Written by Ihor Sokolyk · PropTech specialist, Software Engineer
Automated valuation models (AVMs) are sold as solutions that eliminate human subjectivity. The algorithm does not see race, does not factor in appearance, does not make decisions under the influence of mood. It seems like the ideal tool for a fair market. However, the data tells a different story: in a number of documented cases, algorithms do not simply reproduce bias, they make it harder to detect and eliminate.
How Bias Emerges in a “Neutral” Algorithm
A machine learning algorithm learns from historical data. If that data reflects decades of discrimination in the real estate market, the practice of redlining, unequal access to credit, systematic undervaluation of housing in certain neighborhoods, the algorithm internalises these patterns as the norm.
The mechanism of bias can be indirect. The algorithm does not use race as a direct input parameter. Instead, it relies on proxy variables: neighborhood median income, crime rate, school ratings, density of ageing infrastructure. These variables statistically correlate with the demographic composition of the population and through them bias enters the model while maintaining the outward appearance of objectivity.
Researchers at UC Berkeley (Bartlett, Morse, Stanton and Wallace, Journal of Financial Economics, 2022) found that Latinx and Black borrowers with the same credit risk pay higher mortgage interest rates, and that algorithmic FinTech lenders did not remove the gap: on GSE-backed loans their rate differences were similar to those of traditional lenders.
Documented Cases: What the Research Shows
Freddie Mac’s 2021 analysis of more than 12 million purchase appraisals found that appraised values fall below the contract price more often in neighborhoods with a higher share of Black or Latino residents. That study measured appraisals made by people, not algorithms. A model trained on such historical valuations can inherit the same distortions, and for AVMs the next study documents a related disparity.
A study by Zhu, Neal and Young, published in 2024 in Cityscape, the research journal of the US Department of Housing and Urban Development (HUD), examined AVMs in Atlanta and Memphis. The authors did not find systematic undervaluation, but AVM errors, measured as a percentage of the sale price, were larger in majority-Black neighborhoods than in majority-White ones. The gap remained even after they added data on property condition and applied more sophisticated ML techniques.
Researchers publishing in Humanities and Social Sciences Communications (Nature Portfolio, 2025) documented another type of asymmetry: in a study of apartment prices in four South Korean cities, ML valuation models tended to overestimate prices in better-educated neighborhoods and underestimate them in less-educated ones. An algorithm that fits existing market prices well can thus lock in and amplify social inequality.
Why This Is Harder to Fix Than Human Bias
The bias of a human appraiser is visible and can be challenged. The bias of an algorithm is hidden behind technical complexity. Even when the outcome is discriminatory, tracing which specific parameter or combination of parameters led to a lower valuation is technically difficult. This is particularly true in large neural networks where the decision-making logic is not directly interpretable.
US regulators in CFPB statements explicitly noted that algorithms are capable of masking biased inputs behind a facade of apparent objectivity, making detection and challenge more difficult.
| Parameter | Human Bias | Algorithmic Bias |
|---|---|---|
| Visibility | Manifests in appraiser behavior that can be observed | Hidden inside a mathematical model, inaccessible to direct observation |
| Scale | Limited to one specialist or team | Scales automatically across thousands of valuations simultaneously |
| Explainability | The appraiser can explain their decision | Black box decisions are often not interpretable |
| Challenge | The client can seek a second appraiser | Proving systemic bias requires a technical audit |
| Regulatory protection | Fair housing and fair lending laws apply | The same laws (FHA, ECOA) apply to algorithms; the 2024 AVM quality control rule adds specific standards |
| Self-reproduction | Depends on the specific individual | Embedded in training data and reproduced with each new version of the model |
| Detection | Through complaints, inspections, comparison of valuations | Requires dedicated statistical audit or access to source code |
The table highlights the core problem: algorithmic bias does not simply reproduce human bias, it makes it systemic, scalable and shielded by pseudo-objectivity. This is precisely why regulators in the US and EU have begun introducing specific requirements for AVM systems that never existed for human appraisers.
Planning to use AI in valuation or other real estate workflows? Explore how ORIL approaches AI strategy, data readiness and AI-powered PropTech products. Explore AI for PropTech →
What Regulators Are Doing: New Standards in the US
In June 2024, six US federal regulators — CFPB, OCC, the Federal Reserve, FDIC, NCUA and FHFA — jointly approved a new rule: Quality Control Standards for Automated Valuation Models.
The rule requires companies that use AVM systems for mortgage lending to comply with five quality control standards: ensuring valuation accuracy, protecting against data manipulation, avoiding conflicts of interest, mandatory random model testing and compliance with anti-discrimination laws.
In parallel, HUD in July 2024 filed charges of racial discrimination against a lender, a real estate appraisal company and an appraiser for violations of the Fair Housing Act in the appraisal process. This sent a signal to the market: regulatory pressure in this area is growing and technological neutrality is no longer a sufficient defense.
EU AI Act and AVM Systems: What This Means in Practice
The EU AI Act, which entered into force in August 2024 and applies in stages, introduces a classification of AI systems by risk level. After the 2026 amendments (the AI Omnibus), the requirements for stand-alone high-risk systems apply from 2 December 2027. Automated real estate valuation systems can fall under the high-risk category when they are used to assess the creditworthiness of individuals, a use the Act lists as high-risk alongside, for example, employment screening.
What this means for PropTech teams developing or planning to bring AVM products to the European market:
- Mandatory documentation. Providers of high-risk AI systems are required to maintain technical documentation describing the model architecture, training data and bias risk assessment methodology.
