Contestable Interpretation Interface (CII)

• CII is a structured interface that enables stakeholders to contest AI outcomes by linking explainability with actionable recourse.
• It integrates by-design and post-hoc mechanisms, such as intuitive dashboards, what-if simulators, and audit trails to support transparent human–AI interactions.
• Empirical evidence across domains like healthcare and finance shows that CIIs improve contestability metrics and regulatory compliance.

A Contestable Interpretation Interface (CII) provides a structured, multidimensional environment that empowers stakeholders to interrogate, challenge, and induce change in the outputs or behavior of AI systems. CIIs formalize the link between explainability and actionable recourse, embedding both proactive and reactive mechanisms into the human–AI interaction cycle. Central to this paradigm is the transformation of regulated, theoretical notions of contestability into operational interface, architectural, and workflow patterns, as instantiated across domains such as healthcare diagnostics, credit scoring, and collaborative content production (Moreira et al., 2 Jun 2025).

1. Formal Definitions and Theoretical Foundations

Contestability in explainable AI is defined as a multidimensional, dynamic property of an AI system $\mathcal{A}$ that (a) empowers stakeholders $S$ to challenge individual decisions $\mathcal{D}$ or the system’s future behavior, and (b) couples explainability $\mathcal{E}$ to actionable recourse $\mathcal{C}$ (Moreira et al., 2 Jun 2025). Formally, three levels are distinguished:

• Explanation-Level Contestability (XLC):

$$\text{Contest}_{xlc}(\mathcal{E}) \Leftrightarrow \forall d\in \mathcal{D},\, \forall s\in S,\, \exists r = \mathcal{E}(d) \in \mathcal{R}: r \rightarrow c \in \mathcal{C},\, \text{given}~\mathcal{P}(s)$$

• System-Level Contestability (SLC):

$$\text{Contest}_{slc}(\mathcal{A}) \Leftrightarrow \exists \mathcal{E},\mathcal{C}~\text{s.t.}~\forall s\in S,\, \exists x\in\mathcal{X},\, d=\mathcal{A}(x),\, r=\mathcal{E}(d),\, c\in\mathcal{C}: \phi(c,d,\mathcal{A})=\text{True}$$

• Aggregate Contestability:

$$\text{Contest}(\mathcal{A}) = \alpha\,\text{Contest}_{xlc}(\mathcal{E}) + \beta\,\text{Contest}_{slc}(\mathcal{A}) + \gamma\,\min_{s\in S}SR(\mathcal{A},s)$$

Here $SR(\mathcal{A},s)$ is the stakeholder-specific contestation success rate, and $\alpha+\beta+\gamma=1$. This formalism aligns operationalization with regulatory mandates and diverse stakeholder requirements (Moreira et al., 2 Jun 2025).

Related frameworks augment this foundation with community-driven value pluralism, framing the CII as an infrastructure supporting explicit, machine-interpretable “value profiles” and participatory conflict resolution (Mayer, 7 Jul 2025). Computational argumentation further grounds the explain–contestation–revision cycle in dynamic argumentation frameworks where contestation, dialogue, and model revision are formally modeled and maintained (Leofante et al., 2024).

2. Architectural Models and Interface Layers

Every CII implements a layered architecture separating by-design (proactive) and post-hoc (reactive) mechanisms, each mediating between users, explanation engines, and contestation workflows (Moreira et al., 2 Jun 2025):

A. By-Design Mechanisms ($S$0):

• Native model introspection (e.g., live feature-weight sliders, editable rules panels)
• Coupled explanation-recourse widgets (“what-if” simulators, counterfactual explorers)
• Real-time logging, override hooks, role-based transparency dashboards

B. Post-Hoc Mechanisms ($S$1):

• Embedding external explanation modules (e.g., SHAP, LIME) with “why” and “how-to-appeal” links
• Third-party appeal channels, upload/documentation tools, retrospective analytics dashboards

Community-Defined Value Profile Integration (Mayer, 7 Jul 2025):

CIIs built atop value pluralism model user/community preferences as tuples $S$2 with explicit rule aggregates and meta-rule enforcement engines. Profile activation, weighting, and forking mechanisms allow dynamic modulation of system behavior to reflect plural community priorities.

Computational Argumentation Engine (Leofante et al., 2024):

CIIs structured as dynamic argumentation frameworks deploy modules for explanation extraction, argument storage, contestation ground transformation, dialogue management (FIPA-style acts), and model revision. The system recomputes argument semantics after each contestation, maintaining a coherent, auditable contest–explain–revise protocol.

