Andy Clark: Generative Models, Extended Minds, and the Future of Human Intelligence

Suppose that human brains are organs of prediction, in which a generative model delivers predictions of sensory flows. How might brains like that breed extended minds? And what happens when we then populate our worlds with artificial systems (Generative AIs) whose bedrock operating principle is also generative-model based prediction – typically applied to very large data sets? Do they become parts of our extended minds too? In this talk, I explore this territory via three observations. First, the fundamental drive to use perception and action to minimize long-term average prediction error neatly merges epistemic (knowledge-seeking) and practical concerns from the very start – so acting to gain information and acting to achieve a goal are generated in exactly the same way and for the same reasons. This is why predictive brains breed what I have sometimes called ‘extended minds’ – minds whose machinery extends beyond that of the biological brain. Second, in advanced agents this fundamental merging of practical and epistemic concerns slowly synergized with the use of sketches, diagrams, material symbols, and (more broadly speaking) culture and science. Third, these new tricks opened up whole new realms as advanced agents also learn ways to encounter, refine, and challenge their own internal predictive models by turning them into public objects apt for sharing, stress-testing, and ‘productive breaking’. Art, science, and culture emerged as key tools for both hacking our own predictive brains. Hacking the predictive brain while creating extended thinking systems (potentially involving generative AIs in both key roles) poises human thought and reason for an unlimited journey.

David J. Chalmers: Do Large Language Models extend the Mind?

Gloria Andrada & Richard Menary: Cognitive Injustice
Do unfair social systems harm human cognition? Recent research on epistemic injustice claims that it does. In this paper we introduce a form of injustice that we call cognitive injustice  and that has been overlooked in the literature. We consider cognitive injustice to be continuous with epistemic injustice, but quite distinctive in nature. Cognitive injustice happens when unjust social and material environments attenuate the cognitive abilities that an agent develops in virtue of the process of enculturation. It is fundamental because it affects the very formation of cognitive abilities. We argue that cognitive injustice can take at least three different forms: i) Cognitive abilities are not developed, ii) they are not developed properly, and  iii) the environment is such that agents cannot exercise those abilities.

Ned Block: Is Consciousness extended and if not, why not?
This talk will argue that consciousness requires connections of a sort that do not exist between the brain and the environment, precluding extended consciousness.

Guido Cassinadri & Marco Fasoli: Rejecting the Extended Narrative: A Critique of Two Normative Arguments for Extended Cognition
Given the explanatory stalemate between ‘embedded’ (EMB) and ‘extended’ (EXT) cognition, various authors have proposed normative arguments to overcome such a deadlock in favour of EXT. According to what we call the “extended narrative” (EXT narrative, see Cassinadri 2022), we should embrace EXT and dismiss EMB, because the former leads to morally preferable consequences with respect to the latter. In this article we argue that two crucial arguments provided by these authors fall short and that, given the current lack of "a mark of the cognitive”, their failure prompts us to provisionally embrace EMB when assessing tool users. In sections 1.1 and 1.2, we present respectively King (2016) and Vold’s (2018) ‘argument from assessment of capacities’ (AAC) and Clowes (2013), Farina and Lavazza’s (2022a) ‘cognitive diminishment argument’ (CDA). The former states that EXT is better at attributing cognitive credit to individuals with learning disabilities who use assistive tools to complete their learning tasks, whereas the latter states that EMB implies the cognitive diminishment of the agent. In section 2.1, we argue that AAC and CDA assume a flawed characterisation of EMB based on a ‘principle of intracranialism’ (PI), which ignores the complex dynamics involved in cognitive integration. Moreover, in section 2.2, we argue that AAC and CDA suffer from the ‘agential bias’, by failing to acknowledge that human agency and cognition are characterised by a relational dependence on external resources. In section 3.1, we present more complex scenarios of tool-use in order to demonstrate that the EXT narrative is uninformative in addressing these cases, since its deployment of EXT effectively produces a black box. To conclude, in section 3.2 we argue that AAC and CDA fail to present EXT as descriptively and normatively superior to EMB.

