Luca Rivelli

This paper sketches a theoretical conception of modularity by generalizing Herbert Simon's idea of near-decomposability, showing that it could constitute a framework for the unification of the notion of modularity in the history and philosophy of biology. To put the framework to test, first a main problematic area is highlighted--the evolution of modularity--focusing on a historical and rational reconstruction of two ways of viewing it that appeared in the second half of the 20th century: one, due to Herbert Simon, rooted in a generic Darwinian mindset, the other, by Stuart Kauffman, inspired by a systemic-oriented approach tending to demote the importance of natural selection. It is shown that, under the light of the general view of modularity proposed here, these two apparently incompatible views can be interpreted as fundamentally homologous. The paper then engages with some current prominent views on modularity in biology, in order to show that the proposed framework is largely compatible with them, and able to accommodate cases of emergent modularity. The facilitating role of modularity in mechanistic and functional explanations is also highlighted. As a conclusion, it seems the proposed sketch of a theoretical view of modularity, open to further improvement, already shows potentiality as a unifying framework for the notion of modularity in philosophy and history of biology, and possibly, other disciplines.

This paper addresses the accelerating crisis of ethical governance in an age of complex socio-technical change, particularly in the domain of Artificial Intelligence. It poses a foundational philosophical question: when, if ever, is AI assistance in ethical deliberation legitimate? An answer is developed through three theses: i) the Ethical No-Free-Lunch (ENFL) principle, which establishes the indispensability of human normative intervention and accountability; ii) the Discovery/Justification Separation inspired by Reichenbach's work, which restricts AI use to the exploratory "context of discovery"; iii) the Algorithmic Mediated Control Framework (AMCF), which mandates that only scrutable, human-vetted deterministic algorithms generated with AI assistance, and not the AI itself, be entrusted with critical societal processes. From these theses, five legitimacy criteria for AI-assisted ethical deliberation are derived. Finally, the paper proposes the "AI-assisted Iterative Method for Ethical Deliberation" (AIMED), an actionable multi-stage workflow that fulfills the exposed criteria for ethical AI-assisted deliberation. This method integrates digital literature analysis, structured human–AI dialogue, human-only verification, and continuous feedback. The paper explicitly addresses several potential objections. It is shown how the AIMED framework aligns with and provides a concrete implementation for major international regulatory guidelines, such as the EU AI Act and the NIST AI Risk Management Framework. By situating the AIMED within traditions of proceduralism, the governance of inductive risk, and human–AI collaboration, the paper argues that this framework offers a philosophically justified, practically implementable model of AI-assisted ethical governance, that can be seen as an actionable instance of Digital Humanism.

A conditional argument is put forth suggesting that if qualia have a functional role in intelligence, then it might be possible, by observing the behavior of verbal AI systems like large language models (LLMs) or other architectures capable of verbal reasoning, to tackle in an empirical way the "strong AI" problem, namely, the possibility that AI systems have subjective experiences, or qualia. The basic premise is that if qualia are functional, and thus have causal roles, then they could affect the production of discourses about qualia and subjective consciousness in general. A thought experiment is put forth envisioning a possible method to probabilistically test the presence of qualia in AI systems based on this conditional argument. The method proposed in the thought experiment focuses on observing whether ideas related to the issue of phenomenal consciousness, such as the so-called "hard problem" of consciousness, or related philosophical issues centered on qualia, spontaneously emerge in extended dialogues involving LLMs specifically trained to be initially oblivious of such philosophical concept and related ones. By observing the emergence (or lack thereof) in the AI's verbal production of discussions related to phenomenal consciousness in these contexts, the method seeks to provide empirical evidence for or against the existence of consciousness in AI. An outline of a Bayesian test of the hypothesis is provided. Three main investigative methods with different reliability and feasibility aimed at empirically detecting AI consciousness are proposed: one involving human interaction and two fully automated, consisting in multi-agent conversations between machines. The practical and philosophical challenges involved by the idea of transforming the proposed thought experiments into an actual empirical trial are then discussed. In light of these considerations, the proposals put forth in the paper appears to be at least a contribution to computational philosophy in the form of philosophical thought experiments focused on computational systems, aimed at refining our philosophical understanding of consciousness. Hopefully, it could also provide hints toward future empirical investigations into machine consciousness.

