data-driven bias: require AVM developers to publish accuracy metrics for their underlying datasets, standardize transparency requirements across vendors, improve data coverage in historically underserved neighborhoods, and scrutinize the use of comparable sales from areas where values were depressed by discriminatory lending and appraisal practices. For algorithmic bias: add fairness as an explicit constraint during model training (accept a small reduction in aggregate accuracy to achieve equitable error rates across racial groups), mandate public reporting of AVM performance by race and geography, and create clear regulatory guidelines for evaluation.
Maryland has established a Task Force on Property Appraisal and Valuation Equity that recommends transparency improvements and keeping appraisers in the loop for automated analyses. California's Civil Rights Department secured financial settlements from the firm Clear Capital after demonstrating appraisal bias against a Black and Latino family. State-level action is happening, though it varies enormously in scope and enforcement capacity.
At the federal level, the landscape shifted on Inauguration Day 2025, when the current administration revoked a 2023 executive order that had required AI developers to share safety test results with the government. Because the interagency AVM rule implements a Dodd-Frank mandate and not an executive order, it survived. But the appetite for aggressive AI fairness regulation at the federal level has, by most accounts, cooled.
Limitations
This study analyzes a single year of data (2018) in two metro areas (Atlanta and Memphis) with predominantly Black and white populations. Results may not generalize to markets with larger Hispanic or Asian American populations, where AVM dynamics could differ, and the fBISG race imputation, while performing well for Black and white homeowners, is less reliable for smaller groups. Only one AVM from a major property records provider was tested; others may perform differently. AEI's appraisal waiver data tracks whether waivers were used, not whether the AVM underlying each waiver produced equitable outcomes by race. That 5 percent undervaluation figure is a regression coefficient, not a universal constant; individual outcomes vary widely around the mean. And because the 2018 data predates the October 2025 rule, this study does not test whether the rule has improved AVM performance, though given the rule's lack of specific requirements, the absence of evidence is not surprising.