Human-Centered Explainable AI

• Human-Centered Explainable AI is a research paradigm that centers human needs, context, and participatory design to produce context-aware AI explanations.
• It leverages interdisciplinary methods from human-computer interaction, cognitive science, and social sciences to enhance model transparency and trust calibration.
• HCXAI systems use adaptive, interactive explanation layers that tailor outputs to user expertise, ensuring clear, actionable insights in diverse application domains.

Human-Centered Explainable AI (HCXAI) is a research paradigm and design strategy that centers human stakeholders, values, organizational context, and cognitive processes in the development, deployment, and evaluation of AI explanations. It extends traditional Explainable AI (XAI) by not only exposing the internal mechanisms of AI models, but also aligning those explanations with users’ goals, expertise, sociotechnical context, and value systems. HCXAI leverages interdisciplinary insights from human-computer interaction, cognitive science, design, and social sciences to foster truly actionable, trustworthy, and context-appropriate explanations for diverse user groups.

1. Conceptual Foundations and Definitions

The definitional core of HCXAI is the explicit prioritization of human needs, interpretive practices, and social context in designing and evaluating AI explanations. Formally, an HCXAI system is structured as a pair $(f, E)$, where $f: X \to Y$ is the AI’s prediction function and $E: (f, X, Y) \to Z$ is the explanation function producing human-interpretable rationales for the model's decisions (Maity et al., 2024). This explanation function is not arbitrary: it is governed by principles such as transparency (surfacing model logic in user-interpretable terms), trust calibration (supporting appropriate user reliance), user engagement (interactive, queryable explanations), and accountability (enabling auditing and recourse) (Mangold et al., 14 Oct 2025, Ehsan et al., 2020, Ma, 2024).

Crucially, HCXAI diverges from model-centric XAI by moving explanations beyond low-level feature attributions or attention maps to high-level concepts, domain knowledge, and social processes that humans utilize for understanding and judgment (Sheth et al., 2021). The field integrates reflective sociotechnical approaches, participatory design, and value-sensitive methodologies to ensure that explanations are meaningful and contextually adapted (Ehsan et al., 2020). It is increasingly recognized that explainability is not an algorithmic property but an emergent property of human–AI collaboration, shaped by users' tasks, expertise, mental models, and the social-organizational settings in which AI is embedded (Ma, 2024, Ghai, 2023, Mangold et al., 14 Oct 2025).

2. Frameworks, Taxonomies, and System Architectures

System Decomposition and Lifecycle

HCXAI research separates systems into two interacting layers: the Core AI System, responsible for data processing and prediction, and the Explanation Component, which transforms model outputs into human-comprehensible accounts articulated through graphical, numerical, and textual modalities (Mangold et al., 14 Oct 2025). Architectures are typically modular, supporting plug-and-play of different XAI methods, explanation schemes, and user interaction patterns (Nguyen et al., 2024, Ma et al., 2024).

A representative research framework involves:

• Intrinsically explainable or post-hoc-enhanced foundational models (e.g., decision trees, neural networks with built-in attribution).
• Human-centered explanation layers that select, adapt, and present explanations according to user expertise and cognitive load.
• Dynamic feedback loops wherein real-time user interaction is monitored, and explanations are refined accordingly, supporting continuous learning and human-AI mutual calibration (Silva et al., 14 Apr 2025, Paraschou et al., 13 Jun 2025).

Social Transparency and Sociotechnical Context

The Social Transparency (ST) framework, now widely cited in HCXAI, expands the scope of explanations to the socio-organizational domain. It addresses questions such as:

1. Who?—Which actors were involved in each AI-mediated decision.
2. What?—What actions those actors undertook.
3. When?—Temporal sequencing of actions.
4. Why?—Underlying motives and goals.

Recent work extends this to a 5W model, introducing "Which?"—clarifying which social attributions (roles, personas) are justified for an AI system in context, and which are being assigned by users. This addition targets the risk of social misattribution, particularly acute with LLMs (Ferrario et al., 2024).

Taxonomies of Explanation and Evaluation

HCXAI taxonomies classify explanations by:

• Modality: graphical (saliency maps), numerical (feature weights), textual (rationales).
• Scope: local (instance-level), global (model-level).
• Approach: causal, contrastive/counterfactual, example-based, feature-based, rule-based.
• Interactivity: static vs. interactive, enabling exploratory or demand-driven querying (Mangold et al., 14 Oct 2025, Ma, 2024).

Evaluation metrics are similarly categorized into affective (trust, satisfaction), cognitive (mental model, understandability, task load), usability (ease of use, effectiveness), interpretability (transparency), and explanation-specific metrics (usefulness, satisfaction with the explanation) (Mangold et al., 14 Oct 2025, Rong et al., 2022).

3. Methodological Innovations

Participatory, User-Centered Design

HCXAI emphasizes front-loaded, participatory design methodologies:

• Segmentation and identification of target users (by expertise, domain, or role).
• Contextual inquiry, interviews, and field studies to elicit user goals, tasks, mental models, and explanation needs.
• Persona-based frameworks that reflect characteristic user archetypes (e.g., power-users, casual users, privacy-oriented users) (Weitz et al., 2022).
• Iterative prototyping and user testing, with feedback integrated throughout the system's lifecycle (Nguyen et al., 2022, Ma et al., 2024).
• Mixed-methods evaluation combining quantitative metrics (task accuracy, trust ratings, workload) and qualitative insights (thematic analysis, think-aloud protocols).

