Emanuele Ratti

This book provides a unified account of models and model-building practice in AI-based science, particularly machine learning (ML). It analyzes the relationship between ML model-building practices and scientific domain knowledge, develops an account of ML models as technical artifacts structured around five levels of abstraction that captures their representational capabilities, and shows how this framework can be used to reformulate contemporary debates in philosophy of science and AI in more fruitful ways.

We provide an analysis of theory-ladenness in machine learning (ML) in science, where ‘theory’ (that we call ‘domain-theory’) refers to the domain knowledge of the scientific discipline where ML is used. By constructing an account of ML models based on a comparison with phenomenological models, we show (against recent trends in philosophy of science) that ML model-building is mostly indifferent to domain-theory, even if the model remains theory-laden in a weak sense, which we call theory-infection. These claims, we argue, have far-reaching consequences for the transferability of ML across scientific disciplines, and shift the priorities of the debate on theory-ladenness in ML from descriptive to normative.

This chapter offers an overview of how virtue-based concepts have been used by philosophers of science to shed light on epistemic aspects of science. In the epistemology of science, the word _virtue_ has referred to two different concepts. First, virtue can be understood as excellence, where excellence is a quality of a model, a theory, or a hypothesis. Second, virtue can be understood more narrowly as a stable character trait and/or disposition of scientists themselves. The first meaning is connected to the long-standing debate on the qualities that make a scientific theory a good scientific theory. The second meaning is connected to a much more recent conversation exploring the connections between virtue epistemology and philosophy of science. I explore how these two meanings of _virtue_ have been developed, and I highlight underexplored areas that can advance our understanding of the relation between virtue theory and philosophy of science.

Responsible artificial intelligence (AI) requires integrating ethical awareness into the full process of designing and developing AI, including ethics-based auditing of AI technology. We claim the Capability Approach (CA) of Sen and Nussbaum grounds AI ethics in essential human freedoms and can increase awareness of the moral dimension in the technical decision making of developers and data scientists constructing data-centric AI systems. Our use of CA focuses awareness on the ethical impact that day-to-day technical decisions have on the freedom of data subjects to make choices and live meaningful lives according to their own values. For internal auditing of AI technology development, we design and develop a light-weight ethical auditing tool (LEAT) that uses simple natural language processing (NLP) techniques to search design and development documents for relevant ethical characterizations. We describe how CA guides our design, demonstrate LEAT on both principle- and capabilities-based use cases, and characterize its limitations.

Background Artificial Intelligence (AI) is being designed, tested, and in many cases actively employed in almost every aspect of healthcare from primary care to public health. It is by now well established that any application of AI carries an attendant responsibility to consider the ethical and societal aspects of its development, deployment and impact. However, in the rapidly developing field of AI, developments such as machine learning, neural networks, generative AI, and large language models have the potential to raise new and distinct ethical and social issues compared to, for example, automated data processing or more ‘basic’ algorithms. Methods This article presents a scoping review of the ethical and social issues pertaining to AI in healthcare, with a novel two-pronged design. One strand of the review (SR1) consists of a broad review of the academic literature restricted to a recent timeframe (2021–23), to better capture up to date developments and debates. The second strand (SR2) consists of a narrow review, limited to prior systematic and scoping reviews on the ethics of AI in healthcare, but extended over a longer timeframe (2014–2024) to capture longstanding and recurring themes and issues in the debate. This strategy provides a practical way to deal with an increasingly voluminous literature on the ethics of AI in healthcare in a way that accounts for both the depth and evolution of the literature. Results SR1 captures the heterogeneity of audience, medical fields, and ethical and societal themes (and their tradeoffs) raised by AI systems. SR2 provides a comprehensive picture of the way scoping reviews on ethical and societal issues in AI in healthcare have been conceptualized, as well as the trends and gaps identified. Conclusion Our analysis shows that the typical approach to ethical issues in AI, which is based on the appeal to general principles, becomes increasingly unlikely to do justice to the nuances and specificities of the ethical and societal issues raised by AI in healthcare, as the technology moves from abstract debate and discussion to real world situated applications and concerns in healthcare settings.

