Re-engineering the Concept of Scientific Discovery in the Age of AI
Artificial intelligence (AI) is transforming the landscape of scientific inquiry challenging traditional paradigms and raising questions about the nature of scientific discovery. As AI systems evolve, their integration into scientific methods marks a revolutionary shift, providing unprecedented opportunities for scientific progress across various disciplines. This project aims to explore these changes, addressing the role of AI as an agent of discovery and its implications for scientific practice and philosophical views on scientific discovery. In particular, this project addresses key epistemic questions: What epistemic virtues do AI systems exhibit that enable them to act as agents of discovery? How do these virtues influence the nature of knowledge generated by AI-driven discoveries? How should we conceptualize scientific discovery, in light of the pivotal role of AI systems as agents of scientific discoveries? A better understanding of the concept of scientific discovery in the age of AI-driven science is beneficial to ensuring that AI-based scientific discovery are robust, objective, and that they truly generate novel and genuine scientific knowledge.
AI, scientific objectivity, and the well-being of science
The integration of artificial intelligence (AI) into the contemporary scientific method constitutes a revolutionary change in the progress of science. Its pervasive role in scientific research includes several crucial tasks, including the collection and creation of data, their refinement and extrapolation of meaningful representations from them, the generation of new hypotheses, and experimental design and optimization. AI-based tools are also crucial for the development of automated experimentation platforms and the creation of self-driving laboratories.
In the last two years, AI has become fundamental for revolutionary scientific discoveries in various fields including physics. A prominent example is the success of NASA's Kepler space telescope, which discovered the existence of 300 exoplanets, previously completely unknown, using the deep learning classifier Exo-Miner (Valizadegan et al. 2022). The development of deep learning is also proceeding at an incredible pace with regard to its application to the detection of gravitational waves (Huerta et al. 2021). In the field of materials physics, scientists at Brookhaven National Laboratory have successfully used AI-based autonomous methods to support a self-assembly technique to discover, without any human intervention, three new nanostructures, including a first-of-its-kind structure (Doerk et al. 2023).
What does the future of science look like? What will be the impact of the increasing integration of AI methods on scientific knowledge? And can AI really achieve epistemic successes such as scientific understanding or a scientific objectivity? As Messeri and Crockett (2023) have pointed out, the risk is that AI not only does not provide greater scientific understanding, but that it does not provide scientific objectivity at all. In fact, the fear is that the type of objectivity that characterizes AI is fake because it is partial. The AI systems that are often integrated into scientific research are typically designed on the basis of an optimization process, and therefore are aimed at solving problems in the most efficient way, identifying the most likely solution, i.e. the one that provides the most accurate predictions. This factor and the ongoing process of replacing human capital with AI risk leaving science in a state of "monoculture", which is partial and even dangerous because it leads to the depletion of the fertile and diverse intellectual ground on which science has historically thrived, fueled by a variety of scientific methods and cognitive resources.
However, this risk can be mitigated and controlled. My proposal is that the Achilles heel of AI, which is its relying, typically, on an optimization process, can become an asset: the optimization process, in fact, can be designed, with due caution, in such a way that it takes into account or even manages to implement the well-being of science.
Understanding with Machine Learning
Whether machine learning (ML) algorithms can provide scientific understanding is a pressing question given their increasingly central role in the scientific method. In this project, I explore this issue through an investigation of ML’s application in accelerator research, with a particular focus on how it is implemented to the electron beam ion source (EBIS) at Brookhaven National Laboratory (BNL). More specifically, I examine how BNL physicists are employing two ML packages, on the one hand the GPTune to optimize EBIS beam intensity in real time, and on the other hand the XGBoost to analyze and interpret EBIS experimental data. This is a philosophically interesting case, as BNL scientists argue that ML algorithms not only enhance experimental optimization but also provide genuine understanding of the underlying physical system (Gu et al. 2024), a claim that may appear suspicious to philosophers. I will argue that while ML algorithms may not offer a kind of understanding that appeal to law-like explanations, they do furnish a form of understanding, understood as an intellectual virtue, which I term ‘attributive’.
The replication crisis has spawned discussions on the meaning of replication. In fact, in order to determine whether an experiment fails to replicate, it is necessary to establish what replication is. This is, however, a difficult task, as it is possible to attribute different meanings to it. This paper offers a solution to this problem of ambiguity by engineering a concept of replication that, if compared to other proposals, stands out for being not only broadly applicable but also sufficiently specific. It features a minimal level of operationalism, which would otherwise limit its applicability, while it heavily relies on replication’s specific epistemic functions, which are inter-disciplinary. Another merit is its context sensitivity, which enables it to differentiate instances of replication from non-instances of replication in every scientific discipline according to the discipline’s own standards.
Replicability is usually considered to be one of the cornerstones of science; however, the growing recognition of nonreplicable experiments and studies in scientific journals—a phenomenon that has been called ‘replicability crisis’—has spurred a debate on the meaning, function, and significance of replicability in science. Amid this discussion, it has become clear that replicability is not a monolithic concept; what is still controversial is exactly how the distinction between different kinds of replicability should be laid out terminologically and conceptually, and to what extent it bears on the more general debate on the centrality of replicability in science. This paper’s goals are to clarify the different uses of the terms related to replicability and, more importantly, to conceptually specify the kinds of replicability and their respective epistemic functions.
Edouard Machery’s article “What is Replication?” deserves particular critical attention. For if it is correct, his Resampling Account of Replication has the power to reshape the current debate on replication in psychology. Indeed, with his new proposal, philosopher Machery claims to replace the “vague characterization of replication in psychology” (Machery, 2020, p. 559) with an account that deflates one of the central debates on replication—the debate which contraposes direct and conceptual replications and asks which one is preferable. In this commentary, I argue that there are deep-rooted reasons for why the distinction is meaningful, and that the Resampling Account of Replication just offers a misleading “semantic shift.” (PsycInfo Database Record (c) 2023 APA, all rights reserved)
The Hubble constant controversy. With C. D. McCoy.
We propose that the epistemic functions of replication in science are best understood by relating them to kinds of experimental error/uncertainty. One kind of replication, which we call “direct replications,” principally serves to assess the reliability of an experiment through its precision: the presence and degree of random error/statistical uncertainty. The other kind of replication, which we call “conceptual replications,” principally serves to assess the validity of an experiment through its accuracy: the presence and degree of systematic errors/uncertainties. To illustrate the aptness of this general view, we examine the Hubble constant controversy in astronomy, showing how astronomers have responded to the concordances and discordances in their results by carrying out the different kinds of replication that we identify, with the aim of establishing a precise, accurate value for the Hubble constant. We contrast our view with Machery's “re-sampling” account of replication, which maintains that replications only assess reliability.