Alessandro Facchini

How can we ensure that AI systems are aligned with human values and remain safe? We can study this problem through the frameworks of the AI assistance and the AI shutdown games. The AI assistance problem concerns designing an AI agent that helps a human to maximise their utility function(s). However, only the human knows these function(s); the AI assistant must learn them. The shutdown problem instead concerns designing AI agents that: shut down when a shutdown button is pressed; neither try to prevent nor cause the pressing of the shutdown button; and otherwise accomplish their task competently. In this paper, we show that addressing these challenges requires AI agents that can reason under uncertainty and handle both incomplete and non-Archimedean preferences.

The research program of eXplainable AI (XAI) has been developed with the aim of providing tools and methods for reducing opacity and making AI systems more humanly understandable. Unfortunately, the majority of XAI scholars actually classify a system as more or less opaque by confronting it with traditional AI systems such as linear regression models or rules-based systems, which are usually assumed to be the prototype of transparent systems. In doing so, the concept of opacity remains unexplained. To overcome this issue, we view opacity as a concept whose meaning depends on the context of application, and on the purposes and characteristics of its users. Based on this, in this work, we distinguish between access opacity, link opacity and semantic opacity, hence providing the groundwork for a taxonomy of the concept of opacity for AI systems.

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.

Background: The integration of artificial intelligence (AI) into health care has become a crucial element in the digital transformation of health systems worldwide. Despite the potential benefits across diverse medical domains, a significant barrier to the successful adoption of AI systems in health care applications remains the prevailing low user trust in these technologies. Crucially, this challenge is exacerbated by the lack of consensus among experts from different disciplines on the definition of trust in AI within the health care sector. Objective: We aimed to provide the first consensus-based analysis of trust in AI in health care based on an interdisciplinary panel of experts from different domains. Our findings can be used to address the problem of defining trust in AI in health care applications, fostering the discussion of concrete real-world health care scenarios in which humans interact with AI systems explicitly. Methods: We used a combination of framework analysis and a 3-step consensus process involving 18 international experts from the fields of computer science, medicine, philosophy of technology, ethics, and social sciences. Our process consisted of a synchronous phase during an expert workshop where we discussed the notion of trust in AI in health care applications, defined an initial framework of important elements of trust to guide our analysis, and agreed on 5 case studies. This was followed by a 2-step iterative, asynchronous process in which the authors further developed, discussed, and refined notions of trust with respect to these specific cases. Results: Our consensus process identified key contextual factors of trust, namely, an AI system’s environment, the actors involved, and framing factors, and analyzed causes and effects of trust in AI in health care. Our findings revealed that certain factors were applicable across all discussed cases yet also pointed to the need for a fine-grained, multidisciplinary analysis bridging human-centered and technology-centered approaches. While regulatory boundaries and technological design features are critical to successful AI implementation in health care, ultimately, communication and positive lived experiences with AI systems will be at the forefront of user trust. Our expert consensus allowed us to formulate concrete recommendations for future research on trust in AI in health care applications. Conclusions: This paper advocates for a more refined and nuanced conceptual understanding of trust in the context of AI in health care. By synthesizing insights into commonalities and differences among specific case studies, this paper establishes a foundational basis for future debates and discussions on trusting AI in health care.

Based on a gambling formulation of quantum mechanics, we derive a Gleason-type theorem that holds for any dimension n of a quantum system, and in particular for \. The theorem states that the only logically consistent probability assignments are exactly the ones that are definable as the trace of the product of a projector and a density matrix operator. In addition, we detail the reason why dispersion-free probabilities are actually not valid, or rational, probabilities for quantum mechanics, and hence should be excluded from consideration.

We investigate a typology of socially and ethically risky phenomena emerging from the interaction between humans and large language model (LLM)-based conversational systems. As they relate to the way in which humans attribute social identity components, such as role and face, to LLM-based conversational systems, we term these phenomena 'social misattributions.' Drawing on classical theories of social identity and recent debates in the philosophy of technology, we argue that these social misattributions represent higher-order forms of anthropomorphisation of LLM-based conversational systems that are not justified by their technical capabilities and follow from the social context of conversational interactions. We discuss the risks these misattributions pose to human users, including emotional manipulation and unwarranted trust, and propose mitigation strategies. Our recommendations emphasise the importance of fostering social transparency and exploring approaches, such as frictional design, that are currently promoted in the research domain of human-centred artificial intelligence.

