Appraisal Bias and AVMs: What Triggers the Valuation Gap

Who This Risk Hits, and What Actually Triggers It

Appraisal bias risk activates the moment a value opinion enters a lending decision, which is almost every time a home changes hands or a homeowner tries to tap equity. A purchase appraisal comes in below the agreed sale price and the deal wobbles unless the buyer produces more cash or the seller cuts the price. A refinance appraisal comes in low and the loan-to-value ratio breaches the lender’s ceiling, killing a cash-out refi or a rate-and-term deal that would have lowered a monthly payment. A home equity line gets capped below what the homeowner expected because the automated collateral evaluation behind it pulled comparable sales from a stretch of the neighborhood that undersells the block.

Three groups carry most of the exposure. Homeowners and buyers in majority-Black, majority-Latino, and mixed-race neighborhoods face the risk most directly, because that is where the documented valuation gap concentrates. Lenders and appraisal management companies (AMCs) carry compliance and reputational risk any time their assigned appraiser or their AVM vendor produces a value that a fair-lending examiner or a plaintiff’s attorney can later show correlates with neighborhood racial composition rather than property characteristics. And mortgage technology vendors building or licensing AVMs carry model risk, because a scoring engine trained on a historically skewed comparable-sales dataset will keep producing skewed outputs long after anyone remembers which comps seeded the bias.

The trigger is rarely a single bad actor making an explicit decision. It is usually a chain of smaller, defensible-looking choices — which three or six comparable sales an appraiser selects from a pool of dozens, how much weight an algorithm assigns to a “condition” adjustment, which zip code boundary an AVM uses to define its comp radius — that compound into a gap large enough to change whether a loan closes.

Human Appraiser Bias: Comparable Selection and the Subjectivity Problem

A traditional appraisal is, underneath the formal language, a judgment call dressed up in a standardized form. The appraiser walks the property, picks a handful of “comparable” recent sales from the surrounding market, and then adjusts each comp up or down for differences in size, condition, lot, and amenities until the adjusted comps cluster around a final value. Every one of those steps — which comps to include, which to exclude, how many dollars a renovated kitchen is worth, how far outside the immediate block to search for comps — is left to professional judgment, and professional judgment is exactly where bias tends to hide.

Freddie Mac’s 2021 research into racial and ethnic valuation gaps found that purchase appraisals came in below the negotiated contract price in roughly 7.4% of transactions in majority-white census tracts, compared with roughly 12.5% in majority-Black tracts and roughly 15.4% in majority-Latino tracts. That is not a rounding error. It means a buyer or seller in a majority-Latino neighborhood was more than twice as likely to hit an appraisal gap that could sink financing, delay closing, or force a price renegotiation, purely as a function of neighborhood composition rather than the physical condition of the home.

Three mechanics drive most of the gap. The first is comp geography: appraisers working under time pressure from an AMC often pull comps from the nearest sales that meet the software’s automatic search radius, and because housing markets remain heavily segregated by neighborhood, that radius can quietly exclude higher-selling blocks a few streets over while including lower-selling ones. The second is subjective condition and “curb appeal” scoring, where identical renovations get rated differently depending on unstated assumptions about what a neighborhood is “supposed” to look like. The third, documented in multiple fair-housing complaints and at least one federal lawsuit, is direct racial cue-taking — an appraiser (or a homeowner’s own experience trying to influence the appraiser) noticing family photos, religious items, or other markers of the occupants’ race and adjusting the value opinion accordingly, consciously or not.

None of this requires an appraiser to hold explicit racial animus. A national study cited by the Brookings Institution and later folded into federal task force findings found that homes in majority-Black neighborhoods are worth roughly 23% less than comparably built and renovated homes in neighborhoods with few or no Black residents, a devaluation the study’s authors attributed largely to the racial composition itself rather than to school quality, crime, or amenities. When that historical devaluation becomes the comp pool, each new appraisal inherits it.

The Uniform Appraisal Dataset and the Language Problem

Appraisal reports are not just numbers. They include a narrative section, and researchers combing through Uniform Appraisal Dataset (UAD) submissions have repeatedly found subjective language — references to a neighborhood being “up and coming,” “transitional,” or “desirable” for a specific demographic — that correlates with lower valuations in diverse tracts. Fannie Mae’s Collateral Underwriter tool now runs natural-language processing across submitted reports specifically to flag potentially biased or discriminatory language before a loan closes, which is itself an admission of how common the pattern was.

