The most intuitive way to make an algorithm "fair" is to simply hide the sensitive information. If a model cannot see a person’s race, gender, or age, the logic goes, it cannot possibly discriminate against them. This approach, known in the literature as Fairness through Unawareness (or anti-classification), is a common starting point for practitioners.
However, a massive body of research has demonstrated that this "blindness" is almost entirely ineffective. As Lydia T. Liu argues in The implicit fairness criterion of unconstrained learning, standard machine learning naturally prioritizes group calibration. This means that as models become more accurate, they don't become more "neutral"—they actually become better at identifying and treating groups differently, even when those group labels are missing.
The Mechanism of Redundant Encoding
The reason blindness fails is not due to a lack of effort in scrubbing data, but because of the high-dimensional nature of modern datasets. Information about protected attributes is rarely contained in a single column; instead, it is "smeared" across the entire feature set through redundant encoding.
In On formalizing fairness in prediction with machine learning, Mykola Pechenizkiy and Pratik Gajane explain that structural inequalities ensure that sensitive information is mirrored in "proxy variables." For example, in Fairness in Algorithmic Profiling: A German Case Study, Christoph Kern and colleagues found that models predicting unemployment could easily reconstruct a person's migration status from their industry or employment history.
Even in visual data, where we might assume a model only sees pixels, the bias is pervasive. Nicole Meister and Olga Russakovsky show in Gender Artifacts in Visual Datasets that models can predict gender from background objects or even the average color of an image, rendering "blinding" techniques like blurring faces useless.
Superhuman Inference from Latent Proxies
The most striking evidence of this failure comes from medical AI. Researchers led by Imon Banerjee and Leo Anthony Celi discovered that deep learning models can accurately predict a patient's self-reported race from chest X-rays—a task human radiologists cannot perform. In Reading Race: AI Recognises Patient's Racial Identity In Medical Images, they demonstrate that this "superhuman" ability persists even when the images are heavily degraded.
This suggests that models aren't just finding simple proxies; they are identifying subtle, non-linear patterns across the entire data manifold. Richard J. Chen and Tiffany Chen frame this in Algorithm Fairness in AI for Medicine and Healthcare as a dataset shift problem. Because the data itself is a product of biased clinical workflows and societal structures, the "signal" of the protected attribute is inseparable from the "signal" of the task.
Why Blinding Can Be Harmful
Paradoxically, removing sensitive attributes can sometimes make models less fair. Omer Ben-Porat and Moshe Tennenholtz argue in Protecting the Protected Group: Circumventing Harmful Fairness that by ignoring group-specific contexts, models may produce inaccurate classifications that fail to account for the unique circumstances of disadvantaged populations.
Furthermore, Philippe Besse notes in A survey of bias in Machine Learning through the prism of Statistical Parity that because of "social determinism," the model's predictions will remain biased as long as the underlying social correlations exist. Blinding simply hides the reason for the bias, making it harder to audit and correct.
Beyond Blindness: Active Decoupling
If hiding the data doesn't work, the alternative is to force the model to "unlearn" the sensitive information. This has led to several active mitigation strategies:
- Adversarial Unlearning: Edward Raff and Jared Sylvester proposed Gradient Reversal Against Discrimination, which uses an adversarial layer to penalize the model if it learns any information that could identify a protected attribute.
- Causal Modeling: Matt Kusner and Chris Russell introduced Counterfactual Fairness, which uses causal graphs to ensure a decision remains the same in a hypothetical world where the individual's protected attribute was different.
- Latent Decoupling: Methods like Fair Neighbor Embedding by Jaakko Peltonen attempt to "factor out" sensitive groupings from high-dimensional representations while preserving the useful structural relationships.
The consensus across these 225 papers is clear: fairness cannot be achieved by omission. To build equitable systems, researchers must move from "fairness through unawareness" to active, audit-driven interventions that acknowledge and decouple the deep-seated correlations within our data.