Thirty Percent of Appraisers Valued Your Home at Exactly the Number They Were Told. The Algorithm Wasn't Told.
In 2021, a couple in Jacksonville, Florida, received an appraisal of $330,000 on their home, a number they believed drastically understated its value. So they tried something that should not have worked: the Black woman who co-owned the home removed her family photos, hid her Toni Morrison books, and had her white husband answer the door for the second appraisal. The new number came back at $465,000, a $135,000 swing triggered by nothing more than which face answered the door. The New York Times published the story, the Appraisal Institute issued a statement acknowledging that "unconscious bias is real and exists in all industries," and then nothing structurally changed about how human appraisers do their jobs.
What changed instead was the regulation of algorithms.
On October 1, 2025, a rule jointly issued by six federal agencies took effect, imposing quality control standards on automated valuation models, the algorithmic systems that companies like Zillow and Redfin use to estimate your home's worth. The rule requires mortgage originators and secondary market issuers using AVMs to adopt policies ensuring high confidence in estimates, protection against data manipulation, avoidance of conflicts of interest, random sample testing, and compliance with nondiscrimination laws. It is a reasonable rule that addresses real risks, and it regulates the system that, by nearly every available metric, is already more accurate and less biased than the one it supplements.
The Number That Should Embarrass an Industry
Thirty percent of human home appraisals land at exactly the agreed-upon contract price between buyer and seller. Not approximately, not within a reasonable margin, but exactly on it, down to the dollar. In rural areas, the number is worse: ninety percent of appraisals confirm or exceed the contract price when the appraiser knows what that price is.
These figures come from research cited by the Brookings Institution and the Federal Housing Finance Agency, and they describe a system where the person hired to independently verify a home's value is, in roughly one-third of cases, producing the number that makes the deal close rather than the number the evidence supports. Researchers at the FHFA confirmed what this means in practice: mortgages where the appraisal matches the exact contract price are more likely to default, because the appraiser's rubber stamp inflated the collateral value, the lender extended credit based on that inflation, and the borrower got a mortgage their home could not actually support.
Compare that to what the algorithms produce: Zillow's Zestimate carries a median error rate of 1.83 percent for on-market homes, and Redfin's estimate comes in at 1.99 percent, both figures that suggest the machines are actually generating independent assessments rather than confirming a number somebody handed them. When Fitch Ratings tested seven AVM providers, all seven predicted within ten percent of the actual sale price for at least 95 percent of the properties in their test set, and the FHFA found that in rural areas, where human appraisers are scarcest and most stretched, AVMs actually outperform human appraisals on accuracy.
Off-market, the picture is less flattering for algorithms, with Zillow's error rate climbing to 7.01 percent and Redfin's to 7.67 percent. For a $500,000 home, a seven percent error is $35,000, which is not nothing, but a human appraiser who knows the contract price and confirms it anyway is operating in a different category of failure entirely. That appraiser is not making an error in the statistical sense; the appraiser is producing a biased estimate by design, one that serves the transaction rather than the truth, and the system has tolerated it because it makes deals close.
Five Standards for Algorithms. Zero for the Person with the Clipboard.
Published in the Federal Register on August 7, 2024, and effective since last October, the AVM rule requires covered institutions to meet five quality control factors. I practiced construction law for five years before covering policy, so I read the full rule, and it is short, principled, and deliberately flexible in ways that suggest the agencies understood the technology would outpace any prescriptive requirements they tried to lock in. They left implementation to the institutions themselves, reasoning that a "principles-based" approach would adapt as technology evolves.
But here is the exercise nobody at the CFPB appears to have performed publicly: grading human appraisers against those same five standards.
Standard 1: Ensure a high level of confidence in estimates. When thirty percent of human appraisals reproduce the contract price exactly, that is not high confidence in an independent assessment; it is confirmation dressed as valuation, and an algorithm that hit the input price thirty percent of the time would trigger an immediate audit.
Standard 2: Protect against data manipulation. In the appraisal context, the contract price functions as an input that manipulates the output, because appraisers who know the target number are demonstrably influenced by it, and the Brookings data on rural areas, where ninety percent of appraisals confirm the price, shows an industry where the supposed independent check has collapsed into ceremonial confirmation. An AVM that operated this way would be pulled from production.
Standard 3: Avoid conflicts of interest. Appraisers are typically paid through channels connected to the party that needs the deal to close. Not always the buyer directly, and reforms after the 2008 crisis introduced appraisal management companies to create a buffer, but the fundamental economic incentive has not changed: low appraisals kill deals, dead deals mean unhappy clients and fewer referrals, and the Dodd-Frank Act addressed this partially without eliminating the structural problem. An AVM owned by the lender that uses it would face scrutiny under this standard, yet the human appraiser doing essentially the same thing does not.
Standard 4: Require random sample testing and reviews. No federal agency systematically tests the accuracy of individual human appraisers at scale, because the Appraisal Foundation sets standards and state boards handle licensing and complaints, but nobody performs the equivalent of the random-sample backtesting that the AVM rule now demands. After Freddie Mac examined 12 million appraisals and found persistent racial disparities, the response was a task force rather than a testing regime.
Standard 5: Comply with nondiscrimination laws. According to Freddie Mac's analysis of those 12 million purchase-transaction appraisals between 2015 and 2020, homes in majority-Black census tracts were appraised below the contract price 12.5 percent of the time, compared with 15.4 percent in majority-Latino tracts and 7.4 percent in majority-white tracts. Bureau of Labor Statistics data shows that 98 percent of home appraisers are white, and the Brookings Institution calculated the total cost of devaluation across 113 metro areas with at least one majority-Black neighborhood at approximately $162 billion.
