BRENNEN

Industries / Construction and property / Real estate / Appraisal

Bias and review

Bias and review is the checking of appraisals after they are written, by a second appraiser, by lenders' and loan buyers' software, and by owners who ask for a reconsideration of value, and the evidence and law on whether values are pulled by race or by other forces, such as the agreed sale price.

6

Milestones achieved, from 1977 to 2025.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1977

    Court ends race-based guidance

    The Federal Housing Administration's 1938 Underwriting Manual told appraisers that "A change in social or racial occupancy generally contributes to instability and a decline in values", and in 1973 the American Institute of Real Estate Appraisers' course material still said that "homogeneity of the population contributes to stability of real estate values". In 1976 the Justice Department sued the institute and three other defendants under the Fair Housing Act of 1968. Under the 1977 settlement it became improper to base a value on the premise that racial, ethnic or religious homogeneity "is necessary for maximum value".

    LessonThe Fair Housing Act did not change the textbooks by itself; a lawsuit nine years later did.

    nationalfairhousing.org

    Credit

    US Department of Justice; US District Court for the Northern District of Illinois (United States v. American Institute of Real Estate Appraisers, 442 F. Supp. 1072)

  2. 2019

    Dutch values match the price

    The Dutch central bank, De Nederlandsche Bank (DNB), looked at more than 200,000 Dutch home appraisals from 2012 to 2017: a third came in exactly at the purchase price, and just over 5% below it. The US studies of the same pull, by Fannie Mae economists, are in the Mortgage lending file.

    LessonA review that only asks whether a value is supported can miss a value pulled toward the price.

    dnb.nl

    Credit

    Remco van der Molen and Rob Nijskens, DNB (March 2019)

  3. 2021

    Software reads every report

    In 2021 Fannie Mae began scanning the appraisal reports it receives for prohibited and subjective language; by its figures, as reported by HousingWire, about 79% of the appraisers sent a warning letter in 2021 had no new findings in 2022. Freddie Mac turned its warning on subjective or possibly biased wording into a "fatal" message, code FRE 4611, which was live by January 2024.

    LessonThe scans test the wording of a report, not whether its value is fair.

    housingwire.comacesquality.com

    Credit

    Fannie Mae; Freddie Mac

  4. 2023

    Appraisal records go public

    On 16 October 2023 FHFA released the first public file of individual appraisal records, a 5% random sample of the appraisals behind Fannie Mae and Freddie Mac loans, with 1.37 million records and 37 fields; version 2.1, of 4 February 2025, removed the field for borrowers' race and ethnicity. Readers disagree on what the data show. The AEI Housing Center said in November 2023 that the gaps by race are "almost entirely or entirely due to differences in geographies". Brookings researchers, using about 46 million appraisals from the full dataset, found homes in majority-Black neighborhoods 1.9 times as likely to appraise below the contract price after controls, and estimated that appraisal bias explains 9% to 19% of those homes' lower values.

    fhfa.govvaluationreview.combrookings.edu

    Credit

    Federal Housing Finance Agency

  5. 2024

    Rules for contesting a value

    A reconsideration of value (ROV) is a lender's request that the appraiser look again at a value, using information the borrower says was missed. On 18 July 2024 five US agencies, the Consumer Financial Protection Bureau (CFPB), the Federal Deposit Insurance Corporation (FDIC), the Federal Reserve, the National Credit Union Administration (NCUA) and the Office of the Comptroller of the Currency (OCC), issued final guidance on ROVs, and Fannie Mae and Freddie Mac set their own rules for borrower-requested ROVs. From 3 September 2025 the two companies trimmed theirs, dropping a "dual disclosure" requirement and no longer requiring lenders to record that an ROV was started.

    LessonRules set by guidance and by the companies' guides can be loosened as quickly as they were written.

    federalreserve.govthewbkfirm.com

    Credit

    CFPB, FDIC, Federal Reserve Board, NCUA and OCC; Fannie Mae; Freddie Mac

  6. 2025

    Federal push reversed

    The federal task force on Property Appraisal and Valuation Equity (PAVE) was set up on 1 June 2021. On 10 July 2025 the Department of Housing and Urban Development (HUD) and the Office of Management and Budget ended its core policies, calling them "burdensome", and withdrew the FHA letters on appraisal review and ROVs (ML 2024-07) and on fair housing in appraisals (ML 2021-27). On 17 July 2025 a federal judge in Maryland ruled for the appraiser in a Baltimore case in which the homeowners alleged racial discrimination: the appraiser's firm had valued their home at $472,000 and, seven months later, at $750,000 by a second appraiser, who later said he had overvalued it by about $50,000; the judge wrote that statements in the complaint were ones "Plaintiffs failed to prove".

    LessonAgency letters can be withdrawn by agency letters, while the fair lending laws stay in force; and a gap between two appraisals did not by itself prove bias in court.

    consumerfinancialserviceslawmonitor.comarchives.hud.govhousingwire.com

    Credit

    HUD and OMB; Judge Stephanie Gallagher, US District Court for the District of Maryland

Watch for

  1. Contesting a value becomes law

    Section 704 of the 21st Century ROAD to Housing Act requires the FHA, FHFA, the Department of Agriculture and Veterans Affairs to "implement and maintain procedures for consumer-initiated reconsiderations of value". Congress passed the Act on 24 June 2026, and it was enacted on 11 July 2026.

