False Shadows Created by Proxies When Proxy Indicators Distort the World December 2025

False Shadows Created by Proxies When Proxy Indicators Distort the World December 2025
AI tries to decipher the world from given data, but when proxy indicators that should not be used are mixed in, it reconstructs reality in a false way. This is proxy bias. Although direct information on crime itself should be used as an indicator to measure criminal tendencies, in reality, approximate data such as arrest records and number of contacts with the police are often used. In practice, however, approximate data such as arrest histories and number of police contacts are often used. AI then learns from past biases, such as statistically more frequent arrests of people of a certain race or region who have historically been targeted for public safety, and carries this into future predictions.
The case of the U.S. recidivism prediction system COMPAS is emblematic of this problem. According to a study by the research institute ProPublica, black defendants were more likely than whites to be judged as high risk even if they did not actually reoffend. This is because the AI extended the structural bias created by past public safety policies as a result of using arrest history as a proxy for criminal risk.
Proxy bias occurs in many domains, including financial scoring, insurance premium calculation, and job matching. For example, using zip codes as a proxy for credit risk quantifies historical disparities in neighborhoods and fixes disadvantaged residents. In the healthcare sector, there are examples of algorithms that use healthcare spending as a proxy for health demand that exacerbate racial disparities.
The EU AI Act and the OECD AI Principles emphasize the relevance and causal validity of data and treat the use of inappropriate proxy indicators as a serious risk. The essence of proxy bias is not in the data but in the choice of indicators, and which concepts to measure and what to use as indicators is a central ethical and technical issue.