The past few years have witnessed a growth of interest in the historical and philosophical dimensions of bioinformatics as a discipline. Despite the importance of bioinformatics in addressing the issues raised by the growing amount of biological data, data management is often seen as all it has to offer to biology. However, the emphasis on data management may come at the expense of understanding how bioinformatics generates genuine biological knowledge beyond its instrumental value for bench biologists. Some authors have taken the first steps beyond data management, and towards the characterization of bioinformatics as a unique epistemic endeavor by stressing how its experimental practices can be conducive to biological knowledge. In this article, we build upon these attempts, and by using a detailed case study from the field of single cell transcriptomics (i.e., RNA velocity), we provide a fully-fledged characterization of bioinformatics as an experimental discipline.

When does more data help and when does it not in the sciences? In the past decade, this question has become central because of the phenomenon of Big Data. While these discussions started as a result of somewhat naive ideas that have been closely analyzed and mostly rejected in the philosophy of data, the question about the epistemic difference that more or less data makes still matters, especially in light of the impressive performance of data science and machine learning tools, which seem to improve their outcomes when they are trained on large volumes of data. In several areas of the sciences, having more data is also connected to methodological and epistemic benefits and more generally something that research should strive toward. More data is often equated to better science: this elicits crucial questions about the epistemic value of the quantity of data. In this chapter, we discuss this problem in light of current discussions in the life and health sciences and the philosophy of data.

We propose a conceptualization and implementation of AI ethics via the capability approach. We aim to show that conceptualizing AI ethics through the capability approach has two main advantages for AI ethics as a discipline. First, it helps clarify the ethical dimension of AI tools. Second, it provides guidance to implementing ethical considerations within the design of AI tools. We illustrate these advantages in the context of AI tools in medicine, by showing how ethics-based auditing of AI tools in medicine can greatly benefit from our capability-based approach.

In this chapter, we propose a non-traditional RCR training in data science that is grounded in a virtue theory framework. First, we delineate the approach in more theoretical detail by discussing how the goal of RCR training is to foster the cultivation of certain moral abilities. We specify the nature of these ‘abilities’: while the ideal is the cultivation of virtues, the limited space allowed by RCR modules can only facilitate the cultivation of superficial abilities or proto-virtues, which help students to familiarize themselves with moral and political issues in the data science environment. Third, we operationalize our approach by stressing that (proto-)virtue acquisition (like skill acquisition) occurs through the technical and social tasks of daily data science activities, where these repetitive tasks provide the opportunities to develop (proto-)virtue capacity and to support the development of ethically robust data systems. Finally, we discuss a concrete example of implementing this approach. In particular, we describe how this method is applied to teach data ethics to students participating in the CODATA-RDA Data Science Summer Schools.

In the past few years, calls for integrating ethics modules in engineering curricula have multiplied. Despite this positive trend, a number of issues with these ‘embedded’ programs remains. First, learning goals are underspecified. A second limitation is the conflation of different dimensions under the same banner, in particular confusion between ethics curricula geared towards addressing the ethics of individual conduct and curricula geared towards addressing ethics at the societal level. In this article, we propose a tripartite framework to overcome these difficulties. Our framework analytically decomposes an ethics module into three dimensions. First, there is the ethical dimension, which pertains to the learning goals. Second, there is the moral dimension, which addresses the moral relevance of engineers’ conduct. Finally, there is the political dimension, which scales up issues of moral relevance at the civic level. All in all, our framework has two advantages. First, it provides analytic clarity, i.e. it enables course instructors to locate ethical dilemmas in either the moral or political realm and to make use of the tools and resources from moral and/or political philosophy. Second, it depicts a comprehensive ethical training, which enables students to both reason about moral issues in the abstract, and to socially contextualize potential solutions.

