Bias and fairness in automated valuation

The PAVE task force, a group of thirteen federal agencies and offices, made the point in its March 2022 action plan. It noted that a model may rely on biased data that could replicate past discrimination, or on data that includes protected characteristics or close proxies for them. Because models run at scale, a discriminatory result could cause widespread harm.

The same plan also said that, used properly, models have the potential to reduce human bias and improve consistency. Both statements can be true at once.

The plan said studies differ on whether automated models reduce the influence of racial and ethnic bias, and that some researchers have observed these models to be less accurate in Black neighborhoods. It also said bias in automated models had not been studied as closely as bias in appraisals, and that more research was needed to learn whether models might carry past misvaluation forward.

The PAVE plan described these models as building on the sales comparison approach, with selection, adjustment and weighting often based on statistical patterns in previous transactions. The result is often an estimate for a property in average condition, produced without observing the property.

The plan made a related point about appraisal waivers. Skipping the appraiser removes one possible source of bias, but the estimated values that shape waiver eligibility may still reflect past bias.

In the rule, the agencies noted that existing nondiscrimination laws already apply to both appraisals and automated models. The new factor makes that obligation part of how an institution governs its models. The rule took effect on October 1, 2025.

Under the CFPB's Regulation B, a report produced by an automated valuation model is one example of a valuation that counts. The PAVE plan also said the CFPB would encourage lenders to tell borrowers how to ask for reconsideration of a valuation they believe is inaccurate.