PaperlayerView exact revision on arXiv

Source-linked AI summary

Responsible AI by Design in Practice

Richard Benjamins, Alberto Barbado, Daniel Sierra

arXiv:1909.12838v2cs.CY

TL;DR

Organizations face AI risks such as discrimination and insufficiently understandable decisions, yet practical implementation guidance remains limited. The paper presents Telefónica’s Responsible AI by Design methodology, combining principles, checkpoints, tools, training, and governance. Its experience indicates that integrating these elements with awareness and self-control can reduce risks while supporting organizational adoption.

  • Problem

    Organizations lack practical technical and organizational experience for implementing responsible AI despite broad consensus about its risks.

    Source: Principles of Artificial Intelligence · p. 2

  • Method

    The paper presents Responsible AI by Design as a methodology combining AI principles, checkpoints, tools, training, and governance.

    Source: by Design” · p. 3

  • Results

    Telefónica’s methodology integrates principles, awareness and training, development questions, tools, and governance to address undesired AI side effects.

    Source: Conclusions · p. 8

  • Takeaways & Limitations

    Responsible AI implementation depends on combining organizational practices with technical tools and assigned responsibilities.

    Source: by Design” · p. 3, Conclusions · p. 8

    Explainability lacks a formal common reference, and explanations remain constrained by the features available in supervised-learning datasets.

    Source: Explainability · p. 7

Abstract

from arXiv · show

Recently, a lot of attention has been given to undesired consequences of Artificial Intelligence (AI), such as unfair bias leading to discrimination, or the lack of explanations of the results of AI systems. There are several important questions to answer before AI can be deployed at scale in our businesses and societies. Most of these issues are being discussed by experts and the wider communities, and it seems there is broad consensus on where they come from. There is, however, less consensus on, and experience with how to practically deal with those issues in organizations that develop and use AI, both from a technical and organizational perspective. In this paper, we discuss the practical case of a large organization that is putting in place a company-wide methodology to minimize the risk of undesired consequences of AI. We hope that other organizations can learn from this and that our experience contributes to making the best of AI while minimizing its risks.

Introduction

AI creates concerns including discrimination, opacity, privacy, and malicious use, but organizations have less practical guidance for addressing them. Telefónica responds with AI principles intended to guide responsible uptake while distinguishing company-level responsibilities from issues requiring government action.

Related work

Principles of Artificial Intelligence

Towards a methodology for “Responsible AI

The methodology operationalizes responsible AI through principles, checkpoints, tools, training, and governance. It applies these ingredients to fairness, explainability, privacy, and security across the AI development lifecycle.

by Design”

  • Responsible AI by Design combines principles, questions and checkpoints, technical tools, training, and governance with assigned responsibilities and accountabilities.Design requires collaboration across Engineering, Security, Legal, Business, IT, Human Resources, and Procurement, with top-management endorsement.Source: by Design” · p. 3

Training content

Fair AI

  • Fairness definitions include unawareness, group, individual, and counterfactual fairness, while independence, separation, and sufficiency cannot all be achieved simultaneously.Source: Fair AI · p. 4
  • Fair AI evaluates and mitigates bias using fairness criteria, metrics, and pre-processing, in-processing, or post-processing techniques.Examples include reweighing, adversarial debiasing, and threshold adjustment for Equalized Odds.Source: Fair AI · pp. 4–5

Transparent & Explainable AI (xAI)

  • Explainable AI distinguishes inherently interpretable white-box models from opaque black-box models and uses model-agnostic, model-specific, or surrogate explanations.LIME explains individual classifications, LRP attributes deep-learning predictions to input features, and surrogate decision trees summarize black-box behavior.Source: Transparent & Explainable AI (xAI) · p. 5

Privacy & Security by Design

  • Privacy and security tools test re-identification risk and robustness against adversarial attacks that manipulate model outputs through minimal input changes.Source: Privacy & Security by Design · pp. 5–6

Implementation of the methodology

Telefónica implements Responsible AI by Design through awareness, role-specific training, supporting tools, and agile governance. The approach emphasizes self-control and delegated responsibility rather than extensive committee-based controls.