In the science and values literature, scholars have shown how science is influenced and shaped by values, often in opposition to the ‘value free’ ideal of science. In this paper, we aim to contribute to the science and values literature by showing that the relation between science and values flows not only from values into scientific practice, but also from (allegedly neutral) science to values themselves. The extant literature in the ‘science and values’ field focuses by and large on reconstructing, post hoc, how values have influenced science; our reconstruction of the case studies, instead, aims to show that scientific concepts and methods _too, because of specific identifiable characteristics,_ can promote some values rather than (or at the expense of) others. We explain this bidirectional relation in analogy to debates on the normativity of technical artifacts and on feminist approaches in science, and we illustrate our claims with cases from the health sciences and machine learning. While our arguments in this paper also draw on post hoc reconstructions, we intend to show where, in the science in the making, we should engage not only with the question whether a practice is value-laden, but also how specific conceptual and methodological choices can influence values down the road. All in all, these considerations expand the ways in which philosophers can contribute to more value-aware scientific practices.

In the last few years, biologists and computer scientists have claimed that the introduction of data science techniques in molecular biology has changed the characteristics and the aims of typical outputs (i.e. models) of such a discipline. In this paper we will critically examine this claim. First, we identify the received view on models and their aims in molecular biology. Models in molecular biology are mechanistic and explanatory. Next, we identify the scope and aims of data science (machine learning in particular). These lie mainly in the creation of predictive models which performances increase as data set increases. Next, we will identify a tradeoff between predictive and explanatory performances by comparing the features of mechanistic and predictive models. Finally, we show how this a priori analysis of machine learning and mechanistic research applies to actual biological practice. This will be done by analyzing the publications of a consortium—The Cancer Genome Atlas—which stands at the forefront in integrating data science and molecular biology. The result will be that biologists have to deal with the tradeoff between explaining and predicting that we have identified, and hence the explanatory force of the ‘new’ biology is substantially diminished if compared to the ‘old’ biology. However, this aspect also emphasizes the existence of other research goals which make predictive force independent from explanation.

The notion of biological function is fraught with difficulties—intrinsically and irremediably so, we argue. The physiological practice of functional ascription originates from a time when organisms were thought to be designed and remained largely unchanged since. In a secularized worldview, this creates a paradox which accounts of functions as selected effect attempt to resolve. This attempt, we argue, misses its target in physiology and it brings problems of its own. Instead, we propose that a better solution to the conundrum of biological functions is to abandon the notion altogether, a prospect not only less daunting than it appears, but arguably the natural continuation of the naturalisation of biology.

A central distinction in Curzer and Epstein (2022) is the one between “protect the disadvantaged” and “protect the data”. This can open up discussions about the relationship between ethics and epistemology in the practice of science. Focusing on the disadvantaged to the exclusion of good scientific practices, Curzer and Epstein argue, can harm everyone impacted by medical science, including the disadvantaged. For this reason, they propose that “ethical data scientists should strive for accurate data and scientifically sound data analysis” (2022, p 2) with attention to minimizing data processing errors, bias, and outside influence, and that includes identifying errors caused by tendencies to neglect disadvantaged and historically underrepresented communities and groups. While we agree with several points made by Curzer and Epstein, we also have three main points of concern. We believe that Curzer and Epstein partially mischaracterize our original paper (Ratti & Graves, 2021) as “protect the disadvantaged.” We view as a controversial that “advancing knowledge” is a pure epistemic goal achievable by pure epistemic means. We see as important the social context to which data science is going to make a difference, and why ethics and epistemology are necessarily intertwined because of that context.

It has been argued that ethical frameworks for data science often fail to foster ethical behavior, and they can be difficult to implement due to their vague and ambiguous nature. In order to overcome these limitations of current ethical frameworks, we propose to integrate the analysis of the connections between technical choices and sociocultural factors into the data science process, and show how these connections have consequences for what data subjects can do, accomplish, and be. Using healthcare as an example, attention to sociocultural conversion factors relevant to health can help in navigating technical choices that require broader considerations of the sociotechnical system, such as metric tradeoffs in model validation, resulting in better ethical and technical choices. This approach promotes awareness of the ethical dimension of technical choices by data scientists and others, and that can foster the cultivation of 'ethical skills' as integral to data science.

