Marianna B. Ganapini

[This piece is coming out in a special issue of AI Magazine] Claims that artificial intelligence has reached or is approaching artificial general intelligence (AGI) are increasingly widespread in both industry and academia. These claims typically rest on either behavioral benchmark tests or mechanistic interpretability findings. The engine driving this AGI optimism (and the fears that surround this milestone) rests on a view of the mind with origins in a philosophical, cognitive theory called functionalism: the view that cognitive and conscious properties depend on functional organization rather than on physical substrates. This article briefly traces the origins of this idea and examines some recent developments in this tradition that focus on AI welfare and even on possible AI rights. The conclusion of the paper is an invitation to consider the extraordinary consequences of extending welfare and even moral worth to non-biological, non-living systems. These consequences arguably demand a higher standard of epistemic justification in putting forward claims about AI consciousness and AGI than the current discourse, both in industry and in academia, has met.

[Preprint] The rise of Large Language Models (LLMs) has sparked debate about whether these systems exhibit human-level cognition. In this debate, little attention has been paid to a structural component of human cognition: core beliefs, truths that provide a foundation around which we could build a stable worldview. These commitments resist virtually any attempt at debunking, as abandoning them would represent a fundamental shift in how we see reality. In this paper, we ask whether LLMs hold anything akin to core commitments. Using systematic probes across five domains (science, history, geography, biology, and mathematics), we find that most LLMs fail to maintain a stable representation of reality. Though some recent models, tested under the identical protocol, showed improved stability on some key commitments, they still failed under pressure. These results document an improvement in argumentative skills across model generations but indicate that all current models fall short of human-level cognition.

This paper addresses the following question: under what conditions is online content moderation permissible? Often stringent content moderation policies are invoked when online misinformation is linked to serious harm. Here I show the complexities involved in allowing for this kind of content moderation. The conclusion will be that, when a clear link between misinformation and serious harm cannot be established, social media companies may instead adopt lighter forms of content moderation, such as accuracy nudging, to curb misinformation.

Il libro è uno strumento per orientarsi nel complesso mondo dell’intelligenza artificiale (IA) e delle sue implicazioni etiche. La tesi centrale è che il futuro dell’IA sia duplice: l’IA sarà parte di noi, ma anche altro da noi. Questa ambivalenza richiede di superare narrazioni apocalittiche o entusiastiche. L’IA riflette le tensioni del nostro tempo: tra controllo e delega, tra tecnica e umanità. Per affrontarla serve un atteggiamento critico e responsabile, capace di definire e difendere valori fondamentali. Al centro di questa relazione c’è il tema della fiducia: possiamo davvero fidarci dell’IA? Quando è tale fiducia giustificata? Per capirlo, ci concentreremo sui temi dei diritti e della giustizia. Solo un’IA che protegga la nostra privacy e autonomia e che abbia un impatto equo sulla società sarà degna di fiducia. Questi sono i criteri con cui orientare il nostro rapporto, sempre più complesso, con le macchine.

Inspired by the ”thinking fast and slow” cognitive theory of human decision making, we propose a multi-agent cognitive architecture (SOFAI) that is based on ”fast”/”slow” solvers and a metacognitive module. We then present experimental results on the behavior of an instance of this architecture for AI systems that make decisions about navigating in a constrained environment. We show that combining the two decision modalities through a separate metacognitive function allows for higher decision quality with less resource consumption compared to employing only one of the two modalities. Analyzing how the system achieves this, we also provide evidence for the emergence of several human-like behaviors, including skill learning, adaptability, and cognitive control. Authors: M. Bergamaschi Ganapini, M. Campbell, F. Fabiano, L. Horesh, J. Lenchner, A. Loreggia, N. Mattei, F. Rossi, B. Srivastava & K. B. Venable

Emotions-recognition technology is a developing field and is likely to become an incredibly lucrative industry. The premise behind this technology is that it will provide some objective ways to understand and access human emotions. Contrary to this, I show that the current debate around AI emotion recognition misses a crucial aspect of folk psychology, namely that it is predominantly a social practice through which we shape minds and behaviors. Once we recognize that, we face the following dilemma: either we conclude that AI is not in the business of recognizing emotions after all or, if we say it is, then AI assessments are not only not value-free but are in fact likely to reproduce social stereotypes and discrimination.

