Explainability (OECD)

The OECD AI Principle (1.3) requiring that people affected by AI-based outcomes are able to understand how and why particular decisions or recommendations were reached. Explanations must be contextually appropriate, enable meaningful contestation, and illuminate causal factors—operationalised through LIME, SHAP, attention visualisation, and inherently interpretable model architectures.

Semantic Classification

Content

• People affected by AI-based outcomes should be able to understand how and why particular decisions or recommendations were reached, with explanations provided in ways appropriate to the context and enabling meaningful contestation of AI-influenced decisions.

Wearables

Source

• Primary: OECD AI Principles 2024 revision (Principle 1.3)

• Related: GDPR Article 22, EU AI Act Article 13

  Context

  Explainability represents the second component of OECD Principle 3, addressing the critical need for people to understand AI system outputs that affect them. This principle recognises that transparency about system existence is insufficient without the ability to comprehend specific decisions.

  Key Characteristics

• Outcome-focused: Explaining specific decisions or recommendations

  • Contextually appropriate: Matched to decision significance and audience

  • Contestability-enabling: Supporting effective challenge of decisions

  • Causal insight: Illuminating what factors influenced outcomes

  • Actionable understanding: Enabling informed responses and recourse

    Relationships

  • Parent Concept: OECD AI Principle 3 (Transparency and Explainability)

  • Related Terms:

    • Transparency (OECD) (AI-0161)
    • Interpretability (source material)
    • XAI - Explainable AI (source material)
    • Accountability (AI-0165)
    • Right to Explanation (source material)

  • Contrasts With: Black-box AI, inscrutable decisions

    Explainability Dimensions

    What-Explanation

  • What decision or output was produced

  • What classification or prediction resulted

  • What actions the system recommends

    Why-Explanation

  • Why this particular outcome occurred

  • What factors most influenced the decision

  • Why alternative outcomes were not chosen

    How-Explanation

  • How the system reached its conclusion

  • What process was followed

  • How inputs were weighted and combined

    What-If-Explanation

  • How outcomes would change with different inputs

  • What thresholds affect decisions

  • What recourse options exist

    Implementation Considerations

    1. Method selection: Choosing appropriate explainability techniques for model types
    2. Audience adaptation: Tailoring explanations to user technical literacy
    3. Contextual significance: Providing deeper explanations for high-stakes decisions
    4. Accuracy-explainability balance: Managing trade-offs in model selection
    5. Evaluation: Testing whether explanations genuinely aid understanding

    OECD Framework Alignment

  • Dimension: Task and Output

  • Principle Number: P3 (part 2 of 2)

  • Actor Responsibility: Providers (design), deployers (provision to affected persons)

    Technical Approaches

    Model-Agnostic Methods

  • LIME (Local Interpretable Model-agnostic Explanations)

  • SHAP (SHapley Additive exPlanations)

  • Counterfactual explanations

  • Sensitivity analysis

    Model-Specific Methods

  • Feature importance (tree-based models)

  • Attention visualisation (transformers)

  • Layer-wise relevance propagation (neural networks)

  • Rule extraction (from neural networks)

    Inherently Interpretable Models

  • Decision trees and rule-based systems

  • Linear models with transparent coefficients

  • Generalised additive models (GAMs)

  • Sparse models with few features

    Explainability Levels

    Global Explainability

  • Overall model behaviour across all inputs

  • Feature importance rankings

  • Decision boundary descriptions

  • General operating principles

    Local Explainability

  • Explanation for specific individual decision

  • Factors influencing this particular outcome

  • Instance-specific feature contributions

  • Nearest decision boundary distance

    Cohort Explainability

  • Behaviour for groups with shared characteristics

  • Sub-population-specific patterns

  • Conditional relationships and interactions

    Regulatory Context

    Explainability principles inform:

  • EU AI Act transparency and explainability requirements (Article 13)

  • GDPR right to meaningful information about automated decision logic (Article 13-15)

  • High-risk AI system human oversight requirements (Article 14)

  • Instructions for use content (Article 13, Annex IV)

    Challenges and Limitations

  • Complexity-explainability trade-off: Most accurate models often least explainable

  • Fidelity: Simplified explanations may not reflect true model behaviour

  • Stability: Explanations can vary for similar inputs

  • Manipulation: Explanations can be gamed to appear reasonable

  • Completeness: Full explanations may be impossibly complex

  • User comprehension: Even provided explanations may not be understood

    Contextual Variation

    Explainability requirements vary by:

  • Decision stakes: Higher stakes demand richer explanations

  • Domain: Medical vs entertainment contexts require different detail

  • User sophistication: Technical vs general public audiences

  • Legal requirements: Regulated sectors may mandate specific explanations

  • Cultural context: Explanation expectations vary across cultures

    2024 Revision Updates

    The 2024 OECD revision strengthened explainability by:

  • Explicitly linking to contestability and challenge rights

  • Emphasising affected person perspective

  • Connecting to meaningful understanding rather than mere information

  • Clarifying context-appropriateness requirements

    Evaluation Metrics

    Explainability effectiveness assessed through:

  • User comprehension testing

  • Actionability of provided information

  • Support for successful contestation

  • Stakeholder satisfaction surveys

  • Task completion with explanations

    Human Factors

    Effective explanations consider:

  • Cognitive load: Avoiding information overload

  • Mental models: Aligning with user expectations

  • Trust calibration: Supporting appropriate trust levels

  • Decision support: Enabling better human judgment

  • Learning: Supporting user skill development over time

    Related Standards

  • ISO/IEC TR 29119-11:2020 - Software testing (AI system testing)

