Value-Sensitive Design (VSD)

• Value-Sensitive Design is a framework that systematically integrates human values into technology via iterative conceptual, empirical, and technical investigations.
• It translates abstract values like privacy and fairness into concrete design requirements, ensuring stakeholder priorities guide system behavior.
• Innovative methods such as participatory design fictions and multi-objective optimization foster transparent and ethical AI system development.

Value-Sensitive Design (VSD) is a theoretically grounded and methodologically rigorous framework for integrating human values systematically into the design, development, and deployment of technology. Originally articulated by Batya Friedman in 1996, VSD has since evolved to encompass a tripartite, iterative methodology that engages stakeholders throughout conceptual, empirical, and technical investigations, ensuring that ethically salient and domain-specific values are operationalized from the earliest stages of system conception through iterative refinement and evaluation (Liao et al., 2019).

1. Core Principles and Methodological Structure

VSD is defined as “a theoretically grounded approach to the design of technology that accounts for human values in a principled and comprehensive manner throughout the design process” (Liao et al., 2019). The standard VSD process in HCI and AI is instantiated via three interlocking, iterative types of inquiry:

• Conceptual investigations: Identify which values are at stake, whose values these are (direct and indirect stakeholders), and how these values should be interpreted in a given socio-technical context.
• Empirical investigations: Elicit, prioritize, and observe how stakeholders experience and negotiate identified values via qualitative and quantitative methods, including interviews, focus groups, participatory workshops, surveys, and scenario-based exercises.
• Technical investigations: Map values onto design requirements, prototype system features, and analyze whether proposed or implemented technical solutions fulfill those values or introduce new value tensions.

This triadic process is inherently iterative, supporting cycling between phases as new value tensions or implementation barriers are surfaced. VSD addresses both general ethical values (e.g., privacy, autonomy, accountability, fairness) and classical software/system requirements (e.g., usability, reliability) (Aizenberg et al., 2020).

2. Value Identification and Translation: From Abstractions to Design Requirements

A hallmark of rigorous VSD is explicit translation from abstract values to operational system norms and concrete design requirements (Aizenberg et al., 2020). Van de Poel’s values hierarchy is widely adopted for this purpose, structuring:

• Values (Level 1): High-level abstractions such as privacy, autonomy, dignity, fairness, trust.
• Norms (Level 2): Contextualized rules or constraints, such as data minimization, informed consent, or contestability, specifying how values are to be realized in a particular domain.
• Design requirements (Level 3): Socio-technical features, workflows, interface components, or algorithmic constraints that instantiate norms (e.g., opt-in consent flows, auditability modules, explainable outputs).

Values are connected to norms and requirements through “for the sake of” relationships (means-end, obstacle removal, sub-goal, or enabler), creating traceability between stakeholder priorities and technical implementation (Aizenberg et al., 2020).

3. Methodological Innovations: Participatory Design Fictions and Value Elicitation

VSD in AI has been augmented with participatory and narrative methods to overcome the epistemic limits of both top-down (deductive, theory-driven) and bottom-up (data-driven, behavioral) value alignment. Notably:

• Participatory Design Fictions (PDF): In this approach, researchers construct “strategically incomplete” fictional scenarios centered around emerging technology artifacts (e.g., an AI-powered nanny bot). Stakeholders are then invited to complete or critique these scenarios, surfacing values, priorities, and trade-offs often invisible to either top-down theorizing or large-scale behavioral aggregation (Liao et al., 2019). Transcribed continuations, artifacts, and group discussions are coded—typically via Grounded Theory—to extract value categories and map them to domain constraints.
• Techniques for Eliciting Values: Using unresolved narrative decision points to force explicit articulation of trade-offs, framing stories in familiar contexts, and leveraging co-creation tools (games, online platforms) to make collective value dynamics observable (Liao et al., 2019). Both direct statements and narrative cues (implicit preferences, hesitations) are systematically analyzed.

A prototypical PDF process includes stakeholder mapping, narrative crafting, diverse engagement (individual, group, performative), iterative scenario refinement, and translation into technical design constraints or reward functions for AI systems.