In the past few years, machine learning (ML) tools have been implemented with success in the medical context. However, several practitioners have raised concerns about the lack of transparency—at the algorithmic level—of many of these tools; and solutions from the field of explainable AI (XAI) have been seen as a way to open the ‘black box’ and make the tools more trustworthy. Recently, Alex London has argued that in the medical context we do not need machine learning tools to be interpretable at the algorithmic level to make them trustworthy, as long as they meet some strict empirical desiderata. In this paper, we analyse and develop London’s position. In particular, we make two claims. First, we claim that London’s solution to the problem of trust can potentially address another problem, which is how to evaluate the reliability of ML tools in medicine for regulatory purposes. Second, we claim that to deal with this problem, we need to develop London’s views by shifting the focus from the opacity of algorithmic details to the opacity of the way in which ML tools are trained and built. We claim that to regulate AI tools and evaluate their reliability, agencies need an explanation of how ML tools have been built, which requires documenting and justifying the technical choices that practitioners have made in designing such tools. This is because different algorithmic designs may lead to different outcomes, and to the realization of different purposes. However, given that technical choices underlying algorithmic design are shaped by value-laden considerations, opening the black box of the design process means also making transparent and motivating (technical and ethical) values and preferences behind such choices. Using tools from philosophy of technology and philosophy of science, we elaborate a framework showing how an explanation of the training processes of ML tools in medicine should look like.

Making public policy choices based on available scientific evidence is an ideal condition for any policy making. However, the mechanisms governing these scenarios are complex, non-linear, and, alongside the medical-health and epidemiological issues, involve socio-economic, political, communicative, informational, ethical and epistemological aspects. In this article we analyze the role of scientific evidence when implementing political decisions that strictly depend on it, as in the case of the COVID-19 pandemic. In carrying out this analysis, we will focus above all on the Italian case. This, on the one hand, precisely because Italy led the way regarding the containment policies of the pandemic. Secondly, the government's action was immediately criticized in various respects. Some were calling into question not only the cumbersome political mechanisms, but also suggesting a scarce ability to take scientific evidence into account. On other fronts, there are those who have criticized Italy for its blind and uncritical faith in science and for the paternalism of its decisions. This debate therefore offers the possibility of dealing with some aspects concerning scientific results and their implementation at the political level from the point of view of a political philosophy of science, roughly in the spirit suggested by John Dupré.

In the past few years, the ethical ramifications of AI technologies have been at the center of intense debates. Considerable attention has been devoted to understanding how a morally responsible practice of data science can be promoted and which values have to shape it. In this context, ethics and moral responsibility have been mainly conceptualized as compliance to widely shared principles. However, several scholars have highlighted the limitations of such a principled approach. Drawing from microethics and the virtue theory tradition, in this paper, we formulate a different approach to ethics in data science which is based on a different conception of “being ethical” and, ultimately, of what it means to promote a morally responsible data science. First, we develop the idea that, rather than only compliance, ethical decision-making consists in using certain moral abilities, which are cultivated by practicing and exercising them in the data science process. An aspect of virtue development that we discuss here is moral attention, which is the ability of data scientists to identify the ethical relevance of their own technical decisions in data science activities. Next, by elaborating on the capability approach, we define a technical act as ethically relevant when it impacts one or more of the basic human capabilities of data subjects. Therefore, rather than “applying ethics”, data scientists should cultivate ethics as a form of reflection on how technical choices and ethical impacts shape one another. Finally, we show how this microethical framework concretely works, by dissecting the ethical dimension of the technical procedures involved in data understanding and preparation of electronic health records.