This work is mainly concerned with the notion of hierarchical modularity and its use in explaining structure and dynamical behavior of complex systems by means of hierarchical modular models, as well as with a concept of my proposal, antimodularity, tied to the possibility of the algorithmic detection of hierarchical modularity. Specifically, I highlight the pragmatic bearing of hierarchical modularity on the possibility of scientific explanation of complex systems, that is, systems which, according to a chosen basic description, can be considered as composed of elementary, discrete, interrelated parts. I stress that hierarchical modularity is also required by the experimentation aimed to discover the structure of such systems. Algorithmic detection of hierarchical modularity turns out to be a task plagued by the demonstrated computational intractability of the search for the best hierarchical modular description, and by the high computational expensiveness of even approximated detection methods. Antimodularity consists in the lack of a modular description fitting the needs of the observer, a lack due either to absence of modularity in the system’s chosen basic description, or to the impossibility, due to the excessive size of the system under assessment in relation to the computational cost of algorithmic methods, to algorithmically produce a valid hierarchical description. I stress that modularity and antimodularity depend on the pragmatic choice of a given basic description of the system, a choice made by the observer based on explanatory goals. I show how antimodularity hinders the possibility of applying at least three well-known types of explanation: mechanistic, deductive-nomological and computational. A fourth type, topological explanation, remains unaffected. I then assess the presence of modularity in biological systems, and evaluate the possible consequences, and the likelihood, of incurring in antimodularity in biology and other sciences, concluding that this eventuality is quite likely, at least in systems biology. I finally indulge in some metaphysical and historical speculations: metaphysically, antimodularity seems to suggest a possible position according to which natural kinds are detected modules, and as such, due to the computational hardness of the detection of the best hierarchical modular description, they are unlikely to be the best possible way to describe the world, because the modularity of natural kinds quite probably does not reflect the best possible modularity of the world. From an historical point of view, the growing use of computational methods for modularity detection or simulation of complex systems, especially in certain areas of scientific research, hints at the envisioning of a multiplicity of emerging scientific disciplines guided by a self- sustained, growing production of possibly human-unintelligible explanations. This, I suggest, would constitute an historical change in science, which, if has not already occurred, could well be on the verge of happening.

Empirical philosophers of science aim to base their philosophical theories on observations of scientific practice. But since there is far too much science to observe it all, how can we form and test hypotheses about science that are sufficiently rigorous and broad in scope, while avoiding the pitfalls of bias and subjectivity in our methods? Part of the answer, we claim, lies in the computational tools of the digital humanities, which allow us to analyze large volumes of scientific literature. Here we advocate for the use of these methods by addressing a number of large-scale, justificatory concerns—specifically, about the epistemic value of journal articles as evidence for what happens elsewhere in science, and about the ability of DH tools to extract this evidence. Far from ignoring the gap between scientific literature and the rest of scientific practice, effective use of DH tools requires critical reflection about these relationships.

This work is concerned with hierarchical modular descriptions, their algorithmic production, and their importance for certain types of scientific explanations of the structure and dynamical behavior of complex systems. Networks are taken into consideration as paradigmatic representations of complex systems. It turns out that algorithmic detection of hierarchical modularity in networks is a task plagued in certain cases by theoretical intractability and in most cases by the still high computational complexity of most approximated methods. A new notion, antimodularity, is then proposed, which consists in the impossibility to algorithmically obtain a modular description fitting the explanatory purposes of the observer for reasons tied to the computational cost of typical algorithmic methods of modularity detection, in relation to the excessive size of the system under assessment and to the required precision. It turns out that occurrence of antimodularity hinders both mechanistic and functional explanation, by damaging their intelligibility. Another newly proposed more general notion, explanatory emergence, subsumes antimodularity under any case in which a system resists intelligible explanations because of the excessive computational cost of algorithmic methods required to obtain the relevant explanatory descriptions from the raw data. The possible consequences, and the likelihood, of incurring in antimodularity or explanatory emergence in the actual scientific practice are finally assessed, concluding that this eventuality is possible, at least in disciplines which are based on the algorithmic analysis of big data. The present work aims to be an example of how certain notions of theoretical computer science can be fruitfully imported into philosophy of science.

In this paper I will reconstruct and analyze a famous argument by Stuart Kauffman about complex systems and evolution, in order to highlight the use in theoretical biology of a kind of non-mechanistic and non-causal explanation which I propose to call, following Kauffman, ensemble explanation. The aim is to contribute to the ongoing philosophical debate about non-causal explanations in the special sciences, kinds of explanation apparently extraneous to the received causal-mechanistic view. Ensemble explanations resemble quite closely the explanations of the structural kind proposed by Philippe Huneman, which—unlike mechanistic explanations—do not explain by exposing the causal structure of a phenomenon, but...

Some scholars claim that epistemology of science and machine learning are actually overlapping disciplines studying induction, respectively affected by Hume's problem of induction and its formal machine-learning counterpart, the “no-free-lunch” (NFL) theorems, to which even advanced AI systems such as LLMs are not immune. Extending Kevin Korb's view, this paper envisions a hierarchy of disciplines where the lowermost is a basic science, and, recursively, the metascience at each level inductively learns which methods work best at the immediately lower level. Due to Hume's dictum and NFL theorems, no exact metanorms for the good performance of each object science can be obtained after just a finite number of levels up the hierarchy, and the progressive abstractness of each metadiscipline and consequent ill-definability of its methods and objects makes science—as defined by a minimal standard of scientificity—cease to exist above a certain metalevel, allowing for a still rational style of inquiry into science that can be called “philosophical.” Philosophical levels, transitively reflecting on science, peculiarly manifest a non–empirically learned urge to self-reflection constituting the properly normative aspect of philosophy of science.