Fairness Through Unawareness

Fairness through unawareness is the mistaken belief that algorithms are fair if they don't explicitly use protected attributes like race or gender. This is false because other data (like zip codes or school names) can serve as proxies for these characteristics and reproduce the same biases. For example, a hiring algorithm that excludes race but uses school names might still discriminate against certain racial groups because some schools predominantly serve specific communities. This approach creates a false sense of fairness while allowing discrimination to continue unnoticed. Reflect on this: What other seemingly neutral data might encode sensitive information about people?

Fairness through unawareness represents one of the most common and dangerous misconceptions in algorithmic ethics. The approach suggests that simply removing protected attributes (race, gender, age, disability status) from datasets automatically produces fair algorithms. However, this ignores the complex web of correlations that exist in real-world data. For instance, zip codes correlate strongly with race and socioeconomic status in many countries, while graduation years can correlate with age. Even seemingly innocent data points like browser history or app usage patterns can reveal protected characteristics.

The mechanisms that make this approach fail operate through proxy variables and statistical dependencies. When an algorithm learns patterns from historical data, it captures societal biases encoded in those patterns. Even without explicit protected attributes, the system can learn to use combinations of neutral features that effectively reproduce discriminatory outcomes. For example, Amazon's experimental hiring tool penalized resumes containing the word "women's" (as in "women's chess club captain") despite not having gender as an explicit input [1].

The consequences extend beyond hiring decisions. In healthcare, algorithms that use healthcare spending as a proxy for health needs systematically underestimate the needs of Black patients, who historically have had less access to and utilization of healthcare services due to systemic barriers [2]. Similarly, predictive policing systems that use historical arrest data as input disproportionately target neighborhoods with existing high police presence, creating a feedback loop that compounds racial disparities. These examples demonstrate how fairness through unawareness not only fails to achieve fairness but can actively reinforce and perpetuate existing inequities while providing deniability to those implementing the systems.

Fairness through unawareness represents one of the earliest and most naive approaches to algorithmic fairness, formalized in early machine learning literature before researchers identified its fundamental flaws. The approach stems from a misunderstanding of both discrimination law and statistical dependence, assuming that formal exclusion of protected attributes constitutes sufficient protection against discriminatory outcomes.

• Barocas & Selbst (2016): Establish that even without explicit protected attributes, algorithms can violate anti-discrimination laws through disparate impact when they reproduce existing patterns of inequality [1].
• Noble (2018): Demonstrates how search algorithms can reinforce stereotypes and marginalize communities of color without explicit racial categorization, through learned associations from biased training data [2].
• Dwork et al. (2012): Introduce formal definitions of algorithmic fairness that go beyond attribute exclusion, emphasizing the need to consider actual outcomes rather than just inputs in determining fairness [3].

Fairness through unawareness fails under virtually every formal definition of algorithmic fairness beyond procedural fairness. It cannot address historical inequities, statistical dependencies, or feedback loops that perpetuate bias. The approach particularly fails in high-stakes domains (healthcare, criminal justice, employment) where proxy variables are abundant and the costs of discrimination are severe. Additionally, this method provides false comfort to implementers, potentially worsening outcomes by creating unwarranted confidence in fundamentally flawed systems while discouraging more robust fairness interventions.