Craft-Based Explainability in AI

• Craft-based explainability is a context-driven, multidisciplinary framework that customizes AI explanations to meet specific audience, regulatory, and impact needs.
• It integrates diverse methodologies such as post hoc and by-design techniques to balance technical accuracy with operational and legal demands.
• The approach emphasizes cost-benefit analysis, addressing design trade-offs to ensure explanations are both effective and societally responsible.

A craft-based approach to explainability in AI situates the design, interpretation, and delivery of explanations as a context-sensitive, multidisciplinary, and iterative “craft” rather than as a fixed, universal procedure. This perspective views explainability as a set of expert practices, blending technical methods with legal, economic, operational, and user-centered considerations to produce explanations that are tailored to specific real-world settings. Rather than a one-size-fits-all solution, craft-based explainability emphasizes a pragmatic balance between system complexity, stakeholder needs, regulatory demands, and the tangible societal benefits and costs of providing explanations.

1. Contextual Drivers of Explainability

A foundational tenet of the craft-based approach is that explainability must be defined by the context in which an AI system operates. The required form and depth of explanation are governed by four principal factors:

• Audience/Recipient Factors: The expertise, roles, and informational needs of the explanation’s recipients significantly influence explanation design. Technical staff (e.g., auditors, data scientists) may require detailed, algorithmic explanations, while non-specialist operators or regulators may need simplified summaries or user-focused rationales.
• Impact Factors: The level of harm that an AI system could cause in erroneous or unexpected scenarios determines the required robustness and granularity of explanations. Systems in safety-critical domains (e.g., autonomous vehicles, medical diagnostics) necessitate more rigorous, auditable explanations than low-stakes applications.
• Regulatory Factors: Legal and policy frameworks—such as the EU's GDPR or country-specific algorithmic accountability laws—may stipulate minimum explainability requirements, specifying both what must be disclosed and to whom. Courts may mandate disclosure of “parameters and their weights,” decision rules, or the logic underlying automated decisions.
• Operational Factors: The practical role of the system (decision support vs. automation), certification requirements, and degree of human–machine partnership all modulate the type and scope of required explanations.

Contextual awareness thus underpins the craft-based approach: explanations are “crafted” to fit the sociotechnical realities and constraints of particular deployments (Beaudouin et al., 2020).

2. Explainability Methodologies and Tools

A spectrum of technical tools is integrated within the craft-based framework, categorized by their relation to the model lifecycle and degree of “by-design” explainability:

• Post Hoc Approaches:
  • Perturbation-based: Tools like LIME perturb input features, fitting sparse surrogate models (e.g., linear or decision tree) in the local neighborhood of the instance to be explained. The optimization typically minimizes a loss of the form:

  $\min_g \mathcal{L}(f, g, \pi_x) + \Omega(g)$

  where $\mathcal{L}$ measures fidelity to $f$ (the black-box model) in the locality defined by $\pi_x$, and $\Omega(g)$ penalizes model complexity. - Model-agnostic attribution: KernelSHAP guarantees additive explanations consistent with properties derived from Shapley values. - Saliency methods: Sensitivity analyses (e.g., gradients, guided backpropagation, Grad-CAM, SmoothGrad) visualize critical regions or features in the input.

• By-Design and Hybrid AI:

  • Objective modification: Augmenting loss functions to promote sparse, stable representations (e.g., spatially localized filters, game-theoretic regularizations).
  • Predictor modification: Contextual Explanation Networks (CENs) and Self-Explanatory Neural Networks (SENN) embed human-interpretable proxies directly in the model decision pipeline.
  • Symbolic–statistical hybridization: Integrate domain rules and constraints (e.g., rules-to-network frameworks) to increase explicitness and verifiability.

These methodologies are not exclusive but are weighed according to contextual, legal, and operational requirements within the craft-based decision process (Beaudouin et al., 2020).