Implementation of the methodology

  • The implementation begins with an awareness campaign covering AI, its organizational uses, challenges, principles, methodology, training, and tools.Source: Implementation of the methodology · p. 6
  • Training starts with employees closest to AI product development and procurement, then expands across the organization with technical depth tailored to roles.A one-hour online course covers AI concepts, bias, principles, tools, questionnaires, governance, and practical assessments.Source: Implementation of the methodology · p. 6
  • Specific tools are needed to make justifiable statements about fairness and explainability.Source: Implementation of the methodology · p. 6
  • The organization uses agile governance and delegates responsibility to product and service owners, supported by training and tools instead of extensive committee controls.Existing privacy and security governance models remain in place for those matters.Source: Implementation of the methodology · p. 6

Challenges and research needs

Fair AI still needs unified, practical approaches to fairness, explainability, and organizational risk management. Key challenges include adapting explanations to users and domains, automating emerging methods, and balancing innovation speed against risk reduction.

Fairness

  • Fairness needs a unified metric and framework because existing techniques measure different aspects of group or individual fairness.The paper argues that standardization would simplify evaluation, correction, and comparison in industry.Source: Fairness · p. 6

Explainability

  • Explainability should be adapted to user profiles, transparency requirements, domains, and sometimes regulations such as GDPR.Ideally, explanations also incorporate domain knowledge rather than only feature rankings.Source: Explainability · pp. 6–7
  • Promising domain-aware and knowledge-based explanation proposals remain early-stage research and are difficult to automate for production environments.These approaches include training models with domain explanations and combining detected features with knowledge-based reasoning.Source: Explainability · p. 7
  • Explainability lacks a formal common reference, and feature-only explanations may omit latent or correlated information relevant to interpretation.Existing criteria such as concision and consistency do not yet establish a general standard.Source: Explainability · p. 7

Methodology

  • A Responsible AI methodology must balance rapid innovation with risk controls, motivating self-control through employee awareness and training.The authors expect slight development and go-to-market delays but reduced error risk and higher social acceptance.Source: Methodology · p. 7

Use Cases at Telefónica

Telefónica applies Responsible AI practices through bias auditing and mitigation, user data transparency and anonymization, and explainability features in recommendation and anomaly-detection products.

Fairness

  • Luca Ethics quantifies bias with several fairness metrics and individually optimizes classification thresholds to satisfy equal opportunity while preserving performance.Its two phases are bias measurement followed by post-processing mitigation.Source: Use Cases at Telefónica · p. 7

Privacy and transparency

  • Telefónica provides data transparency through Aura’s transparency center and privacy protection through Spectra anonymization and anti-reidentification tools.The transparency center gives customers access to and control over personal data held and used by the system.Source: Use Cases at Telefónica · pp. 7–8

Explainability

  • Device Recommender generates natural-language explanations, while Luca Comms provides local and counterfactual explanations for OneClass SVM anomaly detection through a visual interface.Device Recommender combines contextual-bandit reinforcement learning with market-research information from conjoint analysis.Source: Use Cases at Telefónica · p. 8

Conclusions

The paper presents Responsible AI by Design as a company-wide methodology combining principles, training, development questions, tools, and governance. It is intended as a practical starting point that will evolve with shared experience rather than a final ethical standard.

Conclusions

  • Responsible AI by Design combines AI principles, awareness and training, development questions, supporting tools, and governance with defined responsibilities and accountability.The methodology is intended for organizations deploying AI at large scale to reduce unintended discrimination and unintelligible decisions.Source: Conclusions · p. 8
  • The authors argue that practical experience should be shared as AI adoption grows, because organizations need workable approaches to minimize AI risks.The paper presents Telefónica’s experience as a contribution that other organizations may learn from.Source: Conclusions · p. 8
  • The principles and methodology are not final ethical standards but a starting point expected to evolve as collective experience accumulates.The authors explicitly describe their organization as having only begun the journey.Source: Conclusions · p. 8
Loading 1909.12838v2…