Algorithms that "Don't See Color": Comparing Biases in Lookalike and Special Ad Audiences

Removing sensitive data alone doesn't eliminate bias: Facebook's removal of demographic features from its ad-targeting algorithm failed to prevent biased outputs, demonstrating that simply deleting protected characteristics from an algorithm's inputs is insufficient to prevent discrimination in real-world systems.

Hidden patterns can perpetuate discrimination: The study provides experimental evidence that algorithms can infer and act on demographic information indirectly through other data points (like user interests, locations, or behaviors), even when explicit demographic data has been removed—a phenomenon known as "proxy discrimination."

Current legal settlements may be inadequate: The Facebook case shows that compliance approaches based on feature removal, which some organizations adopt to address discrimination concerns, do not actually solve the underlying problem and may create a false sense of fairness.

Broader implications for high-stakes decisions: The findings suggest that credit, employment, healthcare, and criminal justice systems relying on similar "color-blind" algorithmic approaches may also be producing discriminatory outcomes, indicating a systemic problem across multiple industries.

Alternative safeguards are needed: Organizations should move beyond simple technical fixes and implement more comprehensive bias mitigation strategies, such as ongoing auditing, diverse training data, fairness testing, and human oversight—particularly for algorithms that determine access to critical life opportunities.