In the past few years, scholars have been questioning whether the current approach in data ethics based on the higher level case studies and general principles is effective. In particular, some have been complaining that such an approach to ethics is difficult to be applied and to be taught in the context of data science. In response to these concerns, there have been discussions about how ethics should be “embedded” in the practice of data science, in the sense of showing how ethical issues emerge in small technical choices made by data scientists in their day-to-day activities, and how such an approach can be used to teach data ethics. However, a precise description of how such proposals have to be theoretically conceived and could be operationalized has been lacking. In this article, we propose a full-fledged characterization of ‘embedding’ ethics, and how this can be applied especially to the problem of teaching data science ethics. Using the emerging model of ‘microethics’, we propose a way of teaching daily responsibility in digital activities that is connected to the higher level ethical challenges discussed in digital/data ethics. We ground this microethical approach into a virtue theory framework, by stressing that the goal of a microethics is to foster the cultivation of moral virtues. After delineating this approach of embedding ethics in theoretical detail, this article discusses a concrete example of how such a ‘micro-virtue ethics’ approach could be practically taught to data science students.

Machine learning (ML) has been praised as a tool that can advance science and knowledge in radical ways. However, it is not clear exactly how radical are the novelties that ML generates. In this article, I argue that this question can only be answered contextually, because outputs generated by ML have to be evaluated on the basis of the theory of the science to which ML is applied. In particular, I analyze the problem of novelty of ML outputs in the context of molecular biology. In order to do this, I first clarify the nature of the models generated by ML. Next, I distinguish three ways in which a model can be novel (from the weakest to the strongest). Third, I dissect the way ML algorithms work and generate models in molecular biology and genomics. On these bases, I argue that ML is either a tool to identify instances of knowledge already present and codified, or to generate models that are novel in a weak sense. The notable contribution of ML to scientific discovery in the context of biology is that it can aid humans in overcoming potential bias by exploring more systematically the space of possible hypotheses implied by a theory.

What makes for a good scientist or a good engineer? How does using a new technology or working in a research lab begin to shape our thought and behavior? How can we best anticipate and navigate the ethical dilemmas created by modern scientific research and technology? Scholars across multiple disciplines have begun turning to a surprising resource to address these questions: discussions of virtue that have their roots in ancient philosophical and religious traditions. This volume gathers a number of these perspectives to show how concepts of virtue can help us better understand, construct, and use the fruits of modern science and technology.

In the last few years, biologists and computer scientists have claimed that the introduction of data science techniques in molecular biology has changed the characteristics and the aims of typical outputs (i.e. models) of such a discipline. In this paper we will critically examine this claim. First, we identify the received view on models and their aims in molecular biology. Models in molecular biology are mechanistic and explanatory. Next, we identify the scope and aims of data science (machine learning in particular). These lie mainly in the creation of predictive models which performances increase as data set increases. Next, we will identify a tradeoff between predictive and explanatory performances by comparing the features of mechanistic and predictive models. Finally, we show how this a priori analysis of machine learning and mechanistic research applies to actual biological practice. This will be done by analyzing the publications of a consortium—The Cancer Genome Atlas—which stands at the forefront in integrating data science and molecular biology. The result will be that biologists have to deal with the tradeoff between explaining and predicting that we have identified, and hence the explanatory force of the ‘new’ biology is substantially diminished if compared to the ‘old’ biology. However, this aspect also emphasizes the existence of other research goals which make predictive force independent from explanation.

The applications of machine learning and deep learning to the natural sciences has fostered the idea that the automated nature of algorithmic analysis will gradually dispense human beings from scientific work. In this paper, I will show that this view is problematic, at least when ML is applied to biology. In particular, I will claim that ML is not independent of human beings and cannot form the basis of automated science. Computer scientists conceive their work as being a case of Aristotle’s poiesis perfected by techne, which can be reduced to a number of straightforward rules and technical knowledge. I will show a number of concrete cases where at each level of computational analysis, more is required to ML than just poiesis and techne, and that the work of ML practitioners in biology needs also the cultivation of something analogous to phronesis, which cannot be automated. But even if we knew how to frame phronesis into rules, still this virtue is deeply entrenched in our biological constitution, which computers lack. Whether computers can fully perform scientific practice independently of humans is an ill-posed question.