For years, a foundational assumption in algorithmic fairness has been the existence of an inevitable zero-sum trade-off: the "price of fairness." Under this traditional paradigm, enforcing equitable outcomes across different demographic groups requires introducing constraints that inherently degrade a model's overall predictive accuracy.
However, a growing body of research challenges this dogma. Instead of a zero-sum conflict, researchers are uncovering a fairness-accuracy synergy. This perspective argues that the perceived trade-off is often an artifact of flawed data representation, biased ground-truth labels, or sub-optimal optimization paths. When these underlying systemic issues are corrected, fairness and accuracy can be optimized simultaneously, yielding a "positive-sum" outcome where improving equity actually drives better overall performance.
Why Bias Degrades Accuracy
To understand why fairness and accuracy can be synergistic, we must first understand why bias occurs in machine learning models. Bias is rarely an optimal strategy for true generalization; rather, it is often a form of structured error or overfitting to a dominant demographic trend.
1. Biased Ground Truth and Structured Noise
When a model is trained and evaluated on biased historical data, its performance metrics become distorted. If the ground-truth labels themselves reflect systemic or selection bias, a model that perfectly mimics the data will appear highly accurate on paper, despite perpetuating real-world harms.
In Bias-tolerant Fair Classification, YIXUAN ZHANG et al. demonstrate that treating label and selection bias as structured label noise allows models to recover "latent fair labels." By targeting these latent fair labels rather than the corrupted ground truth, models can simultaneously improve both true predictive accuracy and fairness metrics, bypassing traditional optimization trade-offs.
2. Under-representation and Poor Generalization
Standard empirical risk minimization (ERM) prioritizes the majority group because it contributes the most to the aggregate loss function. Consequently, the model underfits or misclassifies minority groups, failing to learn the nuanced features required to generalize across the entire population.
In FAE: A Fairness-Aware Ensemble Framework, Vasileios Iosifidis et al. show that explicitly balancing the representation of minority groups enhances the model's ability to generalize across the entire dataset. Rather than dragging down majority performance, correcting this under-representation lifts the model's overall balanced accuracy. Similarly, in decentralized settings, Huimin Zeng et al. demonstrate in Fair Federated Learning with Biased Vision-Language Models that mitigating demographic biases allows models to correctly classify previously misidentified samples from underrepresented populations, boosting total system accuracy.
3. Erasing Human Nuance
Data preprocessing often compresses diverse human perspectives into a single majority-vote label, erasing marginalized viewpoints. In On the Interplay between Human Label Variation and Model Fairness, Timothy Baldwin et al. show that training models on the complete distribution of human annotations—preserving minority opinions instead of discarding them—simultaneously improves both model accuracy and group-level fairness. Capturing human disagreement provides the rich signal necessary to treat different demographic groups equitably.
Mechanisms for Positive-Sum Fairness
Achieving a win-win scenario requires moving away from naive constraints that achieve parity by "leveling down"—that is, reducing the accuracy of the majority group to match the lower accuracy of a minority group. Instead, positive-sum interventions focus on lifting disadvantaged groups while preserving or enhancing overall performance.
Traditional "Leveling-Down" Fairness:
[Majority Accuracy ↓] --> Parity Achieved (Lower Overall Utility)
Positive-Sum Fairness Synergy:
[Minority Accuracy ↑] --> Parity Achieved (Higher Overall Utility)
Preventing Performance Degradation via Continual Learning
A common pitfall when fine-tuning a model for fairness is that the optimization process "forgets" the effective patterns it learned for the majority group. To solve this, Lucas Mansilla et al. introduce a continual learning perspective in BM-CL: Bias Mitigation through the lens of Continual Learning. By treating fairness optimization as a sequential task and using knowledge distillation to protect the high-performing features of advantaged groups, the authors demonstrate that models can lift marginalized subgroups without paying a performance tax on the rest of the data.
Furthermore, in Fairness of Deep Ensembles, Estanislao Claucich et al. show that simple deep ensembles naturally shrink performance gaps between demographic groups. They argue that true fairness often requires intentionally unbalancing datasets during training to allocate more capacity and samples to groups facing inherently harder tasks, ensuring no group is left behind.
