Fairness, in its literal sense, describes acting or deciding in such a way that a just proportion is respected. The ethical principle connected to it establishes, with possible variations, that similar situations should be judged and dealt with in a similar way. Some versions of this principle, concerning political and social fairness, emphasize that benefits and burdens should be distributed within society in a proportionate manner, paying particular attention to improving the conditions of the least privileged; this indicates that fairness does not necessarily consist in treating everyone identically, but rather in fairly compensating for existing inequalities in order to restore a just balance. It is also possible to speak of fairness with regard to the use of AI systems to make decisions and predictions. Indeed, a system’s decision or prediction may favor or disadvantage more or less privileged groups. Furthermore, considering that systems of this kind are also employed in contexts that are highly sensitive from a social and economic point of view—such as credit granting or personnel selection—this specific problem is of great significance.

What has been said so far raises a legitimate question: what is “fair”? What is fair to one individual will necessarily be fair to another? The answer to these questions suggests that the concept of fairness is not a universal and static value, but must be considered in context, as it is rooted in each person’s cultural landscape and individual sensibilities.

Although fairness is a subjective and context-dependent value, its adoption as a guiding principle has made it necessary to establish shared criteria through which actions can be taken to compensate for disadvantages and inequalities. Certain sensitive characteristics—such as ethnicity, gender, social position, and economic status—are typically taken into consideration when formulating these shared criteria.

An important example of an analysis of the notion of fairness is the one proposed by the philosopher John Rawls. A central figure in modern political philosophy, in his proposal Justice as Fairness he interprets fairness as a mechanism for correcting inequalities produced by chance or fortune. Rawls therefore structures a thought experiment designed to establish social principles impartially: individuals called upon to choose these principles are in fact placed under the condition of the veil of ignorance, which deprives them of knowledge of their own position in the future society. In this way, impartiality is guaranteed because, being unable to formulate rules that favor their own position, individuals are encouraged to protect the most disadvantaged members of society as well.

Rawls’s veil of ignorance clearly represents a utopia in human reality and, even more so, a technical and structural impossibility with regard to fairness in artificial intelligence. Here, in fact, the concept translates into the requirement that algorithmic systems operate without perpetuating or amplifying existing biases. The attempt to implement the ethical principle of fairness by imposing constraints that ensure the system produces results consistent with specific metrics for measuring its fairness, however, is confronted with the actual complexity of the models.

The first and most obvious difficulty is that AI models can only be developed by using existing data that define their future behavior: if the training dataset reflects historical imbalances or prejudices, for example by insufficiently or inaccurately representing a certain demographic group, the model does nothing but amplify and replicate these inequalities in its future predictions or decisions. In this case, the model is said to have a structural bias (a topic already addressed in the entry “B as in Bias”).

The second, equally evident observation concerns the lack of ethical intention. The algorithm is not capable of empathy and has no will of its own that drives it to be fair: all it does is optimize a function. Therefore, all that can be done is to hope that the system conforms to a mathematical translation of the ethical ideal, which is limited by its very nature. The problem of fairness in AI must therefore be understood not only from a moral point of view but, above all, by taking into account the technical limitations of the models.

Finally, among the many variables to be taken into account are the human components that form part of a model’s life cycle: not only data scientists, programmers, and other professionals involved in the field, but also end users. Taking the latter into consideration opens the way to many other requirements, including regulatory ones: legal issues such as discrimination, reliability, and explainability make attention to the fairness of AI systems no longer a discretionary ethical choice, but rather an imperative concerning the legitimacy of systems of this kind.
 
 Fairness and Legal Rights
 
 As mentioned earlier, discussing fairness is not a simple undertaking; its meaning, in fact, is intrinsically dependent on the context, culture, and specific situation. In computer science, as in law, people have been questioning for years what it means for a system or a decision to be fair. Various scientific communities have attempted to translate this idea into numbers and formulas, producing different measures of fairness. However, none of them can individually capture the entire semantic dimension of the term, and, moreover, it is mathematically impossible to satisfy some of these measures simultaneously. Thus, a system may appear fair according to one model and unfair according to another: everything depends on the point of view adopted to analyze it.

The same difficulty has been encountered in the legal world. Legislators have often used the term without providing a precise definition. A useful distinction, however, is the one between procedural fairness and substantive fairness.

With reference to the procedural dimension, the term refers to the way individuals are treated: it means ensuring dignity, respect, and a fair process. This is the idea underlying the right, enshrined in the Charter of Fundamental Rights of the European Union (Art. 47), to an effective remedy and to an impartial tribunal.

The second interpretation of the concept of fairness, on the other hand, concerns the substance: the conclusion reached at the end of the decision-making process must be fair and non-discriminatory, as established by the right to non-discrimination (Art. 21 of the Charter). This meaning of fairness becomes crucial in certain areas of jurisprudence, such as those concerning employment and technology. One example is a system used for personnel selection: it must assess skills and requirements relevant to the job, and must not base its decision on characteristics irrelevant to the candidate’s assessment, such as ethnicity or sexual orientation.

Similar trends have meant that the incorporation of fairness has become a central element in the development of artificial intelligence systems. The guidelines of the High-Level Expert Group on Trustworthy AI identify it as one of the fundamental requirements for developing trustworthy systems. The AI Act, too, although it does not mention the term in its articles, refers to the principle in its preamble: it requires that systems do not generate discrimination (Recital 27) and reiterates the need for an automated decision to be contestable (Recital 59).

In recent years, academic debate has also focused on the distinction between algorithmic discrimination and human discrimination, examining the ability of current anti-discrimination laws to address automated decisions. In particular, the regulations currently in force explicitly prohibit basing decisions on characteristics such as sex and ethnicity; however, they may prove inadequate when faced with the new forms of differentiation introduced by algorithms. As numerous cases demonstrate, AI systems are capable of identifying neutral characteristics, which are therefore not recognized by law as sensitive, but which are strongly correlated with data whose use is prohibited. In this context, explainability techniques, that is, methods that make it possible to understand how an AI system arrived at a particular decision, play a key role. On the one hand, these techniques make it possible to identify potential biases and verify compliance with the principle of non-discrimination; on the other, they represent a fundamental requirement for guaranteeing genuine access to justice. Indeed, only by knowing the logic underlying an algorithmic decision is it possible to challenge it or request its review.

Benedetta Bonomi & Matteo Barakat

(translated by MIRAI)