Machine-learning proxies black-box complex social domains to enable intervention
Evidence
There is first the problem, already noted, that decision makers may not understand their data sources. The increasing use of data streams with insufficient contextualization in government decision-making through so-called dashboards has provoked concern in the United Kingdom. The problem is deepened by the distinctive way that data generates knowledge through Machine Learning. Data processing generates proxies—regular correlations that stand in for other variables of social knowledge—out of which larger models of the world to be governed are built. The whole point of proxies is to black-box a more complex social domain while facilitating intervention. When an insurer assumes that driving after midnight equates to choosing to take more risk even if a person’s working conditions require it or when algorithms treat geographical areas as proxies for levels of crime risk, complexities on the ground become unreadable. — Nick Couldry & Ulises A. Mejias, The Costs of Connection, 450
Categorization
- DEFINE: Defines ‘proxy’ in the ML-governance sense: regular correlations standing in for unmeasured social variables while black-boxing complexity.
- XREF: Proxy-driven unreadable complexity connects to the attention/surveillance school’s critique of how algorithmic mediation renders lived experience illegible. — connects to: Machine learning increasingly intertwines technical problems with human and social questions; Computer simulation can test hypotheses about large-scale social learning — Cailin O’Connor and James Owen Weatherall, The Misinformation Age: How False…, 47
- SEED: Angle: the insurance example — driving after midnight as risk proxy — illustrates how proxies punish structural labor conditions, a vivid case for an essay on algorithmic governance’s blindness.