Standard machine learning is rarely "fairness-neutral." As Lydia T. Liu demonstrated in The implicit fairness criterion of unconstrained learning, empirical risk minimization implicitly prioritizes group calibration. While this sounds positive, it often forces models to reconstruct hidden demographic information from available features to satisfy internal constraints, leading to violations of other fairness notions like equalized error rates.
The challenge for modern AI is that we often cannot—or should not—use sensitive attributes like race, gender, or age during the prediction phase (inference). Legal frameworks like GDPR and ethical privacy concerns often forbid the collection of protected characteristics at runtime. This has led to a research direction focused on attribute-blind fairness: ensuring equitable outcomes without requiring sensitive data at the moment a decision is made.
Scrubbing the Latent Space
One primary way to achieve attribute-blindness is to ensure the model's internal "thoughts"—its latent representations—simply do not contain information about the sensitive attribute. If the model is physically unable to distinguish between groups in its hidden layers, it cannot discriminate between them in its output.
Alberto Sinigaglia and colleagues propose this in Simple and Effective Specialized Representations for Fair Classifiers. They use Characteristic Function Distance (CFD) to align data representations across groups during training. By mapping data into a group-invariant latent space, the downstream classifier becomes "blind" to protected status. Similarly, Mohsen Ghassemi uses Optimal Transport in Auditing and Enforcing Conditional Fairness via Optimal Transport to enforce invariance, allowing models to satisfy conditional demographic parity without needing sensitive labels at runtime.
In specialized domains like face recognition, Jean-Rémy Conti and others developed "Ethical Modules" in Mitigating Gender Bias in Face Recognition using the von Mises-Fisher Mixture Model. These modules use sensitive data during training to learn corrective mappings on a multidimensional sphere, but remain entirely attribute-agnostic once deployed.
The Power of the Bias Score
If we cannot scrub the representation, we can instead treat fairness as a post-hoc correction. The intuition here is that even if a model is biased, we can "nudge" its predictions if we know how much it tends to over- or under-predict for certain groups.
Yegor Klochkov and Wenlong Chen proved a significant theoretical result in Post-hoc Bias Scoring Is Optimal For Fair Classification: the mathematically "optimal" fair classifier is simply a standard classifier adjusted by a bias score.
Crucially, this bias score can be estimated using an auxiliary model that "guesses" the sensitive attribute from non-sensitive features. As Kazuto Fukuchi explored in Prediction with Model-Based Neutrality, this secondary model allows the main system to remain neutral relative to inferred distributions. By using these estimated probabilities, researchers like Xiangyu Chang in Bayes-Optimal Fair Classification with Multiple Sensitive Features can adjust decision thresholds at test time without ever asking the user for their protected status.
Fairness Without Training Labels
The most difficult frontier is achieving fairness when sensitive attributes are missing even during the training phase. This is often called "oblivious fairness."
One approach is to use proxy features to recover latent sensitive information. Enyan Dai and colleagues introduced FairWS in Learning Fair Models without Sensitive Attributes: A Generative Approach, which uses generative models to estimate hidden group memberships. In graph-based learning, Zichong Wang uses Bayesian inference in fairGNN-WOD: Fair Graph Learning Without Complete Demographics to "guess" demographic proxies and disentangle them from the task at hand.
Alternatively, we can optimize for "worst-case" scenarios. Xuanqian Wang proposed VFair in Towards Harmless Rawlsian Fairness Regardless of Demographic Prior, which achieves fairness by minimizing the variance of individual training losses. The intuition is simple: if the model performs equally well on every individual (or every small subset of data), it is mathematically impossible for it to discriminate against any specific group, even if those groups are never explicitly labeled. Jing Li takes a similar path in Alpha and Prejudice: Improving α-sized Worst-case Fairness via Intrinsic Reweighting, using per-sample gradients to identify and upweight underrepresented groups without ever seeing a demographic label.
Why It Works: The Proxy Principle
The common thread across these papers is the realization that sensitive attributes are rarely truly "hidden." They leave shadows across the entire dataset. Whether through Counterfactual Fair Opportunity auditing or Posterior Sampling for image generation, these methods work because they model the relationship between the "visible" features and the "invisible" protected traits. By either breaking that link (invariance) or accounting for it (bias scoring), we can build systems that are fair by design, rather than fair by observation.