The deployment of machine learning models is expected to bring several benefits. Nevertheless, as a result of the complexity of the ecosystem in which models are generally trained and deployed, this technology also raises concerns regarding its (1) interpretability, (2) fairness, (3) safety, and (4) privacy. These issues can have substantial economic implications because they may hinder the development and mass adoption of machine learning. In light of this, the purpose of this paper was to determine, from a positive economics point of view, whether the free use of machine learning models maximizes aggregate social welfare or, alternatively, regulations are required. In cases in which restrictions should be enacted, policies are proposed. The adaptation of current tort and anti-discrimination laws is found to guarantee an optimal level of interpretability and fairness. Additionally, existing market solutions appear to incentivize machine learning operators to equip models with a degree of security and privacy that maximizes aggregate social welfare. These findings are expected to be valuable to inform the design of efficient public policies.

We study copying of machine learning classifiers, an agnostic technique to replicate the decision behavior of any classifier. We develop the theory behind the problem of copying, highlighting its properties, and propose a framework to copy the decision behavior of any classifier using no prior knowledge of its parameters or training data distribution. We validate this framework through extensive experiments using data from a series of well-known problems. To further validate this concept, we use three different use cases where desiderata such as interpretability, fairness or productivization constrains need to be addressed. Results show that copies can be exploited to enhance existing solutions and improve them adding new features and characteristics.

When deployed in the wild, machine learning models are usually confronted with an environment that imposes severe constraints. As this environment evolves, so do these constraints. As a result, the feasible set of solutions for the considered need is prone to change in time. We refer to this problem as that of environmental adaptation. In this paper, we formalize environmental adaptation and discuss how it differs from other problems in the literature. We propose solutions based on differential replication, a technique where the knowledge acquired by the deployed models is reused in specific ways to train more suitable future generations. We discuss different mechanisms to implement differential replications in practice, depending on the considered level of knowledge. Finally, we present seven examples where the problem of environmental adaptation can be solved through differential replication in real-life applications.

Machine learning plays an increasingly important role in our society and economy and is already having an impact on our daily life in many different ways. From several perspectives, machine learning is seen as the new engine of productivity and economic growth. It can increase the business efficiency and improve any decision-making process, and of course, spawn the creation of new products and services by using complex machine learning algorithms. In this scenario, the lack of actionable accountability-related guidance is potentially the single most important challenge facing the machine learning community. Machine learning systems are often composed of many parts and ingredients, mixing third party components or software-as-a-service APIs, among others. In this paper we study the role of copies for risk mitigation in such machine learning systems. Formally, a copy can be regarded as an approximated projection operator of a model into a target model hypothesis set. Under the conceptual framework of actionable accountability, we explore the use of copies as a viable alternative in circumstances where models cannot be re-trained, nor enhanced by means of a wrapper. We use a real residential mortgage default dataset as a use case to illustrate the feasibility of this approach.

This paper explores the impacts of artificial intelligence (AI) systems on human autonomy and identifies the responsibilities of different stakeholders in addressing them. As AI systems increasingly shape not only the outcomes but also the conditions of everyday decision making, they directly influence individuals’ ability to exercise free and informed choice, a capacity central to democratic participation. Hence, it becomes crucial to examine how the definition, design, and implementation of AI systems may effectively support or constrain autonomy, through a well-defined organizational governance framework. Yet, most existing governance approaches overlook or marginalize the role of autonomy, leaving a critical gap. To address this, we introduce a governance framework structured around a set of diagnostic questions that link design choices to concrete dimensions of human autonomy at both system and user levels, and that map these impacts across the different stages of the decision-making process. The framework clarifies how responsibility for answering these questions can be distributed across different organizational roles, promoting internal reflection and collaboration. A dedicated section illustrates how the framework can be applied in practice to recognize and address autonomy impacts in AI systems within different organizational settings and development processes. This paper provides a practical tool to support autonomy-centred AI governance, grounded in the view that upholding human autonomy today is essential to secure sustainable and democratic futures tomorrow.