Lambros Malafouris is Professor of Cognitive and Anthropological Archaeology at the Institute of Archaeology and Tutorial Fellow at Hertford College, University of Oxford. He is the author of How Things Shape the Mind (MIT press, 2013) and An Anthropological Guide to the Art and Philosophy of Mirror Gazing (Bloomsbury, 2020, with Maria Danae Koukouti). He has edited The Cognitive Life of Things (Cambridge 2010, with Colin Renfrew), The Sapient Mind (2009, Oxford University Press, with Colin Renfrew and Chris Frith), and Material Agency (2008, Springer, with Carl Knappett). Lambros Malafouris is currently PI of HANDMADE funded by the European Research Council (ERC)

Active Inference (AI) has gained significant traction in recent years across the field of cognitive science. It signifies a shift away from the brainbound logic of predictive coding (by which learned predictions are used to infer probable causes of sensory data) and towards a more extended view, responsive to, and embedded in the social and material world (known also as predictive processing PP). For AI the human cognitive system is actively engaging the world (making predictions at the intersection of the brain-body with its environment), rather than passively encoding (making representations inside the brain) the world (Clark 2013, 2016, Gallagher and Allen 2018). Such a shift opens up new possibilities for synergies with Material Engagement Theory (MET) and cognitive archaeology. However, important questions remain: if, as AI rightly suggests, it is not just the Bayesian brain that performs the predictive processing, what exactly are the elements of the cognitive system that participate in that process, and in what ways? What really matters, and what part of the system will take the credit for that processing? Importantly, what is the role played by material things (in the broadest sense of material mediations and material environments) in the process by which our learned modes of perception and ‘prior’ understandings are used to infer what we see and how we experience the world? A meaningful and productive discussion between AI and MET demands that we try to tackle those questions and attain some clarity on the fundamental assumptions, analytical scope, and epistemic objectives of the two frameworks. This would be the main task of this paper. My objective is twofold: to identify the main problems and stumbling blocks that, in my view, limit the applicability of PP from a MET-perspective (and vice versa), and to suggest possible ways to overcome them towards a theory of Predictive Material Engagement.

I start with a brief overview of MET. I explain the main assumptions of that theory for a continuity of brains, bodies and things, and the vision of thinking as thinging (thinking and feeling with and through things) that it embodies. I also present Perspectival Kinaesthetic Imaging (PKI) which is a methodology developed in the context of the ‘HANDMADE - Understanding Creative Gesture in Pottery Making’ project (Malafouris et al., 2023) designed to facilitate the multimodal sensitivity needed for studying the sentient dimensions, conditions and possibilities of creative material engagement. Drawing on research and using examples from my ongoing anthropological work with potters and ceramists I will attempt to show how a focus on the ‘middle’ lived space of material engagement can help us disambiguate the often discussed, but little understood, creative dialogue between maker and material. I will use this example of the attentive and dialogic character of creative thinging (feeling of and for clay) as my main point of intersection with PP and AI.

I am not going to evaluate AI as a general theory. My discussion relates specifically to the potential contribution of this model for explaining material engagement phenomena, traditionally studied in archaeology and anthropology, that manifest at the intersection between cognition and material culture. I identify three areas of potential friction between MET and AI. These are: (a) the lack of clarity (also disagreement) among the major varieties and proponents of AI on some of the fundamental assumptions behind the ontology, mereology, and topology of the model, (b) the overreliance on the mathematical formalism (known also as partially observable Markov decision process POMDP) (Smith et al., 2022; Constant et al., 2021) which demands from the phenomena to be explained total submission to the algorithmic logic of the generative model, last (c) the overambitious, and in my view unhelpful explanatory scope of AI as a unifying and complete account of cognition. That is, an account capable of explaining all aspects of cognition and of unifying explanations across life sciences and cognitive sciences under a single mathematical formalism.

To explain my reasoning around those issues I present two fundamental challenges to the use of AI models in MET and cognitive archaeology: (a) the challenge of misplaced concreteness, and (b) the inference-grounding problem.