AI technologies have recently shown remarkable capabilities in various scientific fields, such as drug discovery, medicine, climate modeling, and archaeology, primarily through their pattern recognition abilities. They can also generate hypotheses and suggest new research directions. While acknowledging AI’s potential to aid in scientific breakthroughs, the paper shows that current AI models do not meet the criteria for making independent scientific discoveries. Discovery is an epistemic achievement that requires a level of competence and self-reflectivity that AI does not yet possess.

Artificial intelligent technologies are utilized to provide online personalized recommendations, suggestions, or prompts that can influence people's decision-making processes. We call this AI-enhanced nudging (or AI-nudging for short). Contrary to the received wisdom we claim that AI-enhanced nudging is not necessarily morally problematic. To start assessing the risks and moral import of AI-nudging we believe that we should adopt a risk-factor analysis: we show that both the level of risk and possibly the moral value of adopting AI-nudging ultimately depend on the complexity of the “moral situation” in which various risk-factors play a causal role in determining levels of uncertainty.

The rapid integration of these artificial intelligence (AI) tools into our daily lives is reshaping how we think and make decisions. We propose that data-driven AI systems, by transcending individual artefacts and interfacing with a dynamic, multiartefact ecosystem, constitute a distinct psychological system. We call this ‘system 0’, and position it alongside Kahneman’s system 1 (fast, intuitive thinking) and system 2 (slow, analytical thinking).System 0 represents the outsourcing of certain cognitive tasks to AI, which can process vast amounts of data and perform complex computations beyond human capabilities. It emerges from the interaction between users and AI systems, which creates a dynamic, personalized interface. Authors: Massimo Chiriatti, Marianna Ganapini, Enrico Panai, Mario Ubiali & Giuseppe Riva.

PENULTIMATE DRAFT. At times, weird stories such as the Pizzagate spread surprisingly quickly and widely. In this paper I analyze the mental attitudes of those who seem to take those absurdities seriously: I argue that those stories are often imagined rather than genuinely believed. Then I make room for the claim that often these imaginings are used to support group ideologies. My main contribution is to explain how that support actually happens by showing that motivated cognition can employ imagination as a seemingly rational tool to reinforce and protect beliefs.

Normativism in the theory of meaning and content is the view that linguistic meaning and/or intentional content are essentially normative. As both normativity and its essentiality to meaning/content can be interpreted in a number of different ways, there is now a whole family of views laying claim to the slogan “meaning/content is normative”. In this essay, we discuss a number of central normativist theses, and we begin by identifying different versions of meaning normativism, presenting the arguments that have been put forth for and against them.

Normativism in the theory of meaning and content is the view that linguistic meaning and/or intentional content are essentially normative. As both normativity and its essentiality to meaning/content can be interpreted in a number of different ways, there is now a whole family of views laying claim to the slogan “meaning/content is normative”. In this essay, we discuss a number of central normativist theses, and we begin by identifying different versions of meaning normativism, presenting the arguments that have been put forth for and against them.

Many agree that one cannot consciously form a belief just because one wants to. And many also agree this is a puzzling component of our conscious belief-forming processes. I will look at three views on how to make sense of this puzzle and show that they all fail in some way. I then offer a simpler explanation that avoids all the pitfalls of those views, which is based instead on an analysis of our conscious reasoning combined with a commonly accepted account of the concept of belief. I conclude that no epistemic norm or aim is actually needed to explain why we cannot deliberatively believe whatever we want.