  • IEEE P7001 - Transparency of Autonomous Systems (under development)

  • ISO/IEC 23894:2023 - AI risk management (explainability component)

    See Also

  • Transparency (OECD) (AI-0161)

  • Interpretability (source material)

  • XAI - Explainable AI (source material)

  • Right to Explanation (source material)

  • Human Oversight (EU AI Act)

    ---

    Part of AI Grounded Ontology - OECD AI Principles Framework Aligned with OECD AI Principles 2024, GDPR and EU AI Act requirements

    Updated Ontology Entry: Explainability (OECD)

    Academic Context

• The OECD AI Principles represent the first intergovernmental standard on artificial intelligence, adopted in 2019 and substantially updated in 2023 and 2024[1][2][3]

• Explainability forms one of five core values-based principles alongside inclusive growth, human rights respect, robustness, and accountability

• The principle emerged from recognition that AI systems operate across borders and require international consensus on trustworthy governance

• Explainability specifically addresses the transparency requirement: stakeholders must understand how AI systems operate, with providers disclosing information about data sources, logic, and decision-making processes in context-appropriate ways[1]

• The principle reflects a rights-based approach to AI governance

• Enables users to challenge outputs where needed, supporting meaningful contestation of AI-influenced decisions

• Balances innovation with protection of human rights and democratic values

• Designed to remain flexible and relevant as AI systems continue to evolve post-deployment, particularly with generative AI applications[2]

  Current Landscape (2025)

• OECD framework adoption and influence

• Over 1,000 policy initiatives across more than 70 jurisdictions follow the OECD AI Principles as of May 2023, with continued expansion through 2025[3]

• The framework has significantly influenced landmark regulatory efforts including the European Union’s AI Act and the NIST AI Risk Management Framework[2]

• OECD member countries are expected to actively support these principles and make best efforts to implement them

• Explainability in practice across governance frameworks

• The EU AI Act operationalises explainability through a tiered, risk-based classification system (unacceptable, high, limited, minimal risk), with high-risk systems requiring explicit transparency mechanisms[4]

• UNESCO’s Recommendation on the Ethics of Artificial Intelligence incorporates explainability whilst emphasising environmental sustainability and gender equality[4]

• Common implementation themes include human oversight, transparency, accountability, and proportionality—with oversight corresponding to potential system impact[4]

• Technical implementation considerations

• Context-appropriate disclosure remains challenging; different stakeholders (engineers, product managers, end-users) require tailored explanations[1]

• Organisations must provide ongoing training to navigate the complex, constantly evolving regulatory landscape, with role-specific guidance on model transparency and accuracy[1]

• The principle acknowledges that meaningful explainability requires balancing technical precision with accessibility for non-specialist audiences

• UK and North England context

• The UK has adopted OECD principles within its AI governance framework, though specific North England implementation details remain limited in current policy documentation

• Manchester, Leeds, Newcastle, and Sheffield host significant AI research and development clusters, though formal explainability-focused initiatives specific to these regions are not yet prominently documented in international governance frameworks

• UK organisations increasingly align with OECD standards as baseline expectations for responsible AI deployment

  Research & Literature

• Primary sources

• OECD (2019, updated 2023–2024). Recommendation of the Council on Artificial Intelligence. OECD Legal Instruments. Available at: https://legalinstruments.oecd.org/en/instruments/oecd-legal-0449[2][3]

• OECD (2025). Governing with Artificial Intelligence: AI in Policy Evaluation. OECD Publications[5]

• Framework documentation

• OECD (2024). AI Principles – OECD. Topic overview and adherent countries. Available at: https://www.oecd.org/en/topics/sub-issues/ai-principles.html[3]

• Bradley (2025). Global AI Governance: Five Key Frameworks Explained. Analysis of OECD recommendations and related governance structures[2]

• Implementation guidance

• AI21 Labs (2025). 9 Key AI Governance Frameworks in 2025. Comparative analysis of explainability across frameworks including OECD, EU AI Act, and UNESCO recommendations[4]

  Future Directions

• Evolving technical standards

• Continued refinement of explainability definitions to accommodate generative AI systems and post-deployment model evolution

• Development of standardised metrics for assessing explanation quality and user comprehension across different contexts

• Regulatory harmonisation

• Governments worldwide increasingly adopt OECD definitions and AI system classifications for interoperable governance[2]

• Anticipated convergence between EU, UK, and international frameworks, though jurisdictional variations will persist

• Research priorities

• Empirical studies on effective explanation formats for diverse stakeholder groups

• Investigation of trade-offs between explainability and model performance

• Development of proportionate oversight mechanisms that scale with system risk levels

• Emerging challenges

• Balancing proprietary concerns with transparency requirements

• Ensuring explainability remains meaningful as AI systems become increasingly complex

• Addressing cultural and linguistic variations in what constitutes “context-appropriate” explanation

  ---

  Note: This entry reflects the current state of OECD AI governance as of November 2025. The principles remain non-binding but highly influential, with implementation varying across jurisdictions. The framework continues to evolve in response to technological developments, particularly in generative AI applications.

  Metadata

• Last Updated: 2025-11-11

• Review Status: Comprehensive editorial review

• Verification: Academic sources verified

• Regional Context: UK/North England where applicable

  Source

• Primary: OECD AI Principles 2024 revision (Principle 1.3)

• Related: GDPR Article 22, EU AI Act Article 13

  Context

  Explainability represents the second component of OECD Principle 3, addressing the critical need for people to understand AI system outputs that affect them. This principle recognises that transparency about system existence is insufficient without the ability to comprehend specific decisions.

Provenance