Human-centered explainable AI (HCXAI) advocates for the integration of social aspects into AI explanations. Central to the HCXAI discourse is the Social Transparency (ST) framework, which aims to make the socio-organizational context of AI systems accessible to their users. In this work, we suggest extending the ST framework to address the risks of social misattributions in Large Language Models (LLMs), particularly in sensitive areas like mental health. In fact LLMs, which are remarkably capable of simulating roles and personas, may lead to mismatches between designers' intentions and users' perceptions of social attributes, risking to promote emotional manipulation and dangerous behaviors, cases of epistemic injustice, and unwarranted trust. To address these issues, we propose enhancing the ST framework with a fifth 'W-question' to clarify the specific social attributions assigned to LLMs by its designers and users. This addition aims to bridge the gap between LLM capabilities and user perceptions, promoting the ethically responsible development and use of LLM-based technology.

We show that the modal µ-calculus over GL collapses to the modal fragment by showing that the fixpoint formula is reached after two iterations and answer to a question posed by van Benthem in [4]. Further, we introduce the modal µ~-calculus by allowing fixpoint constructors for any formula where the fixpoint variable appears guarded but not necessarily positive and show that this calculus over GL collapses to the modal fragment, too. The latter result allows us a new proof of the de Jongh, Sambin Theorem and provides a simple algorithm to construct the fixpoint formula.

Since the so-called Deep Learning revolution, recent advances in Artificial Intelligence (AI) obtained through Machine Learning (ML), that would have been impossible without the accumulation of huge quantities of data and continually improving and accelerating computer hardware, are promising to change much of what we do and how we do it in domains as different as scientific research, education, arts, and management. Those promises, which with the wide accessibility of tools stemming from generative AI have suddenly become very tangible for everyone, have sparked many hopes and expectations, such as the possibility of enhancing productivity and tackling complex problems such as climate change and poverty. At the same time, many ethical and more broadly societal concerns have emerged. In this chapter, we will try to make sense of these fears and see how we can try to counter them. First of all, we will clarify what AI and ML are, what makes ML systems and models so peculiar, and for which tasks and purposes they can be used, taking examples from domains such as management. Elucidating the nature (ontology) of these technological artefacts and the whole ML development cycle will then enable us to clarify where the critical points lie, both from an ethical but also cognitive and epistemic perspectives. In fact, among other things, what one can observe is that often, when deployed, AI systems are under-used or not used at all, one of the reasons commonly put forward being that it is difficult for humans to estimate to what extent to trust outputs coming from algorithms whose behavior is difficult, if not impossible, to understand. We will then see that one way to overcome such issues is to advocate for a human-centered perspective on the design and development of ML systems, in which general ethical principles but also the specific needs, concerns, and values (ethical, epistemic, and cognitive) of the various stakeholders should be taken into account in the design, development, and deployment of a ML model. In doing so, we align with the core tenets of humanistic management.

The high predictive accuracy of contemporary machine learning-based AI systems has led some scholars to argue that, in certain cases, we should grant them epistemic expertise and authority over humans. This approach suggests that humans would have the epistemic obligation of relying on the predictions of a highly accurate AI system. Contrary to this view, in this work we claim that it is not possible to endow AI systems with a genuine account of epistemic expertise. In fact, relying on accounts of expertise and authority from virtue epistemology, we show that epistemic expertise requires a relation with understanding that AI systems do not satisfy and intellectual abilities that these systems do not manifest. Further, following the Distribution Cognition theory and adapting an account by Croce on the virtues of collective epistemic agents to the case of human-AI interactions we show that, if an AI system is successfully appropriated by a human agent, a _hybrid_ epistemic agent emerges, which can become both an epistemic expert and an authority. Consequently, we claim that the aforementioned hybrid agent is the appropriate object of a discourse around trust in AI and the epistemic obligations that stem from its epistemic superiority.

This paper establishes model-theoretic properties of \, a variation of monadic first-order logic that features the generalised quantifier \. We will also prove analogous versions of these results in the simpler setting of monadic first-order logic with and without equality and \, respectively). For each logic \ we will show the following. We provide syntactically defined fragments of \ characterising four different semantic properties of \-sentences: being monotone and continuous in a given set of monadic predicates; having truth preserved under taking submodels or being truth invariant under taking quotients. In each case, we produce an effectively defined map that translates an arbitrary sentence \ to a sentence \ belonging to the corresponding syntactic fragment, with the property that \ is equivalent to \ precisely when it has the associated semantic property. As a corollary of our developments, we obtain that the four semantic properties above are decidable for \-sentences.

