Introduction

There is a need for a more holistic approach to fair AI, that includes technical steps, sociological activities, legal compliance measures and understanding, and ethical considerations. Addressing the need for AI systems free from discrimination requires a multidisciplinary approach that combines social, legal, and technical perspectives. Despite significant advancements in research and technical solutions, a gap remains between socio-legal and technical approaches.

Fair-by-design methodology refers to an approach in the field of AI where fairness considerations are integrated into the AI lifecycle from the outset. This methodology aims to ensure that AI systems are fairly designed, addressing potential biases or discrimination (according to a particular definition of fairness) during the design and development phase rather than attempting to rectify them after deployment.

Fair-by-design methodologies emphasize proactive measures to mitigate biases and promote fairness, such as using diverse and representative datasets, implementing fairness-aware algorithms, and incorporating transparency and accountability mechanisms into AI systems.

We address this complex problem by proposing a meta-methodology – namely, FairBridge – offering a reference for defining AI fairness methodologies that integrate all three perspectives (social, legal, and technical ones). The meta-methodology utilizes a questionnaire-based system where socio-legal and technical domain experts iteratively refine questions and responses, supported by automation.

Fair-by-design Approaches

The design of fair AI approaches involves a detailed analysis of the entire lifecycle of an AI system, as considering the non-discrimination aspects of an AI system requires constant actions and monitoring at every phase of the AI lifecycle. The analysis of the AI lifecycle and its relation to fairness actions begins with the analysis of the lifecycle itself from two different perspectives: i) the socio-legal perspective, and ii) the technical perspective.

The key takeaways are that there are significant gaps between these perspectives:

  • The technical point of view focuses on a narrower – and more limited – AI lifecycle compared to the broader social and legal perspective.

  • It is very rare for technical solutions to cover the entire AI lifecycle, because its narrower definition compared to its socio/legal counterpart.

  • Technical methods typically intervene at very specific phases.

Albeit there are significant differences from the socio/legal/ethical lens and the technical one, the core phases of the AI life cycle are shared by both perspectives: there is a clear identification of the necessity of carefully considering how data has been sampled (collected or generated), analyzing whether it contains any kind of bias, and studying how it has to be processed not to introduce further bias. Both perspectives also place careful emphasis on how AI algorithms should be evaluated for fairness.

More details the overview over the different perspectives and current gaps can be found in the following subsections:

Merging the gap between these perspectives is a crucial aspect of a fair-by-design methodology. FairBridge addresses this precise challenge.

Technical Lens

There is a dearth of fair-by-design methodologies tackled from the engineering/technology perspective. Technological approaches mostly focus on specific phases of the AI lifecycle (e.g., data collection, training of models, evaluation of results, etc.).

AI outside of the ML subfield is extremely underrepresented, and this is especially true from the technological point of view. This is a limitation:

  • There are many AI algorithms that do not fall into the ML categories whose impact to society and economy is non-negligible and whose behavior can be influenced by various biases.

  • We recommend researchers and practitioners to start increasing their attentions to other AI domains as well.

AI Lifecycle - Technical Perspective

Technological methods to enforce fairness in ML are typically subdivided according to the phase in the AI lifecycle in which they can be applied. A broad classification is the following:

  • Pre-processing techniques approach the problem by removing the underlying discrimination from the data prior to modelling. This is argued in the literature to be the most flexible phase of repairing bias in the pipeline, as it makes no assumptions with respect to the choice of applied modelling technique. The methods, that modify the training data are at odds with policies like GDPR’s right to an explanation, potentially introducing new biases. Sufficient knowledge of the data and veracity assumptions are required.

  • In-processing techniques modify the traditional learning algorithms to account for fairness during the model training phase. They require a higher technological effort and integration with standard ML libraries to avoid porting challenges.

  • Post-processing is a set of methods that can be run on the output scores of the classifier as a post-training processing step to make decisions fairer. The accuracy is suboptimal when compared to “equally fair” classifiers and could be the case that test-time access to protected attributes is needed, which may not be legally permissible.

For a detailed survey on technical methods for enhancing fairness of AI approaches, we refer to the recent paper from [Calegari et al.].

Technological methodologies tend to adopt a reductionist approach, aiming at decomposing complex problems into a series of (hopefully easier) sub-problems. Under this solution paradigm, it is more “natural” to devise approaches that focus on specific fairness-related aspects, such as bias detection or mitigation, rather than to create holistic approaches encompassing the entire design process. This is compounded by the fact that a fair-by-design methodology can hardly be founded on merely technological grounds: a fair-by-design approach should encompass several aspects (e.g., dataset creation, data sampling, algorithmic choices, output evaluation, etc.) that should involve human-mediated elements, and thus cannot be entirely decoupled from sociological, economical, cultural and legal subtexts.

Bridging the Gap

Considering the entire AI system lifecycle is fundamental when assessing fairness and mitigating bias in AI systems:

  • It allows for a comprehensive understanding of how bias can infiltrate at various stages, from data collection and model training to deployment and impact assessment.

  • By examining the entirety of the process, we can identify and address potential biases more effectively, ensuring fairness across all stages of development and implementation.

The analysis of the socio/legal and technological lenses revealed how there is still a non-negligible distance between the two areas. It is very rare for technical solutions to cover the entire AI lifecycle, because its narrower definition compared to its socio/legal counterpart. More commonly, technical methods intervene at very specific phases.

The interplay between sociological/legal and technological perspectives is still in its infancy: engineering solutions tend to adopt excessively reductionistic approaches (discarding the big picture) while sociological/legal varied indications and suggestions struggle to coalesce into a set of well-defined and actionable guidelines which can be actually applied

Other gaps between the technological and legal perspectives stem from the relative lack of (effective) communication between legal experts (and lawmakers), ethicists and social scientists on the one hand, and technical experts (i.e., the developers of AI systems) on the other.

  • The socio/legal approaches tend to provide broader requirements and guidelines, refraining from defining how fairness should be measured in practice.

  • The technical approaches typically start with the aim of defining fairness metrics, requiring:

    • a definition of the fairness notions from social, legal, ethical and technical perspectives;

    • a quantitative mechanism to measure them (if possible).

  • Fairness notions vary by context and stakeholder, requiring different actions to achieve. They can be measured quantitatively using fairness metrics, but this leads to numerous metrics each capturing different aspects of fairness.

Summarizing:

  • There is a clear gap in current fair-by-design practice.

  • The integration of social, legal, ethical, and technological perspectives presents two challenges: complexity and interdisciplinarity.

  • Each perspective operates within its own framework:

    • Social, legal, and ethical perspectives focus on human behavior, ethical principles designed for digitalization, and regulation, while technological perspectives prioritize efficiency, functionality, and innovation.

    • Bridging these perspectives requires interdisciplinary collaboration.

    • This is compounded by cultural and contextual differences, which are crucial from the legal point of view.

  • Divergent priorities: technological perspectives often prioritize performance and scalability, whereas social and legal considerations emphasize accountability, equity and the protection of (fundamental) rights, democracy, and the rule of law.

  • Pace of change: technology evolves rapidly, outpacing the ability of social, ethical and legal frameworks to adapt. This misalignment leads to regulatory gaps and ethical dilemmas.

  • Lack of common vocabulary and/or conceptual framework: each discipline has its own vocabulary and ‘language’ and concepts whilst quite often referring to the same elements or objectives. Mapping and matching these diverging vocabulary and concepts are a lengthy but crucial process.