The challenge of misplaced concreteness (I borrow the phrase from A.N. Whitehead) denotes the confusion of actual and ideal that is a common problem with any kind of in silico (computational) modelling and simulation. Specifically, in the case of AI and PP, the challenge is how to avoid the danger of confusing statistical abstractions for the concrete situations or processes that they code. Imagine a potter centring a lump of clay on the wheel. Whatever model you use to simulate the interaction between hand and clay it can only ever give you a partial and incomplete abstraction or representation of the real creative processes enacted when the potter actually discovers and exploits the affordances of clay and the local environment (e.g. the wheel). In itself, this ontological deficit of the generative model is not necessarily a problem. But that depends on two conditions: first, that the limits of the model (what it can do and what it cannot do) are made explicit, and second, that we provide some clear exclusion and inclusion criteria about what we choose to represent (what are the variables that we choose to encode and what are the variables that we leave out). We cannot simply assume, or worse pretend, that any set of parameters chosen for use in the generative model (e.g. related to the potter’s attention, level of precision or skill) somehow ‘correspond’, or adequately and accurately reflect, the concrete situations they are used to encode or simulate. The challenge here for MET and AI is how best to combine forces and perspectives in order to provide a deeper understanding of the material conditions that ‘matter’. This mutual understanding will allow a critical evaluation of the chosen parameters, and by extension, of the actual relation between the generative model and the cognitive process or behaviour under study.

This brings us to the second and related challenge, which I call the inference-grounding problem. This is a version of the classical symbol-grounding problem. It relates to meaning-making and the semiotics of matter, processes with which MET is centrally concerned. More specifically to AI, the inference-grounding problem relates to how generative models get their meanings, which relates to the question raised before about how are those models connected to the real things and processes they refer to (for instance, to the potter’s bodily skills, creative gestures and modes of material imagination)? The problem can be summarised as an inherent conflict between the ontology of AI, as an anticipatory process of engagement with the world that presupposes a body (embodiment) meaningfully situated within an environment, and the ontology of the mathematical formalism that determines the interactive potential (affordances), and which, at least for its virtual or neurofunctional implementations, seems to be immune, or in any case is not affected by the material conditions of situatedness and embodiment. One may argue here, that from a computational point of view, AI is part of the ‘software’ and thus ‘implementation-independent’, meaning, it will do the same job no matter what hardware it is executed on. But such a view, would be in conflict with MET’s commitment to enactive signification and the continuity between cognition and material culture. I suggest, we must be sceptical of any simplistic functionalist version of multiple realisability that undermines the sociomaterial bases of cognition by treating cognitive processes along the software/hardware analogy (where the mind is to the brain as a computer program is to the hardware of the computer on which it runs) so that they can be abstracted out and, in principle, implemented in a variety of systems. That would mean that the details of organismic embodiment and materiality (in the broadest sense of the relevant material environment) does not really matter for cognition. Which effectively cancels the argument for material agency. The challenge we face then, relevant to the inference-grounding problem, is one of reconciling the following opposition: the inherent dependency of AI on the ‘mathematical formalism’ which renders it such a powerful model for explaining subpersonal states at the micro level, is also what reduces and limits its explanatory power in the case of suprapersonal states at the macro-level of human situated action and material engagement.

How can we meet these challenges? In the last part of the paper, I will attempt to map out a way forward for realising the cross-disciplinary scope and potential synergies between AI and MET. My main suggestion will be that for AI and PP models to succeed in unifying representationalist and enactivist (non-representational) visions of mind (Constant, et al., 2021) and in building synergies with MET, it is important to devise ways of connecting the generative model with the local details (variable and invariable structures) of material engagement. Before we come up with effective ways of doing that (and mere encoding or in silico simulation of those details is not the solution) the usefulness and applicability of AI models for understanding human engagement with material world remains uncertain.

It is important to clarify what exactly I mean here. I do not mean developing algorithms and generative models capable of coding the totality of the potter’s ‘creative thinging’ or the relevant material environment. Think again of the skilled creative interaction between potter and clay. According to PP and AI models the potter’s actions can be accounted for by using a Bayesian process of prediction error minimization. On this construal, the potter engages with clay on the basis of prior predictions that are continually updated against incoming information. Very often, the incoming information regarding the actual state or form of clay will not exactly match the potter’s prior predictions. The perceived mismatch generates a ‘prediction error’ and ‘surprise’ that in turn will generate a new prediction recalibrated for accuracy. This loop of ‘prediction error’ and ‘surprise’ leading to ‘recalibration’ of ‘priors’ will continue throughout the creative process.