3. Interaction Patterns and Human-Centered Design

CII interaction patterns are distinguished by highly structured, transparent user workflows (Moreira et al., 2 Jun 2025):

• Challenge Buttons: Directly attached to every explanation, labeled for contestation
• Guided Feedback Wizards: Context-sensitive forms for contestation, document upload, and escalation path selection
• Layered Visualization: Summary explanations, drill-down feature contribution charts, “actions you can take,” and audit timelines
• Conflict Resolution Panels: (in CDAVP-based CIIs) Drag-and-drop profile activation, weight sliders, real-time conflict inspectors, and refer-to-community escalation

Seamful transparency is realized through continuous notifications of ambiguity, “undo/history” controls, and accessible deliberation threads. Human-centered features (multilingual UI, context-dependent wizards, explicit non-retaliation guarantees) maximize stakeholder ease-of-contestation and align with procedural justice requirements.

4. Quantitative Metrics and Evaluation

CIIs are systematically evaluated using the Contestability Assessment Scale (CAS), a continuous composite metric in $S$3 aggregating eight core properties (Moreira et al., 2 Jun 2025):

Property	Range	Weight ($S$4)	Scoring
Explainability	0–2	0.30	None, post-hoc, intrinsic
Openness to Contestation	0–2	0.12	None, expert, all
Traceability	0–10	0.12	5 sub-criteria
Built-in Safeguards	0–1	0.12	Non-retaliation, fail-safe
Adaptivity	0–2	0.10	Static, reactive, proactive
Auditing	0–2	0.10	None, internal, external
Ease of Contestation	0–10	0.07	Checklist: UI, language, access
Explanation Quality	10–50	0.07	System Causability Scale

Higher CAS reflects robust interpretability (properties 1,3,8) and strong actionable recourse (properties 2,4–7) (Moreira et al., 2 Jun 2025). Application to domains such as radiology AI, credit scoring, and recommendation systems shows that even minimally improved contestability mechanisms significantly boost CAS, especially in high-risk settings.

5. Practical Implementations and Empirical Evidence

Case studies reveal the operational impact and adaptation of CIIs:

• Radiology AI (High Risk): Initial CAS ≈ 0.55; MF improvements (multilingual, audit-log access, co-design) raised this to ≈ 0.93 (Moreira et al., 2 Jun 2025).
• Credit Scoring (Medium Risk): CAS increase from ≈ 0.44 to 0.85 after adding what-if simulators, rights notifications, non-retaliation guarantees, and external audit pipelines.
• Collaborative Report Editing (CDAVP-based CII): Journalists with conflicting value profiles co-authoring a report reach conflict resolution through weighted meta-rule enforcement and real-time UI-guided deliberation, with all decisions and actions logged for traceability (Mayer, 7 Jul 2025).

Domain-specific CII implementations (psychiatric diagnosis, Parkinson's care) integrate multi-modal explanation discrepancy metrics, contestable LLM modules, and structured user-contestation channels, further evidencing measurable improvements in interpretability, correction rates, and regulatory alignment (Nguyen et al., 16 May 2025, Nguyen et al., 21 Oct 2025, Nguyen et al., 30 Jul 2025).

6. Integration into Technical, Organizational, and Regulatory Workflows

Comprehensive CII adoption requires multi-layered embedding:

• Technical: Early instrumentation for logging, explainability microservices, contestation APIs, and workflow engines for appeals and retraining (Moreira et al., 2 Jun 2025).
• Organizational: Assigned roles (Liaison Officer, Review Committee), cross-functional oversight boards, and continuous monitoring (monthly CAS re-assessment, low-resource user testing).
• Legal/Compliance: Embedding of policy clarifications and non-retaliation clauses, automated mapping of CAS results to regulatory checklists (e.g., EU AI Act Annex III), and engagement of independent external auditors for review.

Role-sensitive feedback, immutable audit trails, and adaptive justifications—adjusted to context and user persona—support both procedural justice and external review (Nguyen et al., 30 Jul 2025).

7. Research Significance and Ongoing Developments

CIIs provide a rigorously defined, operational blueprint for embedding contestability into AI systems, directly addressing regulatory and ethical imperatives for transparency, recourse, and stakeholder agency (Moreira et al., 2 Jun 2025). Emergent research continues to explore participatory value pluralism, dynamic argumentation, and seamless integration of contestability mechanisms into broader socio-technical ecosystems (Mayer, 7 Jul 2025, Leofante et al., 2024).

Notable areas for further work include: in-clinic and field usability studies, enrichment of domain-specific justification engines, expansion to low-resource, non-technical stakeholder populations, and integration with standardized data systems (e.g., EHR via SMART on FHIR) (Nguyen et al., 30 Jul 2025). The formal underpinnings, modular architecture, and empirical evidence base of CIIs establish a foundational reference for both practical deployments and future research on contestable, accountable AI.