Automated Valuation Models: How Algorithmic Bias Gets Inherited, Not Eliminated

The mortgage industry’s answer to slow, expensive, subjective human appraisals has been to lean harder on Automated Valuation Models — statistical or machine-learning engines that estimate a property’s value from public records, tax assessments, prior sales, and sometimes photos, without a human ever walking through the door. Fannie Mae’s value acceptance program and Freddie Mac’s Automated Collateral Evaluation both let eligible loans skip a traditional appraisal entirely, relying instead on an AVM-generated value plus a risk score. Roughly one in five conventional purchase and refinance loans now closes with some form of appraisal waiver rather than a full traditional appraisal, a share that has grown steadily since the GSEs expanded eligibility in the early 2020s.

AVMs solve the “cue-taking” problem — a model never sees a family photo on a mantel — but they do not solve the underlying data problem, and in some respects they make it worse by hiding it behind a numeric confidence score that looks objective. Every AVM is trained on a historical record of comparable sales and, for many models, historical appraisal values themselves. If that training data carries the same neighborhood-level undervaluation documented in the Freddie Mac research, the model learns to reproduce it with mathematical consistency rather than human inconsistency. A biased human appraiser might occasionally overshoot in the other direction; a biased algorithm applies the same discount every time, to every comparable property, at scale, without anyone in the loop who could notice the pattern and push back.

This is sometimes called a feedback loop, and it is worth walking through concretely. An AVM estimates a home’s value using nearby sale prices. Those sale prices were themselves partly shaped by past appraisals, which were shaped by past comps, which were shaped by valuations going back through decades of formal redlining and informal steering. Every layer of that chain compresses value in the same direction in the same neighborhoods. A model can be statistically well-calibrated — meaning its average error is small across the whole market — while still being systematically wrong in one direction for a specific subset of neighborhoods, and “well-calibrated on average” is exactly the metric most AVM accuracy testing has historically emphasized. Lenders evaluating a model’s overall forecast standard deviation can miss a localized, racially patterned error entirely.

Lenders who lean on automated underwriting for these decisions face a related, less-discussed exposure: when an AVM waives the appraisal, there is no independent human check left to catch the very pattern regulators are worried about. Our guide to automated underwriting for complex mortgages walks through how automated collateral evaluation feeds into the broader underwriting decision and where lenders still insert manual review for higher-risk files.

The Regulatory Patchwork: What the 2024 Interagency AVM Rule Actually Requires

Congress gave regulators the authority to police AVM quality back in the 2010 Dodd-Frank Act, under a provision aimed at the automated valuation tools used in mortgage origination and securitization. It took until 2024 for six agencies — the Office of the Comptroller of the Currency, the Federal Reserve, the FDIC, the National Credit Union Administration, the Consumer Financial Protection Bureau, and the Federal Housing Finance Agency — to jointly finalize the implementing rule. The rule became effective October 1, 2024, and it does not ban AVMs or set a numeric bias threshold. Instead it requires any institution using AVMs in connection with mortgage origination or securitization to adopt policies and procedures reasonably designed to satisfy four quality-control objectives: ensuring a high level of confidence in the value estimates produced, protecting against the manipulation of data used to produce an estimate, avoiding conflicts of interest, and requiring random sample testing and reviews of the AVMs used. A fifth, cross-cutting requirement folds in existing nondiscrimination law, directing institutions to comply with applicable provisions of the Equal Credit Opportunity Act and the Fair Housing Act when they rely on an AVM.

That last piece is where the rule actually bites on bias, because it means a lender cannot treat “the model said so” as a legal shield. If an AVM systematically undervalues homes in majority-minority tracts and a lender keeps using it without testing for that pattern, the lender — not just the AVM vendor — carries fair-lending exposure. The rule leaves the specific testing methodology largely up to each institution, which examiners and consumer advocates have both flagged as a gap: a bank can satisfy the letter of the rule with a testing program that checks overall accuracy without ever segmenting results by neighborhood racial composition, and a compliant-on-paper program can still miss a real, measurable gap.

