Early attempts to measure and mitigate social stereotypes in machine learning models relied on a beautifully simple geometric intuition: the linear subspace hypothesis. The core idea was that semantic concepts could be neatly separated into orthogonal directions within a vector space. For instance, if you want to find the "gender direction" in a word embedding space, you simply compute the vector difference between pairs of words like $\vec{v}_{\text{man}} - \vec{v}_{\text{woman}}$. To remove gender bias from a neutral word like "doctor," you project its vector onto a subspace orthogonal to that gender axis, effectively flattening the unwanted association.
This linear assumption underpins foundational debiasing algorithms. However, a growing body of research demonstrates that this geometric picture is incomplete. Social biases are rarely confined to straight lines; instead, they are encoded within complex, non-linear manifolds that wrap intricately through high-dimensional space. When we use linear projections to scrub these attributes, we don't actually eliminate them—we merely camouflage them.
The "Lipstick on a Pig" Phenomenon
To understand why linear debiasing fails, we must look at the spatial arrangement of embeddings after projection. In a seminal paper, Gonen et al. demonstrated that zeroing out a word's projection along a linear gender axis does not alter its fundamental relationship with its neighbors.
Imagine a cluster of words associated with domestic roles ("receptionist", "nurse", "housekeeper") and another associated with leadership roles ("boss", "executive", "director"). Even if you mathematically shift all these vectors so that their coordinate along the "gender axis" is exactly zero, the domestic words remain closer to each other than to the leadership words. Because these words still cluster together based on their original, biased associations, the discriminatory signal remains fully recoverable. The authors aptly described this superficial fix as putting "lipstick on a pig."
Before Linear Projection:
[Nurse] [Receptionist] [Boss] [Executive]
\ / \ /
-------(Female Axis)--------------------(Male Axis)-------
After Linear Projection (Dropped onto a single neutral line):
[Nurse] [Receptionist] [Boss] [Executive]
|-------Still Clustered-------| |-------Still Clustered-------|
This residual clustering occurs because high-dimensional spaces possess thousands of degrees of freedom. As Du et al. observed, stereotyped words exhibit a persistent geometric proximity that linear methods cannot untangle. If two concepts share hundreds of subtle, non-linear coordinate patterns due to their historical co-occurrence in training data, flattening a single dimension leaves the broader topological landscape unchanged.
Why Bias Bends: The Geometry of Deep Representations
Why does bias assume a non-linear shape in modern AI models? The complexity stems from three primary sources:
1. Deep Neural Transformations
Unlike static word embeddings (such as Word2Vec), modern language models process text through dozens of layers containing non-linear activation functions (like ReLUs) and self-attention mechanisms. As Qian et al. argue, even if a bias starts out linear in the input data, passing it through these deep layers bends and distorts the representation. The discriminatory patterns become latent within the non-linear transformations of the network, making post-hoc geometric flattening insufficient.
2. Multiclass Demographic Intersections
Binary biases (such as masculine vs. feminine) are deceptively easy to model as a single axis. However, human social categories like race, religion, or nationality do not exist in a binary vacuum. Manzini et al. and Silva et al. point out that multiclass demographic relationships are inherently non-linear. Projecting embeddings into a simplified linear space to remove multi-faceted biases causes a severe loss of semantic resolution and leaves complex, overlapping prejudices completely intact.
3. The Limits of Variance
Standard linear dimensionality reduction techniques (like Principal Component Analysis) are designed to capture the dominant axes of variance in a dataset. As Libovický demonstrates, these methods often overlook subtle or non-linear social biases that do not align with the primary axes of variance, allowing complex prejudices to persist undetected in the minor geometric folds of the embedding space.
The Probe Dilemma
The illusion of successful linear debiasing is often maintained by the tools we use to evaluate it. If an engineer tests a debiased model using a linear classifier (a linear probe) and finds that the model can no longer predict whether a word is "male" or "female," they might conclude the model is fair.
However, Kenneweg demonstrated that if you replace that linear probe with a non-linear classifier—such as a Multi-Layer Perceptron (MLP)—the classifier can easily reconstruct the original biased associations from the underlying representations. The information was never deleted; it was merely encoded in a way that required a slightly more flexible mathematical key to unlock. Ethayarajh et al. provided the theoretical foundation for this phenomenon, proving that undesirable associations frequently reside in non-linear manifolds rather than flat subspaces.
It is worth noting that the robustness of the linear hypothesis remains a subject of active academic debate. For instance, Vargas et al. developed kernelized versions of popular debiasing techniques to target non-linear manifolds, yet found that the non-linear variants did not significantly outperform standard linear projections on certain upstream benchmarks. This suggests that while bias can be recovered non-linearly, the bulk of the exploitable variance may still align closely with linear structures depending on the architecture.
Untangling the Manifold
To address the limitations of linear projection, researchers have shifted toward debiasing frameworks that respect—and actively reshape—the non-linear geometry of neural representations.
Kernel-Based Alignment
Instead of operating in the original vector space, frameworks like the Counterfactual-Debiasing (CF-Debias) method proposed by Song et al. leverage kernel functions. Kernels implicitly map the embeddings into a higher-dimensional space where the "curvy" non-linear manifolds of bias are straightened out. Once linearized in this higher-dimensional space, the sensitive attributes can be cleanly separated and neutralized using latent disentanglement.
Information-Theoretic Filters
Rather than relying on geometric projections, methods like FairFil (Cheng et al.) and the text-generation framework proposed by Cheng et al. turn to information theory. They optimize neural networks to minimize the mutual information between a sentence's semantic embedding and a protected demographic attribute. Because mutual information measures any statistical dependency—regardless of whether it is linear or non-linear—this approach strips away bias no matter how deeply it is tangled within the internal layers.
Contrastive Compression
Another powerful mechanism involves changing the training objective itself to compress the bias manifold into a single point. Kenneweg introduced objectives utilizing contrastive word and sentence pairs (e.g., "The doctor walked into the room" vs. "The female doctor walked into the room"). By forcing the deep encoder to produce identical representations for these counterfactual sentences, the model is mathematically prevented from constructing a non-linear manifold for that protected attribute, achieving a deeper, more permanent form of equity.
Key Foundational Work
MDR Cluster-Debias: A Nonlinear Word Embedding Debiasing Pipeline (2020) — contributed 2 concept elements to this direction.
Neutralizing Gender Bias in Word Embedding with Latent Disentanglement and Counterfactual Generation (2020) — contributed 2 concept elements to this direction.
Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them (2019) — contributed 1 concept elements to this direction.