Anthropomorphisation -- the phenomenon whereby non-human entities are ascribed human-like qualities -- has become increasingly salient with the rise of large language model (LLM)-based conversational agents (CAs). Unlike earlier chatbots, LLM-based CAs routinely generate interactional and linguistic cues, such as first-person self-reference, epistemic and affective expressions that empirical work shows can increase engagement. On the other hand, anthropomorphisation raises ethical concerns, including deception, overreliance, and exploitative relationship framing, while some authors argue that anthropomorphic interaction may support autonomy, well-being, and inclusion. Despite increasing interest in the phenomenon, literature remains fragmented across domains and varies substantially in how it defines, operationalizes, and normatively evaluates anthropomorphisation. This scoping review maps ethically oriented work on anthropomorphising LLM-based CAs across five databases and three preprint repositories. We synthesize (1) conceptual foundations, (2) ethical challenges and opportunities, and (3) methodological approaches. We find convergence on attribution-based definitions but substantial divergence in operationalization, a predominantly risk-forward normative framing, and limited empirical work that links observed interaction effects to actionable governance guidance. We conclude with a research agenda and design/governance recommendations for ethically deploying anthropomorphic cues in LLM-based conversational agents.

We study the strictness of the modal μ-calculus hierarchy over some restricted classes of transition systems. First, we prove that over transitive systems the hierarchy collapses to the alternationfree fragment. In order to do this the finite model theorem for transitive transition systems is proved. Further, we verify that if symmetry is added to transitivity the hierarchy collapses to the purely modal fragment. Finally, we show that the hierarchy is strict over reflexive frames. By proving the finite model theorem for reflexive systems the same results holds for finite models.

The high predictive accuracy of contemporary machine learning-based AI systems has led some scholars to argue that, in certain cases, we should grant them epistemic expertise and authority over humans. This approach suggests that humans would have the epistemic obligation of relying on the predictions of a highly accurate AI system. Contrary to this view, in this work we claim that it is not possible to endow AI systems with a genuine account of epistemic expertise. In fact, relying on accounts of expertise and authority from virtue epistemology, we show that epistemic expertise requires a relation with understanding that AI systems do not satisfy and intellectual abilities that these systems do not manifest. However, recent developments of large language models make us argue that in the future AI systems may endowed with a form of epistemic authority. Further, following the Distributed Cognition theory and adapting an account by Croce on the virtues of collective epistemic agents to the case of human-AI interactions we show that, if an AI system is successfully appropriated by a human agent, a hybrid epistemic agent emerges, which can become both an epistemic expert and an authority. Consequently, we claim that the aforementioned hybrid agent is the appropriate object of a discourse around trust in AI and the epistemic obligations that stem from its epistemic superiority.

We argue that there is a simple, unique, reason for all quantum paradoxes, and that such a reason is not uniquely related to quantum theory. It is rather a mathematical question that arises at the intersection of logic, probability, and computation. We give our ‘weirdness theorem’ that characterises the conditions under which the weirdness will show up. It shows that whenever logic has bounds due to the algorithmic nature of its tasks, then weirdness arises in the special form of negative probabilities or non-classical evaluation functionals. Weirdness is not logical inconsistency, however. It is only the expression of the clash between an unbounded and a bounded view of computation in logic. We discuss the implication of these results for quantum mechanics, arguing in particular that its interpretation should ultimately be computational rather than exclusively physical. We develop in addition a probabilistic theory in the real numbers that exhibits the phenomenon of entanglement, thus concretely showing that the latter is not specific to quantum mechanics.

Human-AI complementarity is the claim that a human supported by an AI system can outperform either alone in a decision-making process. Since its introduction in the human-AI interaction literature, it has gained traction by generalizing the reliance paradigm and by offering a more practical alternative to the contested construct of 'trust in AI.' Yet complementarity faces key theoretical challenges: it lacks precise theoretical anchoring, it is formalized just as a post hoc indicator of relative predictive accuracy, it remains silent about other desiderata of human-AI interactions and it abstracts away from the magnitude-cost profile of its performance gain. As a result, complementarity is difficult to obtain in empirical settings. In this work, we leverage epistemology to address these challenges by reframing complementarity within the discourse on justificatory AI. Drawing on computational reliabilism, we argue that historical instances of complementarity function as evidence that a given human-AI interaction is a reliable epistemic process for a given predictive task. Together with other reliability indicators assessing the alignment of the human-AI team with the epistemic standards and socio-technical practices, complementarity contributes to the degree of reliability of human-AI teams when generating predictions. This supports the practical reasoning of those affected by these outputs -- patients, managers, regulators, and others. In summary, our approach suggests that the role and value of complementarity lies not in providing a relative measure of predictive accuracy, but in helping calibrate decision-making to the reliability of AI-supported processes that increasingly shape everyday life.