Human appraisers demonstrably fail four of the five standards the federal government just imposed on algorithms. They arguably fail all five.
The Counterargument That Actually Matters
An algorithm cannot see inside your home, and that limitation is not trivial. No AVM can notice the water stain spreading across the basement ceiling, the unpermitted addition that added a bedroom but violated setback requirements, or the foundation crack your seller concealed with a strategically placed bookshelf. A Zillow Zestimate processes tax records, MLS data, and comparable sales, but it does not process the smell of mold in the crawl space, and interior condition drives enormous differences between apparently identical properties.
A 2,200-square-foot ranch in Dunwoody, Georgia, might be worth $460,000 with original 1985 bathrooms or $530,000 with a recent renovation, and no algorithm can bridge that $70,000 gap from the outside. AVM providers know it: Zillow explicitly states that Zestimates "are not appraisals" and should be used as "a starting point."
A stronger version of this argument holds that we need both systems working together, that the ideal is an AI estimate checked by a human inspector who adds what the data cannot see, and that hybrid model probably is better than either alone. But the current system is not a hybrid model in any meaningful sense. What exists treats the human appraiser as the authoritative source and the algorithm as the supplementary one, and the new rule codifies that hierarchy by regulating the supplement more aggressively than the authority.
What This Means If You Are Buying or Selling a Home
First, understand what changed on October 1, 2025: if your mortgage lender uses an automated valuation model as part of the lending decision, that lender is now required to have quality control policies covering the five factors described above. No specific policies are prescribed, and no private right of action exists, meaning you cannot sue your lender for violating the rule directly, but it creates a compliance obligation that federal examiners can enforce, and the nondiscrimination standard, which did not exist before, gives regulators a new tool for challenging AVMs that produce racially disparate results.
Second, know your rights around human appraisals, because they are broader than most buyers realize. Under the Equal Credit Opportunity Act, you are entitled to a copy of any appraisal performed in connection with your mortgage application, and if the appraisal comes in low, you can request a reconsideration of value from your lender, provide additional comparable sales data, and in some cases request a second appraisal. Several states, including California and Maryland, have passed laws creating additional protections against discriminatory appraisals, including the right to have family photos and personal effects present during the appraisal without penalty.
Third, treat consumer AVMs for what they are: useful starting points with known limitations that vary dramatically depending on whether the home is listed or not. A 1.8 to 2 percent median error rate on listed homes means the Zestimate on a $500,000 listing is likely within $9,000 to $10,000 of the sale price, which is genuinely helpful for comparison shopping, but the 7 percent error rate on off-market homes means the Zestimate on your unlisted home could be off by $35,000 in either direction, which is less helpful and potentially dangerous if you use it to set your sale price without a comparative market analysis from a broker who has actually walked through the property.
Fourth, and this is the uncomfortable recommendation: if you are selling a home in a neighborhood where the data shows appraisals run low, consider asking your lender whether an AVM was used and whether the result differed significantly from the human appraisal, because the AVM rule's nondiscrimination standard creates a new data point worth leveraging. If the algorithm valued your home significantly higher than the appraiser did, that gap is worth documenting and worth raising in a reconsideration request.
What the Rule Should Have Done
The AVM rule is not a bad rule. It addresses genuine risks around algorithmic opacity, data quality, and discrimination in automated systems. The nondiscrimination standard, in particular, represents the first time federal regulators have explicitly required that property valuation algorithms be tested for compliance with fair lending laws. That matters.
But the regulatory asymmetry is hard to justify on the merits. We now have a federal rule requiring random sample testing and conflict-of-interest controls for algorithms while human appraisers continue to operate under a system where thirty percent of them reproduce the number they were given, where 98 percent of them are white, and where the only systematic study of their accuracy, covering 12 million transactions, found persistent racial disparities that nobody has fixed.
Fourteen years passed between the Dodd-Frank authorization and the final rule, and another year elapsed before it took effect. During all fifteen of those years, the human appraisal system that the rule implicitly trusts more has been producing the same documented biases, the same confirmation of known contract prices, the same racial disparities that cost majority-Black neighborhoods $162 billion in lost home value. If the five quality control standards are the right framework for evaluating property valuations, and I believe they are, then applying them to algorithms and not to humans is not regulation in any serious sense of the word. It is theater. The machine got audited. The appraiser got a clipboard and a key.
Limitations of This Analysis
Most AVM accuracy data cited here is self-reported by the providers, because Zillow publishes its own error rates, Redfin publishes its own error rates, and independent verification remains limited. Fitch Ratings tested providers against actual sale prices, which is the strongest external validation available, but the sample and methodology were not published in full detail.
Brookings' appraisal bias data measures appraisals below contract price, which captures one form of undervaluation but may miss others, because an appraisal that meets the contract price in a neighborhood where comparable sales support a higher value is also an undervaluation that would not appear in this metric.
Nine months have passed since the CFPB rule took effect, and no compliance data has been published, making it too early to know whether the rule will produce measurable changes in AVM accuracy or bias, particularly given the current administration's approach to CFPB enforcement priorities.
Finally, the 30 percent figure for contract-price confirmation is drawn from research analyzing large datasets across multiple years, but the denominator and methodology vary across the studies that cite it, which means the directional finding, that appraisers systematically confirm known contract prices, is well-established while the precise percentage carries uncertainty that honest analysis should acknowledge.