4. Operationalizing Values in AI System Design

VSD’s translation pipeline manifests concretely in AI and machine learning via:

• Multi-objective optimization: Turning each stakeholder/ethical desideratum into a separable objective function; e.g., in multi-objective recommenders balancing personal values (taste, cost, nutrition) and societal imperatives (environmental impact, fairness). Optimization is over a Pareto front, permitting user-centered negotiation of explicit trade-offs (Asikis, 2021).
• Reward-engineering and Inverse Reinforcement Learning (IRL): Narrated or empirical stakeholder choices are mapped to probabilistic reward functions, e.g., reward(s,a) ∝ P(a|s;θ), supporting value learning from qualitative or semi-structured data (Liao et al., 2019).
• Transparency and autonomy: Production and presentation of plural, non-dominated system outputs (e.g., alternate recommendations or decisions) expose consequences, preserving autonomy and supporting informed consent (Asikis, 2021).

The preferred evaluation is not a single aggregated metric, but a set of outcome ratios (e.g., cost, similarity, environmental impact), with stakeholder agency in trade-off selection.

5. Case Studies and Domain-specific Extensions

VSD, extended via participatory and empirical methods, has been applied across a range of socio-technical contexts:

• AI for childcare (“Nanny Bot”): Stakeholder engagement surfaced both classic and domain-specific values (e.g., tradition, child’s social development) and revealed significant inter-individual variation in privacy/autonomy trade-offs. Creative, hybrid solutions (supplemental human care for premium experiences) emerged (Liao et al., 2019).
• Sustainable consumption recommendation: System objectives directly instantiate mapped values; diverse Pareto-optimization algorithms (RNSGA-II, MO-NES, G3A) allow users to choose among alternatives according to per-individual value priorities—demonstrating clear sustainability gains with partial adoption (Asikis, 2021).
• Justice in AI for child welfare: “Value source analysis” draws justice principles directly from policy texts through mixed-methods topic modeling and inductive coding, translating them into actionable socio-technical design requirements (e.g., language accessibility, well-being optimization, harm minimization, explicit equity audits) (Rodriguez et al., 11 Nov 2025).

In all cases, iterative, feedback-driven refinement is emphasized. Value hierarchies are revisited as new empirical or technical findings arise.

6. Open Challenges and Future Directions

Several open research problems and methodological questions persist:

• Metric formalization: Translating qualitative value tensions and trade-offs from narrative, stakeholder input, or participatory workshops into operational system constraints and machine-interpretable reward functions remains unresolved. Proposals include integration with topic modeling, grounded theory coding, and advanced IRL (Liao et al., 2019).
• Scalability and representativeness: As value diversity and narrative variety scale in large-population or multi-stakeholder settings, efficient aggregation, outlier detection, and representativeness checking surface as technical bottlenecks (Liao et al., 2019).
• Digital and online platforms: Harnessing digital games or co-design platforms for large-scale, emergent value elicitation is under-explored.
• Continuous, co-adaptive value learning: Methods for closing the loop between periodic participatory value elicitation and live model updating (e.g., via cooperative IRL) are largely prospective.

Fundamentally, VSD’s animating concern remains: whose values are represented, and how can technical and participatory methods be iteratively structured to guarantee that indirect, marginalized, or underrepresented stakeholder values directly condition system behavior?

7. Guidelines and Best Practices for AI Practitioners

Best-practice VSD implementation in AI contexts, as distilled from leading papers, is summarized as follows (Liao et al., 2019):

1. Conduct granular stakeholder mapping, with explicit recognition of indirect and marginalized actors.
2. Construct narrative probes that lay bare unresolved, technology-specific value tensions.
3. Employ multiple, complementary engagement formats (interviews, narrative writing, co-design workshops); favor open, non-leading, value-oriented prompts.
4. Iteratively loop between broad, exploratory conceptual/empirical phases and more focused, trade-off probing narratives.
5. Utilize rigorous, mixed qualitative–quantitative analysis (e.g., Grounded Theory coding, topic modeling, reward function learning) to distill and weight emergent value categories.
6. Translate elicited values to direct system constraints and emergent preference structures; implement and expose multiple alternatives.
7. Structure scalability via digital participatory infrastructure.
8. Treat all narrative data and artifacts as reusable probes in future system iterations.
9. Routinely audit representation, update narratives, and refocus engagement as systemic values, domains, or user populations shift.

Embedding these techniques into the AI development cycle operationalizes the aspiration of VSD: aligning complex AI system behavior with the evolving, context-specific, and heterogeneous fabric of human values (Liao et al., 2019).