Frederique de Vignemont: canceled

Zoe Drayson: Is Language still the Ultimate Artefact?
Clark and Chalmers (1998) characterize language as “a central means by which cognitive processes are extended into the world”. Language plays this role in virtue of being an external tool which supplements the brain’s computational processes rather than changing them (Clark 1998). In this paper, I explore the extent to which this artefactual view of language is compatible with Clark’s current take on the brain as a predictive processor. Clark (2022) proposes that predictive brains enable extended minds: the predictive neural architecture is what allows our brain to recruit external artefacts into our cognitive processes. I draw on recent work in linguistics and artificial intelligence to suggest that on this picture, our linguistic competence depends more heavily on internal predictive processes than on external symbolic artefacts. I explore the extent to which predictive models of language competence undermine some of the original arguments for extended cognition.

Luis H. Favela: Empirical Evidence for Extended Cognitive Systems
We present an empirically supported theoretical and methodological framework for quantifying the system-level properties of person-plus-tool interactions in order to answer the question: “Are person- plus-tool-systems extended cognitive systems?” Nineteen participants provided perceptual judgments regarding their ability to pass through apertures of various widths while using visual information, blind- folded wielding a rod, or blindfolded wielding an Enactive Torch—a vibrotactile sensory-substitution device for detecting distance. Monofractal, multifractal, and recurrence quantification analyses were conducted to assess features of person-plus-tool movement dynamics. Trials where people utilized the rod or Enactive Torch demonstrated stable “self-similarity,” or indices of healthy and adaptive single systems, regardless of aperture width, trial order, features of the participants’ judgments, and partici- pant characteristics. Enactive Torch trials exhibited a somewhat greater range of dynamic fluctuations than the rod trials, as well as less movement recurrence, suggesting that the Enactive Torch allowed for more exploratory movements. Findings provide support for the notion that person-plus-tool systems can be classified as extended cognitive systems and a framework for quantifying system-level proper- ties of these systems. Implications concerning future research on extended cognition are discussed.

Keith Harris: Extended Cognition and Cognitive Integration
The hypothesis of extended cognition (HEC) alleges that human cognitive processing is sometimes partially realized in features of the agent’s environment. Epistemologists have recently begun to consider the possibility that human agents might extend their cognitive abilities by use of external tools. This possibility is particularly significant for virtue reliabilists, who maintain that the attainment of knowledge requires the exercise of one’s cognitive abilities. HEC thus seems to allow for the development of virtue reliabilism, and to raise the prospect of extended knowledge. Moreover, it is sometimes argued that the marriage of HEC and virtue reliabilism is necessary for virtue reliabilism to account for knowledge attained through use of instruments.

HEC also raises a puzzle for virtue reliabilism. When does a part of the world partially realize one’s cognitive abilities? If no plausible and HEC-friendly answer can be given, this puzzle can become a problem for those who wish to marry virtue reliabilism with HEC. Philosophers have offered various answers to this question but, as I argue, parts of the world external to the human organism never partially realize the cognitive abilities of human individuals. I then argue that virtue reliabilism can be developed not through HEC, but through the related thesis that cognitive processes, not attributable to human individuals, may be realized in systems of human organisms and external props. Thus, there may be knowledge that is produced by cognitive systems comprising human individuals and features of their environments but is not attributable to human individuals.

Mariel Goddu & Beate Krickel: Cognitive Ontology through the Lens of Biology: How an Evo-Devo Approach shows that Cognitive Capacities are Extended

In recent years, thinking about how to carve up the mind into different cognitive capacities has become popular in philosophy and cognitive neuroscience under the heading of “cognitive ontology”. In this chapter, we will connect the cognitive ontology literature with a recent proposal to think about cognition in terms of cognitive characters (Figdor 2022). We will argue that this connection will show that the criteria for individuating cognitive capacities are essentially extended, i.e., include factors outside the organism’s brain and body.

The character concept is an important analytic tool taken from evolutionary-developmental biology. Characters are biological properties that constitute relevant units of evolutionary selection and heritability: they are biological properties that could in principle be shared by various species and have a certain kind of stability that makes them traceable through the phylogenetic tree. The stability and traceability of characters is explained by the existence of specific so-called character identity mechanisms (ChIMs) (Di Frisco, Wagner, Love 2021). Character identity mechanisms are units in developmental processes that are necessary and non-redundant causes of characters. In other words: the causal topology of developmental processes that give rise to characters has a bow-tie structure (Ross, 2021), where the knot of the bow tie is the ChIM. Characters, thus, are determinables of character states (the determinants) that result from mechanisms that have a bow-tie topology. 