Domain Knowledge Integration and Personalization

Advanced HCXAI systems are built to ground explanations in explicit domain and organizational knowledge. Techniques include:

• Mapping output features to knowledge graphs, ontologies, or rule-sets representing expert-validated domain concepts (Sheth et al., 2021).
• Tailoring explanations dynamically to individual user profiles and contexts using explicit personalization functions (e.g., $P(u, c) \to$ explanation variant), and employing progressive disclosure (layered access to explanatory depth) (Meske et al., 8 Aug 2025).
• Supporting both global and local explanation needs, including "What-if" counterfactuals and examples relevant to user tasks.

Adaptive Presentation and Interaction

Emerging best practices for explanation interfaces include:

• Multi-modal presentation (text, graphics, narrativized rationales).
• Adaptive complexity modulation based on cognitive load and user expertise ($D(e, \ell)$—selecting summary or deep-dive presentation) (Silva et al., 14 Apr 2025).
• Interactive components (sliders, Q&A dialogues, example exploration), empowering users to initiate explanation requests as needed (Weitz et al., 2022, Nguyen et al., 2024).
• Social-contextual overlays (social role warnings, attribution dashboards) to calibrate trust and clarify intended vs. perceived AI roles (Ferrario et al., 2024).

4. Evaluation Paradigms and Empirical Findings

Human-Grounded and Application-Grounded Evaluation

Most HCXAI user studies adopt either:

• Human-grounded protocols (simplified tasks, generic participants) for systematic effect estimation and prototype selection.
• Application-grounded protocols (real tasks, domain experts) for ecologically valid assessment in high-stakes domains such as healthcare, finance, and education (Gambetti et al., 14 Feb 2025, Rogha, 2023).

Quantitative measures include observed trust (agreement rates, over/under-reliance), user understanding (forward simulation, feature-importance identification), usability (NASA-TLX, task completion time), fairness (demographic parity, equalized odds), and actionability (decision quality, willingness to act on AI recommendations) (Rong et al., 2022, Ma et al., 2024). Subjective metrics leverage validated scales: Likert ratings on trust, explanation satisfaction, transparency, and fairness (Mangold et al., 14 Oct 2025).

Empirical Synthesis

Empirical research consistently demonstrates:

• Moderate explanation granularity and progressive disclosure aid user understanding.
• Tailoring explanation style and complexity to user expertise and goals improves satisfaction and calibration of trust.
• Rich interactivity (on-demand Q&A, live simulations) increases engagement, particularly for detail-oriented and expert users.
• Cognitive overload and information bloat remain risks if explanation interfaces are insufficiently adaptive or over-detailed.
• No single explanation paradigm fits all users or tasks; personalization, inclusivity, and context-awareness are essential for actionable, trustworthy HCXAI systems (Ma, 2024, Weitz et al., 2022).

5. Socio-Technical and Ethical Considerations

HCXAI foregrounds sociotechnical dynamics and ethical imperatives:

• Embedding value-sensitive design, participatory processes, and ongoing reflection into system development.
• Addressing social misattributions—risks that arise when users ascribe inappropriate social roles or agency to LLMs—by integrating explicit social-transparency mechanisms (such as 5W frameworks) (Ferrario et al., 2024).
• Ensuring fairness in both technical and social senses through human-in-the-loop auditing, multi-level group fairness metrics, and participatory definition of fairness criteria (Ghai, 2023).
• Supporting accountability, privacy, and compliance by making not only model operations but also data provenance and decision justifications auditable and accessible to diverse stakeholders.
• Avoiding over-reliance and automation bias by exposing model uncertainty and limitations, surfacing counterfactuals, and allowing user intervention at critical junctures (Rogha, 2023, Silva et al., 14 Apr 2025).

6. Open Challenges and Prospects

Key open research challenges identified across recent literature include:

• Achieving scalable, domain-general methods for explanation personalization, adaptive complexity management, and inclusivity across cultures, languages, and user abilities (Meske et al., 8 Aug 2025).
• Formalizing social attribution models and building robust taxonomies of allowable system roles—especially as LLMs grow more flexible and capable of simulating human agency (Ferrario et al., 2024).
• Bridging the gap between algorithmic transparency and meaningful, task-relevant, contextualized sense-making—especially in high-stakes and regulated environments (Maity et al., 2024, Gambetti et al., 14 Feb 2025).
• Developing standardized evaluation frameworks, benchmarks, and documentation protocols to accelerate cumulative empirical insight and enable reproducible research (Ma et al., 2024, Ma, 2024).
• Architecting mixed-initiative, lifelong HCXAI systems capable of evolving with users, organizational practices, and societal norms over time (Ehsan et al., 2020, Mangold et al., 14 Oct 2025).

A plausible implication is that the continued convergence of foundational XAI methods, adaptive interface and interaction paradigms, and deep human-participatory processes will be central to realizing the ambitions of HCXAI for trustworthy, fair, and contextually resonant AI systems across domains.