    Achieved whenAll four agencies have issued ROV procedures that lenders must follow, and they are in effect.

    classvaluation.comnclc.org

    Claimed by

    US Congress, June and July 2026, as summarized by Class Valuation, an appraisal management company

  2. A public appraisal database, studied

    The same Act tells the GAO to "complete, within 240 days, a feasibility study on a publicly accessible appraisal-level database", which puts the due date around March 2027.

    Achieved whenThe GAO publishes the study.

    classvaluation.com

    Claimed by

    US Congress, July 2026

  3. Fewer rules, lower costs

    HUD and OMB said the PAVE-era rules were "burdensome policies" that raised costs and limited access to home ownership, and argued that education and credit scores drive appraisal outcomes more than race does.

    Achieved whenFHA or independent data show lower appraisal costs or faster turnaround after July 2025, with no rise in below-contract appraisals in neighborhoods where most residents are Black or Latino.

    consumerfinancialserviceslawmonitor.com

    Claimed by

    HUD and OMB, July 2025

Open

  1. 01

    Open appraisal-gap replication

    A public code repository that rebuilds the appraisal gap from FHFA's appraisal-level file (version 2.1) joined to Census tract data on race, for 2013 to 2022, and then asks how much of it survives controls. It reports the share of purchase appraisals below the contract price by the tract's minority share, first raw, then within county and year, then after controls for price band and the condition and quality ratings, and the share landing exactly at the contract price, DNB's measure of the pull toward the price.

    The brief
    Credit · unclaimed
  2. 02

    A test for bias scanners

    A public set of about 2,000 made-up sentences of the kind written in the neighborhood and site sections of appraisal reports, each labeled by two people as acceptable, subjective or prohibited under the guidance on unacceptable appraisal practices that Fannie Mae and Freddie Mac publish. It includes traps both ways: coded wording that names no group, and fair sentences that a word list would flag, such as an address on Desirable Lane. Word-list scanners and AI models are scored on what they catch and what they flag wrongly.

    The brief
    Credit · unclaimed

Open 01 · the brief

Open appraisal-gap replication

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public code repository that rebuilds the appraisal gap from FHFA's appraisal-level file (version 2.1) joined to Census tract data on race, for 2013 to 2022, and then asks how much of it survives controls. It reports the share of purchase appraisals below the contract price by the tract's minority share, first raw, then within county and year, then after controls for price band and the condition and quality ratings, and the share landing exactly at the contract price, DNB's measure of the pull toward the price.

Why it is still open

FHFA's aggregate dashboards already publish the raw share of appraisals below the contract price for tracts that are 0% to 50%, 50.1% to 80% and 80.1% to 100% minority, nationally and down to single tracts, but no gap after controls, and that is where AEI and Brookings disagree; none of the pages opened links to their code. FHFA removed borrowers' race from the record-level file in February 2025, so only neighborhood-level work is possible. A GitHub search for "uniform appraisal dataset" on 19 November 2026 returned two repositories, neither an analysis.

fhfa.govfhfa.gov

Built with what exists

FHFA's appraisal-level public file, version 2.1, in CSV, with fields such as contract price, appraised value, appraisal-to-contract ratio, condition, quality and tract codes; US Census Bureau American Community Survey 5-year tract tables on race and Hispanic origin, through the Census API; R (tidycensus, fixest) or Python (pandas, statsmodels); a coding agent such as Claude Fable 5.2 (see the Note "Three working devices").

  1. First step

    Download the combined file, keep purchase records, compute the yearly share with an appraisal-to-contract ratio under 100% for each tract group, and compare it with FHFA's dashboard.

  2. Done looks like

    One command rebuilds every table from the raw download; the raw below-contract shares for FHFA's three tract groups land within 1 percentage point of FHFA's own dashboard figures for each year, allowing for the file being a 5% sample; the controlled gaps are published with their method beside AEI's and Brookings' readings; and the repository reports record counts and the tracts lost to FHFA's rule suppressing areas with fewer than 11 appraisals.

Open 02 · the brief

A test for bias scanners

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of about 2,000 made-up sentences of the kind written in the neighborhood and site sections of appraisal reports, each labeled by two people as acceptable, subjective or prohibited under the guidance on unacceptable appraisal practices that Fannie Mae and Freddie Mac publish. It includes traps both ways: coded wording that names no group, and fair sentences that a word list would flag, such as an address on Desirable Lane. Word-list scanners and AI models are scored on what they catch and what they flag wrongly.

Why it is still open

Fannie Mae and Freddie Mac scan every report for such wording (see Achieved), and vendors sell AI review that claims to go "beyond the rule-based checks" (PropMix, December 2024) or to find bias "far more accurately than lexicon-based tools" (Reggora, 2023), but neither report gives accuracy figures. Real appraisal narratives are not public; FHFA's file has 37 fields and no text. The Mortgage lending file's open test of planted errors in made-up reports targets mismatches in data and photos, not wording.

valuationreview.commannpublications.com

Built with what exists

The two companies' published guidance as the rulebook; an AI model to draft candidate sentences, with people writing the hardest cases; Label Studio for blind double labeling; Gemini Decision 1 to sort each sentence into the three labels with a probability for each (see the Note "A thousand questions, one pass"); Claude as a second reviewer (see the Note "Seven more flagged"); every tool run 20 times, as in Microsoft's ThinkingBox tests (see the Note "All twenty times").

  1. First step

    Write 100 sentences, a third under each label, have two people label them blind, and measure how often they agree.

  2. Done looks like

    2,000 sentences with the two labelers' agreement rate published, and, for a plain word list and at least two AI models, the share of prohibited and subjective sentences caught, the share of acceptable sentences wrongly flagged and the spread across 20 runs.