We will apply this idea to identify cognitive capacities. We will argue that an important class of cognitive ChIMs are feedback loop learning mechanisms that include elements of the environment as crucial components. Thus, the identity of some cognitive characters will depend on environmental factors––and thus, some cognitive characters are extended. We will show how this conclusion differs from similar views such as passive externalism, active externalism, as well as Khalidi’s (2022) recent account of extrinsic cognitive capacities.

Holger Lyre: Active Content Externalism and Socially Extended Cognition
My talk has three goals: First, I argue that TXM, commonly understood as a form of vehicle externalism, is accompanied by a new and interesting form of content externalism that can aptly be called "active content externalism." Second, I show that the mechanisms of shared intentionality can be viewed as coupling mechanisms of cognitive extension into the social domain, and third, that social externalism, contrary to what was originally thought, is actually a version of active content externalism.

Albert Newen: The Situated Self: Situatedness as enabled by Integrated Patterns of Characteristic Features
How can we best characterize the self? It is now common sense, that we have to distinguish a plurality of aspects of the self (Neisser 1988, Bermudez 1998, Newen & Vogeley 2003, Musholt 2015). This comes with the still debated question 1. how many selves there are in a healthy person? and 2. how can the different selves or (respectively) aspects of the self be integrated such that we can understand our everyday unitary self-experience? This also invites for 3. Can features beyond the brain-body unit (or the an agent) be part of the self, i.e. can the self involve extrabodily features as constitutive features of the self?  Concerning 1, I defend the position that we have to presuppose one embodied self (contra Metzinger’s no self position and against theories of multiple  selves in healthy persons). Concerning 2, I propose the pattern theory of self delivers an adequate answer: the self is an integrated pattern of characteristic features. This framework enables us also to deal with the question 3: Intuitively, some extrabodily features (or entities) can be part of the self, namely one’s partner, one’s parents, and sometimes also some favorite objects (house, car, book). Loosing them often is reported to feel like loosing a part of oneself. How can we adequately describe this feeling with a naturalistic theory of the self? The aim is to deliver answers to the three central questions about the situated self.

Robert Rupert: Distributed Cognition and Self-knowledge
The compelling presentation of distributed perspectives on the mind – as one encounters in, for example, Dennett’s Consciousness Explained and Clark’s Being There – contributed significantly to the rise of situated approaches to cognition, partly by damping any felt need to find a single location (inside the skull, presumably) where the mind resides and where genuine cognition takes place. Distributed perspectives on the mind also gave rise to difficult questions about self-knowledge. If the mind were local, taking the form of, say, a single stream of conscious states, a straightforward account of self-knowledge appears ready to hand: a state of self-knowledge obtains when a conscious state accurately represents (possibly self-referentially) an event in the stream of consciousness of which that state is a part. But, if mind and cognition are distributed mongrels, we find ourselves in need of a different account of self-knowledge. In this talk, I first articulate a distributed account of cognitive systems, understanding a cognitive system to be the thing that stands a chance of being the cognitive self. Then, I provide a framework for self-knowledge as it might appear in this sort of radically distributed cognitive self (regardless of whether that self extends beyond the organism, a matter on which my characterization of cognitive systems is neutral).

Lawrence Shapiro: Ruminations on Rubber Hands
The rubber hand illusion was designed to provide insights into the phenomenological experience of body-ownership. But it also offers an opportunity to reflect more generally on topics of philosophical and psychological interest, such as the justification for top-down processing models, conceptions of representation, and cognitive penetration. I will argue that conclusions psychologists have drawn about these issues do not follow from the data they cite and that the rubber hand illusion may in fact be two distinct illusions. These lessons are then applied to full body illusions.