[Multiple authors] In this paper, we propose a general architecture that is based on fast/slow solvers and a metacognitive component. We then present experimental results on the behavior of an instance of this architecture, for AI systems that make decisions about navigating in a constrained environment. We show how combining the fast and slow decision modalities allows the system to evolve over time and gradually pass from slow to fast thinking with enough experience, and that this greatly helps in decision quality, resource consumption, and efficiency.

Multiple Authors - please see paper attached. AI systems have seen dramatic advancement in recent years, bringing many applications that pervade our everyday life. However, we are still mostly seeing instances of narrow AI: many of these recent developments are typically focused on a very limited set of competencies and goals, e.g., image interpretation, natural language processing, classification, prediction, and many others. We argue that a better study of the mechanisms that allow humans to have these capabilities can help us understand how to imbue AI systems with these competencies. We focus especially on D. Kahneman’s theory of thinking fast and slow, and we propose a multi-agent AI architecture (called SOFAI, for SlOw and Fast AI) where incoming problems are solved by either system 1 (or "fast") agents (also called "solvers"), that react by exploiting only past experience, or by system 2 (or "slow") agents, that are deliberately activated when there is the need to reason and search for optimal solutions beyond what is expected from the system 1 agent.

Why do people share or publicly engage with fake stories? Two possible answers come to mind: (a) people are deeply irrational and believe these stories to be true; or (b) they intend to deceive their audience. Both answers presuppose the idea that people put the stories forward as true. But I argue that in some cases, these outlandish (yet also very popular) stories function as signals of one's group membership. This signaling function can make better sense of why, despite their unusual nature or lack of a factual basis, some of these stories are so widespread.

Many agree that one cannot consciously form a belief just because one wants to. And many also agree this is a puzzling component of our conscious belief-forming processes. I will look at three views on how to make sense of this puzzle and show that they all fail in some way. I then offer a simpler explanation that avoids all the pitfalls of those views, which is based instead on an analysis of our conscious reasoning combined with a commonly accepted account of the concept of belief. I conclude that no epistemic norm or aim is actually needed to explain why we cannot deliberatively believe whatever we want.

In this paper, I will focus on a type of confabulation that emerges in relation to questions about mental attitudes whose causes we cannot introspectively access. I argue against two popular views that see confabulations as mainly offering a psychological story about ourselves. On these views, confabulations are the result of either a cause-tracking mechanism or a self-directed mindreading mechanism. In contrast, I propose the view that confabulations are mostly telling a normative story: they are arguments primarily offered to justify one’s attitudes, and they are produced by our argumentative reasoning mechanism driven by the biological goal of presenting ourselves as good reasoners and as reliable sources of information.

One main argument that has been offered in support of the Knowledge Account of Assertion is that it successfully makes sense of a variety of Moorean-paradoxical claims. David Sosa has objected to the Knowledge Account by arguing that it does not generalize satisfactorily to make sense of the oddity of iterated conjunctions of the form “p but I don’t know whether I know that p”. Recently, Martin Montminy has offered a defense of the Knowledge Account. In this paper, I show that both Montminy’s and Sosa’s arguments fail. First, I argue that Montminy does not offer a good reply to Sosa; then I show that Sosa’s objection actually does not constitute a real threat to the Knowledge Account.

The error theory is a metaethical theory that maintains that normative judgments are beliefs that ascribe normative properties, and that these properties do not exist. In a recent paper, Bart Streumer argues that it is impossible to fully believe the error theory. Surprisingly, he claims that this is not a problem for the error theorist: even if we can’t fully believe the error theory, the good news is that we can still come close to believing the error theory. In this paper I show that Streumer’s arguments fail. First, I lay out Streumer’s argument for why we can’t believe the error theory. Then, I argue against the unbelievability of the error theory. Finally, I show that Streumer’s positive proposal that we can come close to believing the error theory is actually undermined by his own argument for why we can’t believe the error theory.