Such a PP account of the making process is general enough to be applicable in all kind of scenarios, spanning different individuals, pottery techniques and materials: the potter’s actions can be accounted for by using the same Bayesian process of prediction error minimization. However, such a general PP account also threatens to misrepresent the synchronic and diachronic variability, and inherent unpredictability of creative processes as they become enacted in real practice and in different situations. It should be noted that from the point of view of creative practices, the uncertainty that characterise pottery making is not experienced necessarily as a problem (one that PP will minimise or remove) but also as a major resource for improvisation (Malafouris 2023). For many potters improvisation is a way to tame surprise and deal with uncertainty or accidents. This means that for certain aspects of the creative process the principle of error minimisation does not apply. Accidents and mistakes are welcome. This potential conflict between prediction and improvisation (see also Gallagher 2023) highlights also my previous point regarding the relation between the computational (sub-personal) level and the phenomenological (personal) level of description and analysis.

Understanding the local details of action and its material environment that constitute the potter’s creative thinging often necessitate years of anthropological participant observation and/or in depth phenomenological and multimodal analysis. Reducing that detail information to some mathematical formalism is a non-starter from the point of view of MET. Quite the contrary, we should be focusing our efforts on devising methodologies and argumentative strategies that actively allows us to incorporate, juxtapose and combine perspectives and descriptions (phenomenological, anthropological or other) that complement, but are not subsumed to each other, or reduced to the generative model. What we need, and what I am proposing, is not finding ways to encode more of the ‘dark matter’ of the cognitive universe so it can processed by the generative model, but instead, finding ways that complementary types of information can be combined avoiding any substitution or reduction. The guiding epistemic objective here is not one of replication or falsification but instead, of correspondence (between models and actual phenomena) and ecological validity (understanding the ecology of predictive processing).

In short, I propose that MET and AI need to come up with a bypassing epistemic strategy that allows coexistence of the two theories without losing or heavily compromising their distinctive descriptive and analytical power. The logic of that strategy should not be reductive or representational but participatory and “diffractive”, in Barad’s sense of “marking differences from within and as part of an entangled state” (Barad, 2007, 89). The aim is to highlight both the complexity of our hypotheses and the complexity of the phenomena and the empirical observations against which they will be weighed. What this essentially implies is that instead of putting mind’s complexity in the service of algorithm we must put the algorithm into the service of mind. Instead of seeking more efficient ways to model thinging we should be seeking ways of re-thinging AI. Re-thinging AI essentially means situating and temporalizing it. Which also means, returning to the old unifying objective of putting brain, body, and world together again, raised by Clark three decades ago.

I understand that for most researchers subscribing to computationalism, the basic story of the brain as the “generator of fantasies, hypotheses and predictions that are tested against sensory evidence” (2023, 257), to borrow the words of Carl Friston, is powerful enough and needs no amendment. But this basic predictive story, is incompatible with the basic enactive principles as well as the kind of phenomena that MET is used to describe and explain. AI models, advocating an extended or enactive version of the predictive story, also claim to depart from this neurocentric model. However, it remains to be seen, whether AI models move past the seductive power of internalism. In Surfing Uncertainty Clark’s message is not entirely clear: on the one hand the brain is presented as “our driving force in the daily surf through the waves of sensory stimulation”. On the other, Clark is careful to point out that the brain is not “an insulated inference engine so much as an action oriented engagement machine” (2016, 1). What does this emphasis of action mean really? Probably it means different things to different people. My personal reading, which is the reading that stems from a material engagement perspective is that human thinking (past, present and future) needs more than a brain. You need a skilled body (the body of the surfer). You need the wave whose diffractive logic will produce the uncertainty. You also need, an element that matters the most so far as the links between MET and AI is concerned, and which has been traditionally neglected i.e., the surfing board. The surfing board is not merely an object that can be weighed and measured but an active process with its own peculiar temporality, affordances and dynamics. It is the creation and use of the surfing board that provides the material foundation for the very possibility of surfing activity and the modes of predictive engagement associated with it. What kind of ‘prior’ is that? Can it be coded and by whom? In what way? These are critical questions. No surfing board no surfing skill. No skill no surfing of uncertainty. As for uncertainty, is the only certainty of human life. But how exactly we deal with uncertainty is an open question.