3. Organizational and Design Frameworks

The development of explainability within a craft-based paradigm is operationalized using structured multi-phase frameworks:

• Context Characterization: Elicit and document all relevant audience, regulatory, impact, and operational needs. This step grounds the explanation in its application environment.
• Technical Assessment: Survey and select from available post hoc, hybrid, and by-design methods. Evaluate their suitability for both global explanations (model-wide logic) and local explanations (instance-specific reasoning).
• Output and Cost-Benefit Calibration: Decide the levels of explanation to be delivered—balancing global and local fidelity, interpretability, and stakeholder requirements—while explicitly considering the costs associated with explanation production (design, accuracy trade-offs, audit log creation, IP violation, security, loss of flexibility, and innovation slowdown).

This methodological sequence acknowledges that explanations must meet technical standards, legal obligations, and economic constraints—thus forming a context-specific, justifiable rationale rather than a generic obligation.

Phase	Key Output	Principal Stakeholders
Context	Audience, harm, regulatory, operational	Domain experts, legal
Technical	Candidate methods for global/local explain	Engineers, scientists
Output/Cost	Calibrated explanations, justified costs	Management, regulators

4. Socio-Economic Cost-Benefit Analysis

A distinguishing feature of the craft-based approach is explicit recognition of the multifaceted costs and societal benefits connected to explainability:

• Design costs: Developing explanations tailored to use-case and regulatory specifics can be resource-intensive.
• Prediction-accuracy tradeoff: Interpretable models or added constraints can result in (potentially) lower accuracy; in critical domains, this may impact societal outcomes.
• Auditability costs: Storage and retrieval of decision records create operational and sometimes privacy burdens.
• Trade secret/confidentiality: Disclosure of methodological details may undermine proprietary or competitive standing.
• Security and flexibility: Detailed transparency may facilitate adversarial attacks or unduly constrain system adaptability over time.
• Slowed innovation: Overly demanding requirements for explicability can hinder agile development and innovation.

Social benefits—greater trust, accountability, improved system auditing, fairness, and compliance—must demonstrably exceed these costs for an explanation strategy to be warranted and justified (Beaudouin et al., 2020).

5. Applications and Implications in High-Stakes Domains

The craft-based approach underpins explainability design decisions in a range of real-world, high-stakes, or regulated environments:

• Safety-Critical Systems: Autonomous vehicles, medical devices, or anti-fraud platforms require contextualized explanations capable of supporting forensic analysis, certification, and error investigation.
• Regulatory Compliance: Legal interpretations (e.g., GDPR, French transparency laws) guide the disclosure forms and inform cost-benefit trade-offs between explanations and proprietary information.
• Risk Management, Certification, Oversight: Organizations rely on explainability strategies derived from the framework to structure impact assessments, compliance reviews, and ongoing oversight (e.g., independent review committees).
• Flexible Design in Lower-Risk Settings: Lesser-explained outputs may be justified in low-stakes systems, guided by contextual assessments to ensure that required trust and safety levels are nonetheless achieved.

This flexible, context-driven approach avoids both under- and over-disclosure, matching the explanation burden to stakeholder needs and the societal importance of the AI’s decisions (Beaudouin et al., 2020).

6. Future Directions and Systemic Implications

Craft-based explainability emphasizes the need for integrated, multidisciplinary development processes. It encourages ongoing engagement among technical teams, legal specialists, operational stakeholders, and broader society. Research and deployment in this paradigm will increasingly focus on:

• Automated tools for context elicitation and requirement traceability
• Multi-stakeholder participatory design processes
• Dynamic cost-benefit modeling and impact assessment
• Evolving technical toolkits capable of producing explanations at varying granularity and for heterogeneous audiences

By committing to ongoing refinement and feedback within specific operational and regulatory contexts, the craft-based approach seeks not only technically sound and auditable systems, but also AI deployments that are trusted, fair, accountable, and societally legitimate.