What is algorithm bias?
Recent studies show how human subjectivities infiltrate machine learning and make it imperative that we critically assess machine agency.
A Harvard study found that Google searches serve up ads for arrest records to minority web surfers (Sweeney 2013 Harvard)
A Carnegie Mellon study found that Google searches tend to present high income job ads to men moreso than women (Datta et al. 2015)
A University of Washington study found that Google extracted fewer percentage pictures of women CEOs than women CEOs actually exist in the world (Kay et al. 2015)
Furthermore, researchers at the University of Maryland found that text and data mining models can predict hidden attributes—-as well as make predictions from micro-decisions, micro-patterns, micro-attributes about age, trust, gender, sexu
al orientation, drug use, alcoholism —-that are strong indicators of behavioral and personal characteristics from seemingly irrelevant data (Golbeck 2012, 2015).
We want to use big data. It only makes sense to aggregate a glut of information for helping provide better and more tailored products and services. But these same strengths of aggregation, pattern detection and prediction can also impose and expose personal, medical and behavioral characteristics that could be used against web surfers in ways that they neither condone nor anticipate.
The “correctives” for some of the biases of algorithms might involve constructing intricate, multi-tiered rule systems for internal validation checks meant to model “fairness” judgment. We are creating a second and third order ranking for word strings beyond counting frequencies and ranking sentiment.
