Human-Centered Fairness Framework

• The human-centered fairness framework is a systematic approach that integrates stakeholder values, lived experiences, and technical fairness metrics.
• It formalizes multi-objective optimization by explicitly visualizing trade-offs among accuracy, fairness, privacy, and controllability.
• Applied in high-stakes sectors like healthcare, criminal justice, and lending, it empowers stakeholders to negotiate and audit algorithmic decisions.

A Human-Centered Fairness Framework refers to the systematic integration of stakeholder values, lived experience, and contextual judgment into the formal specification, evaluation, and deployment of algorithmic fairness in sociotechnical systems. Moving beyond purely technical criteria, such frameworks embed human judgment into the end-to-end machine learning lifecycle, explicitly balancing accuracy, fairness (across multiple populations and according to multiple competing metrics), privacy, and controllability. The most rigorous human-centered frameworks provide (1) formal constructs to define and visualize multi-objective trade-offs, (2) interactive mechanisms for stakeholder or end-user elicitation and negotiation, (3) privacy-preserving audit and explanation loops, and (4) empirical or procedural assurance that human values are not only solicited but binding on the system. These frameworks are applied in high-stakes domains such as criminal justice, healthcare, lending, and public services, where the consequences of algorithmic decisions are directly experienced by diverse and often vulnerable communities and subject to regulatory scrutiny (Sanchez et al., 11 Nov 2025, Rahman et al., 7 Dec 2025, Luo et al., 16 Jul 2024, Ferreira et al., 3 Apr 2025).

1. Fundamental Constructs and Objectives

Human-centered fairness frameworks are designed to bridge the gap between advanced statistical fairness criteria and real-world value alignment in machine learning. Their core objectives are:

• To make explicit and navigable the trade-offs between accuracy, multiple formal fairness definitions (e.g., demographic parity, equalized odds), and privacy.
• To operationalize fairness as a fundamentally sociotechnical judgment, requiring both formal certification (frontiers or constraints) and the embedding of human agency—either at the stakeholder, end-user, or collective level—into the workflow (Sanchez et al., 11 Nov 2025, Rahman et al., 7 Dec 2025).
• To ensure that transparency and normative intent are systematically reflected and auditable, both at the model development phase and at deployment or decision time.

Typical instantiations include FAIRPLAI, which integrates privacy-fairness-accuracy frontiers with policy tuples binding stakeholder input, and frameworks that provide structured interfaces for end-user inspection, contestation, and redress (Sanchez et al., 11 Nov 2025, Luo et al., 16 Jul 2024, Ferreira et al., 3 Apr 2025).

2. Mathematical Foundations and Trade-off Formulation

Human-centered fairness frameworks formalize model selection as a multi-objective optimization problem:

$\min_{\theta}~ L(\theta;D)$

subject to

$M \text{~satisfies~} (\epsilon, \delta)\text{-DP}$

$\Delta_{\text{fairness}}(\theta) \leq \tau$

where $L(\theta;D)$ represents predictive loss, $M$ is a privacy mechanism (e.g., DP-SGD), $(\epsilon, \delta)$ are differential privacy parameters, and $\Delta_{\text{fairness}}$ is a selected disparity metric (such as demographic parity or equalized odds) bounded by $\tau$ (Sanchez et al., 11 Nov 2025).

The feasible set of models—privacy–fairness frontier—is computed via Lagrangian relaxation or grid search over $(\epsilon, \tau)$, producing explicit Pareto surfaces:

• x-axis: privacy budget $\epsilon$ (smaller values imply stronger privacy)
• y-axis: fairness disparity (e.g., $|\Pr(\hat Y=1|A=0)-\Pr(\hat Y=1|A=1)|$)
• z-axis or color: accuracy.

Stakeholders interact directly with this surface, specifying their own constraints via a policy tuple $\mathcal{P} = (F, \tau, \epsilon, M)$ (Sanchez et al., 11 Nov 2025). Unifying frameworks further allow the weighting and negotiation among 8+ fairness metrics grouped by granularity (individual/group), societal stance (infra-marginal/intersectional), and evaluation regime (outcome/EOO) (Rahman et al., 7 Dec 2025).

3. Human-in-the-Loop and Stakeholder Integration

A hallmark of human-centered frameworks is continuous stakeholder engagement, both at design time and during post-deployment audits:

• Interactive Visualizations: Frontiers and model tables enable users to select fairness criterion, target disparity, privacy level, and required performance.
• Natural Language Summaries: Translation layers present trade-offs and selected models in clear prose, with high translation fidelity between technical and lay description (Sanchez et al., 11 Nov 2025).
• Audit Loops: Stakeholders can request DP-protected explanations (LIME/SHAP with noise), counterfactuals, or cohort analysis; privacy accountant mechanisms ensure total $\epsilon$ budget is respected.
• Contestability and Redress: Individual-centric frameworks empower users to inspect their own decisions, query “Am I being treated fairly?” via both group-defined (statistical parity, conditional statistical parity) and individualized (recourse, counterfactual stability, actionable minimal changes) assessments, escalating via structured dialogue to third-party audit if necessary (Ferreira et al., 3 Apr 2025, Luo et al., 16 Jul 2024).
• Consensus and Negotiation: Multi-party interfaces implement Explain–Ask–Review–Negotiate protocols, supporting preference elicitation (e.g., ranking top-3 metrics, declaring thresholds), team-based review, and negotiation (majority voting, hybrid compromise, Borda aggregation) to reach actionable consensus in metric and threshold selection (Luo et al., 16 Jul 2024).