Shannon Spaulding: Extending Social Cognition
In their seminal paper “The Extended Mind,” Clark and Chalmers hint at the possibility that extended minds could include other people. To the possibility of socially extended cognition they say, “We see no reason why not.” Though Clark and Chalmers’ paper was published 25 years ago, it is only recently that scholars have attempted to flesh out the possibility of socially extended cognition. Shared intentionality, gestures, bodily expressions, and mere social interaction have been proposed as candidates for socially extended cognition. In this paper, I will review these extensions of social cognition and propose one more: socially extended biases, modeled on Keith Payne’s bias of the crowds.

David Spurrett: The Extended Cuckoo
Arguments that cognition or minds can be extended regularly invoke an analogy with Dawkins’ argument that phenotypes can be extended. I argue here that there are two neglected ways in which that pair of boundary-breaking theses are complementary. A large part of the argument of The Extended Phenotype concerns phenotypes expressed in the behaviour of other organisms. But the options Dawkins considers for this extended manipulation are cognitively internalist. If we view cognition as extended we can recognise a wider range of vulnerabilities for exploitation. On the other hand the analogies drawn with Dawkins almost always emphasise the benefit to the individual agent in being cognitively extended. Taking Dawkins’ concerns about manipulation and exploitation more seriously leads to a more contested, less optimistic picture of extended minds. This second line of thinking follows Sterelny’s lead (2003) but I argue that hostility presents worse and more pervasive problems that he allows.

Karina Vold: Cognitive Extension and the Human-Centered Generality of Multi-Modal Models
Recent advancements in artificial intelligence (AI) research have led to the emergence of massive prompt-commanded multimodal models (M*s) that are increasingly taking on the role of cognitive extenders (Schellaert et al. 2023). According to the extended cognition thesis, cognitive extenders are the tools we regularly recruit to aid our completion of certain cognitive tasks. When these tools make important functional contributions, they can become as essential to our cognition as our brains, such that the tool should be seen as a literal extension of our cognition: a cognitive extender (Clark and Chalmers 1998). In this paper, I will argue that when M*s are used as cognitive extenders they present unique opportunities and challenges for human users due to some of their distinctive features. I will begin by explaining how M*s function and how they are being used as cognitive extenders. I will then consider some of the opportunities that M*s present. For example, I argue that unlike most cognitive tools in the past, because M*s have reached a new level of generality, they are proving to be flexible cognitive tools offering support for many human cognitive capacities. Finally, I turn to some of the unique challenges that M*s qua cognitive extenders present for human users, developers, and other relevant stakeholders. For example, we know M*s have some obvious deficiencies: they produce generated content that is not truth-tracking, their generated content reflects and exaggerates existing biases in training datasets, they are not autonomous nor are they value-aligned, etc. Each of these shortcomings present distinct challenges for human users, which should be considered in our responsible use and design of these human-centered M* systems.

Carmen Mossner & Sven Walter: Shaping (Extending) Social Media Minds: Scaffolding Empathy in Digitally Mediated Interactions?
Empathy is an integral aspect of human existence. Without at least a basic ability to access others’ affective life, social interactions would be well-nigh impossible. Yet, several studies seem to show that the means we have acquired to access others’ emotional life no longer function well in what has become our everyday business – social interactions in online spaces. If this is correct, there are two important questions: (1) What makes empathy for those with a debauched online life so difficult? and (2) What can we do to alleviate the negative consequences? We first identify structural differences between offline and online interactions that can explain why digital empathy is harder to achieve, and then consider the idea of combating the resulting difficulties by changing online spaces in ways specifically designed to ‘scaffold’ or ‘extend’ empathy where our evolved mechanisms fail.
Section 2 argues that empathy is essentially a matter of interpreting the behavior of embodied subjects. Section 3 identifies three factors that are a crucial part of the social foundation of this interpretative endeavor: the empathizer’s affective repertoire, their perceptual input, and their background knowledge. Section 4 argues that digital interactions differ from face-to-face interactions with regard to these factors in ways which render our evolved empathy mechanisms less effective in the digital world. Section 5 toys with the idea that situational factors can serve as ‘empathic scaffolds,’ i.e., as ‘tools’ that can ‘shape’ or ‘extend’ people’s empathic reactions. Section 6 wraps up the main line of reasoning, responds to objections and invites further scholarship.