- Transparency and explainability. The system must be capable of explaining the valuation result in understandable terms, not simply providing a number. This directly concerns the black boxes of deep learning.
- Human oversight. For high-risk systems, the possibility of human intervention and result review is mandatory. Fully autonomous AVMs without a challenge mechanism will not meet the requirements.
- Bias audit. Before market entry and on a regular basis thereafter, mandatory testing of the system for discriminatory outcomes across protected characteristics.
- Registration. High-risk AI systems are subject to registration in a dedicated EU database.
For teams currently building AVM products aimed at the European market, these requirements should be built into the architecture at the design stage, as retrofit compliance is significantly more expensive.
Technical Approaches to Reducing Bias
Researchers and PropTech teams are developing several directions for technical solutions.
- Model auditing. Regular review of valuation outputs for systematic deviations by demographic characteristics or geographic zones. Comparison of AVM results against independent appraisals.
- SHAP values and explainable models. Technical tools that make it possible to understand which specific parameters most influenced a given valuation. They increase transparency and enable the detection of suspicious patterns.
- Alternative data sources. Incorporating non-traditional data rather than relying exclusively on historical transactional data that reflects past discriminatory practices. For example, the physical condition of infrastructure from satellite imagery, actual market activity and similar sources.
- Hybrid approaches. Combining algorithmic valuation with human review in borderline cases or when results deviate significantly from the expected range.
It is important to understand that the problem of bias in AVMs is addressable through technical methods but requires a deliberate decision by the team that designs and implements the system.
Checklist: How to Test an AVM Model for Bias
For PropTech teams developing or auditing automated real estate valuation systems:
– [ ] Training data audit. Verify whether training data includes transactions from neighborhoods that have historically experienced discriminatory practices. Is the sample balanced across geographic and demographic dimensions?
– [ ] Testing across protected groups. Compare valuation results for similar properties in neighborhoods with different demographic compositions. Is there a statistically significant difference in model error?
– [ ] Proxy variable review. Which variables does the model use as inputs? Are any of them metrics that statistically correlate with race, ethnicity or social status (for example, neighborhood median income, crime rate)?
– [ ] SHAP analysis or equivalent. Can you explain which specific parameters most influenced a given valuation? If not, the model requires further work on explainability.
– [ ] Regular retesting. Bias can emerge after the model is retrained on new data. Auditing is not a one-time measure but an ongoing process.
– [ ] Challenge mechanism. Is there a provision for human review of a result? Can the client challenge a valuation and receive an explanation?
– [ ] Regulatory compliance. For the US market: compliance with the CFPB 2024 AVM rule. For the EU market: review of EU AI Act requirements regarding system classification and mandatory documentation.
– [ ] Methodology documentation. Is the model training methodology, data selection and audit results documented internally and available for regulatory review?
Why This Matters for PropTech Teams Outside the US
Algorithmic bias is not a phenomenon limited to the American market. Any team developing or deploying an AVM system trained on historical market data faces the same risk: if the market from which the data was collected had structural inequalities, the model will internalise them.
For the European PropTech market, this question takes on particular significance in the context of the EU AI Act, which introduces the category of high-risk AI systems; automated valuation systems can fall into it when they are used to assess the creditworthiness of individuals.
Responsible design of AVM systems is not only a matter of regulatory compliance. It is a matter of trust in PropTech as an industry and in AI as a tool that should improve rather than reproduce the injustices of the market.
Want to follow developments in AI ethics in PropTech and responsible technology implementation in real estate? Listen to the Innovation Blueprint podcast — conversations with industry practitioners on the technologies shaping the future of the Built Environment. Listen to Innovation Blueprint →
Algorithmic bias in real estate valuation sits at the intersection of technology, law and social justice. It requires attention not only from regulators but from teams building PropTech products: at the design stage, in the selection of training data and in model architecture. The decisions about which data to train an algorithm on and how to verify its fairness are made by specific people and it is they who determine whether AI becomes a tool for a fairer market.
Go deeper
- An introduction to PropTech and how it is changing the real estate market
- Our review of the US PropTech market in 2025: players and trends
Sources:
- CFPB. Quality Control Standards for Automated Valuation Models. June 2024
- Debevoise & Plimpton. Federal Regulators Approve New Rule on AI Use and Bias Risks in Real Estate Valuation. July 2024
- HUD. Fair Housing Act Charge: Racial Discrimination in Appraisal Process. July 2024
- Freddie Mac. Racial and Ethnic Valuation Gaps in Home Purchase Appraisals. Economic & Housing Research Note, September 2021
- Zhu L., Neal M., Young C. Racial Disparities in Automated Valuation Models: New Evidence Using Property Condition and Machine Learning. Cityscape, vol. 26, no. 1, 2024
- An S., Song Y., Jang H., Ahn K. Asymmetric impacts of artificial intelligence on housing price valuation across education levels. Humanities and Social Sciences Communications, vol. 12, article 1884, December 2025
- Bartlett R., Morse A., Stanton R., Wallace N. Consumer-Lending Discrimination in the FinTech Era. Journal of Financial Economics, vol. 143, no. 1, 2022
- UChicago Kreisman Initiative. AI is Making Housing Discrimination Easier Than Ever Before. February 2024
- European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Annex III. 2024
- European Commission. AI Omnibus enters into force. July 2026
More PropTech & energy efficiency
Subscribe to the blog and listen to the Innovation Blueprint podcast — practical insights at the intersection of AI, IoT and real estate.
Thinking about applying AI to a real-world workflow?
ORIL works with real estate companies on AI strategy, data readiness and AI-powered PropTech products.