These mechanisms ground technical trade-offs in lived stakeholder experience and facilitate both legitimacy and procedural accountability.

4. Empirical Validation and Case Studies

Human-centered fairness frameworks have been empirically validated across multiple high-stakes benchmarks and real-world case studies:

• On datasets such as Adult Census, Student Performance, ACTG175 (clinical), CDC Diabetes, and COMPAS, FAIRPLAI demonstrates that strong privacy guarantees can be achieved simultaneously with substantial reductions in fairness disparities; human-in-the-loop choice consistently outperforms automated selection on fairness at fixed accuracy (Sanchez et al., 11 Nov 2025).
• In criminal justice (COMPAS), multi-stakeholder optimization reveals that prioritizing different fairness metrics (e.g., EOO for public safety vs. intersectional group fairness for civil rights) leads to diverse optima, but explicit negotiation across metrics can yield Pareto-improving or complementary trade-offs (Rahman et al., 7 Dec 2025).
• In health and credit contexts, frameworks such as EARN Fairness show that affected parties prefer and can reach consensus on hybrid metrics—with subgroup-based conditional statistical parity commonly chosen as primary and individual fairness metrics (consistency, counterfactual fairness) as secondary, employing explicit negotiation strategies (Luo et al., 16 Jul 2024).

Translation between plain language and formal policy tuples maintains over 85% fidelity, and automated audit tools track both privacy and fairness metric satisfaction after deployment (Sanchez et al., 11 Nov 2025).

5. Auditing, Explanations, and Privacy Guarantees

Robustly human-centered frameworks embed differentially private audit loops:

• Noise Injection: All explanations (feature attributions, counterfactuals) employ DP via noise calibration and subsampling; for each query, a privacy accountant tracks cumulative budget (Sanchez et al., 11 Nov 2025).
• Statistical Testing: DP-protected significance tests (e.g., two-sample z-test for group disparity) yield noisy p-values and confidence intervals, allowing stakeholders to judge the statistical reliability of observed gaps.
• Audit Contracts: All decisions, policy tuples, and selected model frontiers are logged, enabling retrospective review and regulatory auditability.

Audit outputs are exposed in formats accessible to both technical experts (numerical summaries, confidence intervals) and non-expert stakeholders (visual sliders, trade-off sentences).

6. Practical Guidelines, Limitations, and Prospective Directions

Actionable best practices have emerged:

Practice	Description
Early stakeholder engagement	Co-design fairness intent and tolerance levels using controlled vocabulary
Privacy calibration	Map qualitative descriptors (“very strong”, “moderate”) to numeric $\epsilon$ values; document all choices
Privacy budgeting	Partition $\epsilon$ among training, explanation, and audit; maintain DP accountant ledger
Automated constraint monitoring	Post-deployment checks of fairness and utility vs. policy tuple
Audit contract maintenance	Persist full documentation of constraints, frontiers, selected models, and explanations
Frontier updating	Re-train and re-analyze frontiers upon significant data drift or deployment regime change

Limitations include varying interplay between privacy and fairness by data modality (e.g., categorical vs. binary attributes), the need for careful translation between lay intent and technical metrics, and scalability challenges for very large or heterogeneous stakeholder pools. Future research areas identified include prospective studies with live governance, deeper analysis of metric complementarity, and extension to multiclass targets and more complex group structures (Sanchez et al., 11 Nov 2025, Rahman et al., 7 Dec 2025, Luo et al., 16 Jul 2024).

7. Relationship to Broader Fairness and Privacy Frameworks

Human-centered fairness frameworks can be embedded within larger sociotechnical and legal architectures, interfacing with:

• Regulatory requirements (e.g., GDPR, EU AI Act) for contestability, transparency, and privacy
• Domain-specific best practices (healthcare, finance, criminal justice)
• Broader design and audit methodologies for responsible AI, including hybrid approaches that integrate context-dependent equality of opportunity, descriptive frameworks (e.g., ACROCPoLis), and cross-disciplinary perception studies (Tubella et al., 2023, Yaghini et al., 2019).

By making trade-offs explicit, exposing them to human negotiation, and maintaining interpretability and privacy guarantees throughout, these frameworks transform fairness assessment from a purely technical constraint to an iterative, accountable, and legitimate component of sociotechnical system design (Sanchez et al., 11 Nov 2025, Rahman et al., 7 Dec 2025, Luo et al., 16 Jul 2024, Ferreira et al., 3 Apr 2025).