Escaping Local Optima via Post-Hoc Adjustments
Many standard models get trapped in local optima where structural flaws harm both fairness and utility. In Search-based Automatic Repair for Fairness and Accuracy in Decision-making Software, Jie Zhang et al. use search-based post-processing to mutate trained models. They find that specific structural adjustments can simultaneously improve both predictive accuracy and fairness metrics, achieving a win-win outcome in 61% of tested cases.
For large language models suffering from class accuracy imbalances, Ruixi Lin et al. propose a post-hoc discrete optimization method in COBias and Debias. By reweighting output probabilities using nonlinear integer programming, they slash accuracy disparities by 61% while simultaneously boosting aggregate accuracy by 18%, proving that leveling the playing field across categories can unlock latent model performance.
Feature Disentanglement and Prompt Engineering
Aggressively scrubbing sensitive attributes from data can inadvertently destroy critical predictive signals. Modern representation learning avoids this by cleanly separating sensitive information from core task features.
In FairDisCo: Fairer AI in Dermatology via Disentanglement Contrastive Learning, Siyi Du et al. combine disentangled representation learning (to strip away skin-tone bias) with supervised contrastive learning (to preserve critical diagnostic features). This dual approach allows their model to improve both diagnostic accuracy and demographic fairness simultaneously across diverse skin types.
In the realm of in-context learning for tabular data, Jingyu Hu et al. demonstrate in Enhancing Fairness in In-Context Learning that deliberately over-representing minority groups in few-shot demonstrations resolves model under-fitting. This targeted prompt engineering guides the language model to form a more robust decision boundary, maximizing fairness while simultaneously improving predictive accuracy.
Synergy Across Diverse Domains
The alignment of fairness and accuracy is not isolated to standard image or tabular classifiers; it manifests across a wide variety of machine learning architectures and paradigms.
- Federated Learning: In distributed environments, data heterogeneity often amplifies bias. In GIFAIR-FL, Xubo Yue et al. introduce a regularization term that penalizes loss discrepancies across client groups, driving the system toward fairer outcomes without sacrificing prediction accuracy. Similarly, in Boosting Fairness and Robustness in Over-the-Air Federated Learning, Halil Yigit Öksüz et al. utilize a fairness-centric minmax objective optimized for the worst-performing agent. This approach ensures the global model learns robust features from all participating distributions, significantly improving classification accuracy in highly imbalanced or noisy wireless environments.
- Graph Neural Networks: Graph structures can easily propagate and amplify historical biases. In FairEGM, Sean Current et al. employ emulated graph modifications to train GNNs for fairer link prediction and recommendation. Their framework achieves substantial fairness gains with minimal impact on accuracy, occasionally outperforming standard models. To stabilize these predictions, Yaning Jia et al. introduce JacoLip in Stabilizing GNN for Fairness via Lipschitz Bounds, constraining how much a GNN's output can change relative to its input to drastically improve individual fairness metrics while maintaining or slightly improving node classification accuracy.
- Domain Adaptation in Text: In Easy Adaptation to Mitigate Gender Bias in Multilingual Text Classification, a single-author study by Xiaolei Huang treats gender bias as a domain adaptation problem. By isolating gender-specific signals across multiple languages, the author demonstrates that reducing disparate impact does not require a trade-off with predictive power, maintaining or marginally improving macro F1 scores across hate speech and sentiment analysis tasks.
Conclusion
The emerging consensus across these diverse subfields is clear: fairness and utility are not inherently conflicting objectives. When bias is recognized as a symptom of poor generalization, corrupted ground truth, or sub-optimal optimization, fairness interventions cease to be a tax on performance. Instead, they become a powerful tool for building more robust, accurate, and equitable machine learning systems.
Key Foundational Work
Feature and Label Embedding Spaces Matter in Addressing Image Classifier Bias (2021) — contributed 1 concept elements to this direction.
Promoting Fairness through Hyperparameter Optimization (2021) — contributed 1 concept elements to this direction.
FairNN- Conjoint Learning of Fair Representations for Fair Decisions (2020) — contributed 1 concept elements to this direction.