Abstract. As Machine Learning (ML) systems are increasingly being used in critical domains, the need for a coherent framework to account for the discriminatory effects of their outcomes becomes urgent. In computer science, existing approaches tend to emphasise the role of ML design, presenting algorithmic fairness as a matter of making better design choices, especially at the level of the data used to train the model. This dissertation challenges the adequacy of such a design-centric perspective on the problem of unfair ML predictions. It does so by reconnecting it with broader and longer-standing issues within the philosophy of computational artefacts that have largely been overlooked in the current debate in the ethics of artificial intelligence. The primary contribution of this thesis is to reframe the analysis of algorithmic discrimination around the notions of use, maintenance, and repair of ML systems. Specifically, I argue that the correctness criteria of an ML system should be reformulated in terms of the contextual convergence of the diverse normative requirements of the agents who use it. Compared to other accounts of ML normativity, this reconceptualisation avoids succumbing to scepticism about implementation ascriptions, while returning a more dynamic and realistic understanding of how normative requirements circulate, feed back, conflict, and adapt across complex ML systems. Crucially, I claim that this shift has the advantage of allowing for a richer understanding of algorithmic fairness, viewing it as a plurality of data repair practices, rather than a static value embodied by certain ML designs. Two main formal contributions follow from this analysis: the introduction of validity criteria for ML predictions and the development of a novel logical framework to reason about the impact of errors in the input data on the fairness of algorithmic outcomes. 

Abstract. Integrating fairness into machine learning models has been an important consideration for the last decade. Here, neurosymbolic models offer a valuable opportunity, as they allow the specification of symbolic, logical constraints that are often guaranteed to be satisfied. However, research on neurosymbolic applications to algorithmic fairness is still in an early stage. In this work, we bridge this gap by integrating counterfactual fairness into the neurosymbolic framework of logic tensor networks (LTN). We use LTN to express accuracy and counterfactual fairness constraints in first-order logic and employ them to achieve desirable levels of both performance and fairness at training time. Our approach is agnostic to the underlying causal model and data generation technique; for this reason, it may be easily integrated into existing pipelines that generate and extract counterfactual examples. We show, through concrete examples on three benchmark datasets, that logical reasoning about counterfactual fairness has some important advantages, among which its intrinsic interpretability, and its flexibility in handling subgroup fairness. Compared to three recent methodologies in counterfactual fairness, our experiments show that a neurosymbolic, LTN-based approach attains better levels of counterfactual fairness. 

Abstract. In the present work, we develop a novel information-theoretic and logic-based approach to data bias in Machine Learning predictions and show its relevance in the specific context of fairness evaluation. We frame predictions made on biased data as Ulam games, which formalise key aspects of data-driven inference, and from which a variation of the rational non-monotonic consequence relation can be defined. We investigate this framework to model how differential levels of noise in input features impact Machine Learning predictions.  To the best of our knowledge, this is the first game-theoretic formalisation of ML unfairness.

Abstract. In the context of deterministic scientific simulations, formal validity requirements have been typically defined to help us reason about the relationship between the mathematical model underlying the target system and the computational model used to simulate it. With machine learning simulations entering the picture, we argue that these formal requirements need to be reviewed, as the objects to which they apply have significantly changed. This is due to several reasons: the target system is no longer an available system object of investigation; the probabilistic mathematical model is abstracted from the target system through the mediation of the computational model, which however remains largely opaque to us due to its high complexity. For these reasons, we formulate weaker, probabilistic versions of the traditional relations of Simulation, Bisimulation, and Approximate Simulation. Accordingly, we define three corresponding validity criteria that capture a range of cases, from the strongest to the weakest, depending on the extent to which the machine learning model can be assumed to correctly represent its target system.

Abstract. By analogy with Kripke’s claim that no fact of the matter can determine the meaning of a word, the sceptical paradox of implementation is an argument to the conclusion that no fact of the matter can determine the function of a computational artefact. The paradox targets the prevailing view within the philosophy of computer science, according to which the function of a computational artefact is to be identified with the content of its functional specification, a mathematical object that formalises the intentions of the artefact’s designer. In existing formulations, such a view requires the existence of certain semantic intentions in the head of the designer. However, if we accept Kripke’s claim that there is no such thing as a mental state of semantic intention, this leaves ultimately indeterminate what is the function implemented by a computational artefact. Just as Kripke’s solution to the paradox requires replacing the received conception of language with a very different one, so the sceptical paradox for computational artefacts forces us to explore a new perspective on computational artefacts that, this time, does without semantic intentions.

Abstract. This paper introduces a first approach to miscomputations for data-driven systems. First we establish an ontology for data-driven learning systems and categorize various computational errors based on the Levels of Abstraction ontology. Next, we consider those computational errors which are associated to user evaluation and requirements and consider the User Level ontology, identifying two additional types of miscomputation. 

Abstract. The widespread emergence of phenomena of bias is certainly among the most adverse impacts of new data-intensive sciences and technologies. The causes of such undesirable behaviours must be traced back to data themselves, as well as to certain design choices of machine learning algorithms. The task of modelling bias from a logical point of view requires to extend the vast family of defeasible logics and logics for uncertain reasoning with ones that capture some few, fundamental properties of biased predictions. However, a logically grounded approach to machine learning fairness is still at early stages in the literature. In this paper, we discuss current approaches to the topic, formulate general logical desiderata for logics to reason with and about bias, and provide a novel approach. 

Abstract. Algorithmic discrimination is often attributed to flaws in Machine Learning (ML) design, particularly biased training data. Framed in this way, achieving fairness becomes primarily a matter of redesigning ML models, namely by retraining them on higher-quality, fairer datasets. This paper challenges this design-centric view by proposing an alternative epistemology grounded in the notion of "broken-world thinking". From this perspective, ML functionality is not fixed at the moment of design but instead emerges through ongoing, collective practices of maintenance and repair across the system’s lifecycle. To clarify the distinction between understanding algorithmic unfairness in terms of ML redesign and understanding it in terms of ML repair, I identify three key limitations of the design paradigm in addressing ML errors and show how they are naturally accommodated within the broken-world epistemology. On this basis, I argue that the latter offers a richer epistemology and should therefore be preferred in our analysis of algorithmic discrimination. From this standpoint, algorithmic fairness should not be seen as a static property embedded in design choices, but rather as a plurality of reparative practices through which we recognise and take responsibility for the inherent fragility of ML systems.