The rule sits alongside, rather than replaces, earlier action. The interagency Property Appraisal and Valuation Equity (PAVE) task force, convened in 2021 under HUD and the White House Domestic Policy Council, released a 2022 action plan with more than 20 commitments across agencies, including funding for a more diverse appraiser workforce, a modernized Uniform Appraisal Dataset with tighter guardrails against biased narrative language, and a requirement that Fannie Mae and Freddie Mac build a formal Reconsideration of Value process into every loan they purchase. Separately, HUD in 2024 issued formal guidance clarifying that an unlawfully biased appraisal can itself constitute a Fair Housing Act violation, and that lenders who knowingly rely on a discriminatory appraisal without offering a path to challenge it can share in that liability.

Real-World Consequence: A $487,500 Reappraisal Gap in Marin City

The clearest illustration of how large this gap can run in a single transaction comes from a widely reported 2021 case out of Marin City, California. A Black couple seeking to refinance their home received an appraisal of $995,000. Skeptical that the number reflected their home’s actual condition and the surrounding market, they removed family photographs and African art, put away anything that might signal the occupants’ race, and had a white male colleague stand in for the follow-up walkthrough while they stayed away. A second appraisal on the same property, weeks later, came back at $1,482,500 — a difference of $487,500, or roughly 49% higher than the first opinion of value on an unchanged house.

The gap was large enough to become the basis of a federal fair-housing lawsuit alleging that the first appraisal violated the Fair Housing Act and the Equal Credit Opportunity Act. Whatever a court ultimately finds about intent in any single case, the arithmetic of the example is the useful part for anyone assessing this risk: identical square footage, identical lot, identical renovations, and a swing of nearly half a million dollars tied to nothing the underlying comparable sales data should have changed between the two visits. On a refinance, a gap of that size does not just cost the homeowner a bruised ego — it changes the loan-to-value ratio, changes mortgage insurance requirements, changes whether cash-out proceeds clear underwriting thresholds, and can be the difference between a loan closing and a loan being declined outright.

Red Flags: A Quick-Reference Table

Not every low appraisal reflects bias, and not every high AVM confidence score reflects accuracy. The signals below are not proof on their own, but each one is a documented pattern worth investigating before you accept a valuation as final.

How to Respond: A Practical Checklist

Whether you are a homeowner staring at a disappointing number or a loan officer trying to keep a file compliant, the response playbook is largely the same: act inside the window, document everything, and use the formal channels regulators built specifically for this problem.

• Get the full report, not just the number. Request the complete appraisal, including the comparable sales grid, adjustment amounts, and any narrative comments, before you agree to any price renegotiation.
• File a Reconsideration of Value immediately. Fannie Mae and Freddie Mac both require sellers and lenders to accept ROV requests on conventional loans; most lenders set a short window, often just a few business days before closing, so move fast.
• Bring your own comps. A strong ROV includes at least two or three comparable sales the original report omitted, ideally closer in distance and more similar in condition than the ones the appraiser used.
• Ask what valuation method was used. A full appraisal, a desktop appraisal, a hybrid product, and an AVM-based value acceptance each carry different challenge processes; know which one you’re contesting.
• Request the AVM vendor and confidence tier if a model set the value. Under the 2024 interagency rule, lenders should be able to explain how the model was tested; a lender that cannot answer this has a compliance gap worth flagging.
• Document any request to alter your home’s appearance. If an appraiser or data collector asks you to remove personal items before a visit, save the message or note the date and wording immediately.
• Escalate to your state appraiser licensing board if the report itself contains biased language or clearly mismatched comps, since state boards can investigate individual appraiser conduct.
• File with HUD or the CFPB if you believe the valuation reflects discrimination; HUD’s Office of Fair Housing and Equal Opportunity accepts complaints tied to appraisal bias, and the PAVE task force built a dedicated public reporting channel for exactly this purpose.
• Consider a second independent appraisal paid out of pocket if the ROV is denied and the stakes justify the cost, particularly on a refinance where the LTV threshold is close to a cutoff.
• Keep a paper trail across every step — the original report, your ROV submission, the lender’s response, and any correspondence — since this same documentation becomes the evidentiary basis for a licensing complaint or a fair-housing claim if informal channels fail.

