Flexible and Context-Specific AI Explainability: A Multidisciplinary Approach

Valérie Beaudouin, Isabelle Bloch, David Bounie, Stéphan Clémençon, Florence d'Alché-Buc, James Eagan, Winston Maxwell, Pavlo Mozharovskyi, Jayneel ParekharXiv 2020 · arXiv

Explainability in Artificial Intelligence (AI) is often treated as a purely technical challenge—a feature to be added to a model after training. However, as deep learning models increasingly govern safety-critical systems like autonomous vehicles and medical diagnostics, this siloed approach is no longer sufficient. This paper argues that explainability is a multidisciplinary balancing act that must reconcile technical feasibility with legal requirements, audience needs, and economic costs.

The Problem: Siloed Explainability

Current approaches to Explainable AI (XAI) are often fragmented. Computer scientists focus on technical tools like saliency maps, while legal experts focus on regulatory compliance (e.g., GDPR). These fields rarely communicate, leading to a "black box" crisis where the provided explanations may not meet the needs of the actual recipient. A data scientist requires a different level of detail than a patient or a regulator. Furthermore, explainability is not "free"; it introduces significant overheads in terms of model performance, data storage, and intellectual property risks.

The Three-Pillar Framework

To bridge these gaps, the authors propose a unified methodology to determine the optimal level of transparency for any given system. This framework moves away from a one-size-fits-all model toward a context-specific strategy.

The Three-Pillar Framework for AI Explainability

1. Context Factors

Before implementing technical solutions, developers must define the operational environment. This involves identifying the recipient (their expertise and time), the impact (potential for harm), the regulatory framework, and operational imperatives like safety certifications.

Key Contextual Factors Influencing Explanation Requirements

2. Technical Solutions

The framework categorizes tools into two primary buckets:

  • Post-hoc Approaches: Methods such as LIME, Kernel-SHAP, and saliency maps that attempt to explain a model's logic after it has been trained.
  • Hybrid AI (Explainability by Design): This more robust approach incorporates symbolic logic or constraints directly into the model architecture, such as genetic fuzzy logic or modifying objective functions to prioritize interpretable features.

3. Explainability Parameters

Finally, the developer selects the output format. This includes choosing between Global Explainability (how the entire system functions) and Local Explainability (why a specific decision was made), as well as determining the necessary level of detail and data disclosure.

The Economics of Explainability

A core contribution of this paper is the identification of the "Seven Costs of XAI," which are often overlooked in technical literature:

  1. Design Costs: Customizing XAI for specific contexts.
  2. Reduced Accuracy: The inherent trade-off between interpretability and performance.
  3. Storage Overhead: The massive data requirements for maintaining audit logs and traceability.
  4. Trade Secret Exposure: The risk of revealing proprietary logic through transparency.
  5. Security Risks: Making it easier for adversarial actors to "game" or exploit the system.
  6. Decisional Flexibility: The risk that explanations set precedents that "trap" future decision-makers.
  7. Innovation Drag: The overhead of compliance slowing down development cycles.

Conclusion: Social Utility and the Sliding Scale

The paper concludes that the need for explainability exists on a "sliding scale" proportional to the severity of a system's consequences. While a smart toaster requires minimal transparency, an autonomous vehicle requires a high degree of explainability to ensure safety and trust. Ultimately, an explanation is only socially useful when its total benefits—trust, safety, and fairness—outweigh the social and economic costs of its implementation. This framework provides a roadmap for the "AI impact assessments" likely to become mandatory under future regulations like the EU AI Act.

Summary updated

Abstract

The recent enthusiasm for artificial intelligence (AI) is due principally to advances in deep learning. Deep learning methods are remarkably accurate, but also opaque, which limits their potential use in safety-critical applications. To achieve trust and accountability, designers and operators of machine learning algorithms must be able to explain the inner workings, the results and the causes of failures of algorithms to users, regulators, and citizens. The originality of this paper is to combine technical, legal and economic aspects of explainability to develop a framework for defining the "right" level of explain-ability in a given context. We propose three logical steps: First, define the main contextual factors, such as who the audience of the explanation is, the operational context, the level of harm that the system could cause, and the legal/regulatory framework. This step will help characterize the operational and legal needs for explanation, and the corresponding social benefits. Second, examine the technical tools available, including post hoc approaches (input perturbation, saliency maps...) and hybrid AI approaches. Third, as function of the first two steps, choose the right levels of global and local explanation outputs, taking into the account the costs involved. We identify seven kinds of costs and emphasize that explanations are socially useful only when total social benefits exceed costs.
BibTeX
@misc{beaudouin2020flexible,
  author = {Valérie Beaudouin and Isabelle Bloch and David Bounie and Stéphan Clémençon and Florence d'Alché-Buc and James Eagan and Winston Maxwell and Pavlo Mozharovskyi and Jayneel Parekh},
  title = {{Flexible and Context-Specific AI Explainability: A Multidisciplinary Approach}},
  year = {2020},
  eprint = {2003.07703},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2003.07703},
}
Figure