Key Takeaways

• Purchase appraisals fall below the contract price roughly twice as often in majority-Black and majority-Latino census tracts as in majority-white tracts, according to Freddie Mac’s 2021 research.
• Automated Valuation Models remove face-to-face cue-taking but can still encode the same neighborhood-level undervaluation because they train on historical sales and appraisal data shaped by decades of segregation and steering.
• The interagency AVM Quality Control Rule, effective October 1, 2024, requires lenders using AVMs in origination or securitization to test for accuracy, guard against manipulation, and comply with the Equal Credit Opportunity Act and Fair Housing Act — but leaves the specific bias-testing method up to each institution.
• A formal Reconsideration of Value request, filed quickly with alternative comps attached, is the single most effective and accessible tool available to a borrower who believes an appraisal undervalued their home.
• Roughly one in five conventional loans now closes on an AVM-driven value acceptance rather than a traditional appraisal, which means the algorithmic version of this risk now touches more transactions than the human version.

Frequently Asked Questions

What is appraisal bias in mortgage lending?

Appraisal bias is a pattern in which a home’s estimated value is systematically lower than its true market value because of the race or ethnicity of its occupants or the racial composition of its neighborhood, rather than because of the property’s actual size, condition, or location fundamentals. Federal research has documented this pattern at the neighborhood level using large samples of purchase transactions, not just isolated anecdotes.

How do Automated Valuation Models reduce or worsen bias?

AVMs reduce the risk of an appraiser reacting to visible cues like family photos or décor, since a model never enters the home. They can worsen bias in a different way: because they are trained on historical sales and appraisal data that already reflect segregation-era undervaluation, they can apply that same discount consistently across every property in an affected neighborhood, at a scale no individual human appraiser could match.

What is the interagency AVM Quality Control Rule and when does it apply?

It is a joint rule finalized by the OCC, Federal Reserve, FDIC, NCUA, CFPB, and FHFA, effective October 1, 2024, implementing a Dodd-Frank Act requirement. It applies whenever a mortgage originator or securitizer uses an AVM to determine collateral value, and it requires policies ensuring high confidence in value estimates, protection against data manipulation, avoidance of conflicts of interest, random sample testing, and compliance with the Equal Credit Opportunity Act and Fair Housing Act.

What is a Reconsideration of Value and how do I request one?

A Reconsideration of Value, or ROV, is a formal request asking a lender to have an appraisal reviewed, typically because the homeowner or their agent believes the report used flawed comparable sales, contains factual errors, or reflects discriminatory assumptions. You request one directly through your loan officer or lender, generally within a few business days of receiving the appraisal, and a strong request includes specific alternative comparable sales the original report omitted.

Can I take legal action if I believe my appraisal was biased?

Yes. A biased appraisal can support a claim under the Fair Housing Act or the Equal Credit Opportunity Act, and HUD’s Office of Fair Housing and Equal Opportunity accepts complaints specifically about appraisal discrimination. Several borrowers have filed federal lawsuits against individual appraisers and, in some cases, the lenders who relied on their reports without offering a meaningful challenge process.

Do AVMs replace human appraisers entirely?

Not entirely. Roughly one in five conventional loans now closes with an AVM-based value acceptance instead of a traditional appraisal, but higher-risk transactions — jumbo loans, unusual properties, higher loan-to-value purchases — still require a human appraiser, and regulators have kept a role for manual desk review even inside heavily automated underwriting workflows.

References

• Freddie Mac, “Racial and Ethnic Valuation Gaps in Home Purchase Appraisals” (2021)
• U.S. Department of Housing and Urban Development, Property Appraisal and Valuation Equity (PAVE) Task Force Action Plan (2022)
• Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency, FDIC, NCUA, CFPB, and FHFA, Interagency Final Rule on Quality Control Standards for Automated Valuation Models (2024)
• Consumer Financial Protection Bureau, guidance on Reconsiderations of Value in mortgage lending
• Brookings Institution, research on the devaluation of homes in majority-Black neighborhoods
• Fannie Mae, Collateral Underwriter documentation on appraisal risk scoring and narrative-language review
• U.S. Department of Housing and Urban Development, Office of Fair Housing and Equal Opportunity, guidance on appraisal discrimination and Fair Housing Act liability (2024)