1. Introduction
“To put it bluntly, AI is a marketing term. It doesn’t refer to a coherent set of technologies. Instead, the phrase ‘artificial intelligence’ is deployed when people building or selling a particular set of technologies will profit from getting others to believe that their technology is similar to humans, able to do things that, in fact, intrinsically require human judgement, perception or creativity.”
Emily Bender and Alex Hanna, The AI Con
“Technological machines of information and communication operate at the heart of human subjectivity, not only within its memory and intelligence, but within its sensibility, affects and unconscious fantasms.”
Felix Guattari, Chaosmosis
“This is wild interdisciplinarity!” – voilà the gut reaction by a first reviewer to Performing AI (PAI): Governance, Agency and Action, the four-year project of interdisciplinary inquiry for which we had just obtained a substantial grant from the Swiss National Science Foundation (SNSF). As its title indicates, the project is due to investigate how AI “performs” in situ, rather than taking for granted its agential powers, and that in (and through) a multiplicity of registers. Therefore, the project leverages an interdisciplinary approach between – and from within – the social, natural, human sciences and the arts, an approach that shall bring together researchers from science and technology studies (STS), digital sociology, ethnomethodology, performance studies, digital arts and complex systems/artificial life. The reviewer’s reaction also expressed skepticism. What might be the distinctive relevance of a “wildly interdisciplinary” project on AI? Isn’t the multiplicity of situated performances at odds with a common rationale for investigating them? And, last but not least, doesn’t the interdisciplinary inquiry into an interdisciplinary domain – “AI” – presage bland redundancy, rather than nuanced insight? Working out this joint contribution to the first Cambridge Forum on AI afforded us an apt opportunity to address the reviewer’s skepticism, inviting us not only to revisit the project rationale for PAI but also, and more ambitiously, to make explicit the paradigmatic interest of interdisciplinary inquiry into the situated performances of AI more broadly. Therefore, we shall draw upon famous historian of science Thomas Kuhn’s (Reference Kuhn1970) notion of paradigm, the notion according to which scientific communities share intricate sets of “beliefs, values, techniques” as a matter of disciplinary working routines (p. 175). More specifically, we shall articulate two contrasting readings of paradigm: first, as highlighting conceptually “incommensurable worldviews” and, second, as prioritizing the routine work of “disciplinary practices” (Rouse Reference Rouse1987).Footnote 1
“Artificial intelligence,” “interdisciplinarity,” “performance” – all of these terms, used separately or jointly, gloss an intricate array of discourses, materialities and practices. Each term has come to gloss expanding research fields and numerous literature reviews, as well as lingering controversies in and reflexive engagements across fields (Herzig Reference Herzig2004; Salter Reference Salter2020; Sormani Reference Sormani, Borgdorff, Peters and Pinch2019), not to mention public and political attention beyond academia (Bareis and Katzenbach Reference Bareis and Katzenbach2021). Against this backdrop, readers of this position paper might reasonably wonder “what difference” an interdisciplinary inquiry into situated performances of AI makes today, similarly to our skeptical reviewer of the PAI project. As indicated, the paper will explicate the paradigmatic interest of interdisciplinary inquiry into “performing AI” in two related ways:
• First, the paper articulates (some of) the foundational conceptual and methodological premises for what a research program on situated AI must address – that is, as part of a (potentially) “incommensurable worldview” with respect to received views of AI (e.g., technological determinism) or unilaterally disciplinary perspectives (e.g., in terms of computer science).
• Second, the paper explicates the practical instantiation of the outlined program by reflecting on the PAI project so far, its common rationale and contrasting research strands – that is, as an ensemble of “research practices” of both multi- and transdisciplinary orientation, balancing division of labor within disciplinary boundaries and moments of transgression (e.g., via public engagement).
The bulk of the paper describes and reflects on the PAI project as the contingent operationalization of interdisciplinary inquiry into “AI in action” and, therefore, explicates its foundational premises as they take shape. After all, “case studies are the bread and butter of STS” (Sismondo Reference Sismondo2010, 8). Among the foundational premises of the project, and the paradigmatic approach to “performing AI” advocated here, is the core concept of sociotechnical assemblage, on the one hand, and reflexive interdisciplinarity as a foundational methodology, on the other. Let us start with a working definition of each concept and methodology, before outlining the structure of the position paper.
Working Definition #1: “Sociotechnical Assemblage” – Ever since the publication and reedition of Lucy Suchman’s Plans and Situated Action (Reference Suchman1987; Reference Suchman2007), the notion of “situated action” has gained traction. This can be seen in the continued engagement with “AI in situ” across STS (e.g., Sormani Reference Sormani2023; Suchman Reference Suchman2023), cognitive science (Varela et al. Reference Varela, Thompson and Rosch1991) and cognate domains.Footnote 2 It can also be seen in sociological debates around the relevant “definition of the situation” in the context of societal digitization (e.g., Marres and Sormani Reference Marres and Sormani2023). This position paper, in turn, highlights “sociotechnical assemblage,” rather than situated practice(s) as a core concept. Why? To probe “AI in situ,” the notion of sociotechnical assemblage – alluding to the ongoing work of articulating “social” and “technical,” “discursive” and “material” dimensions into a single operating device, infrastructure or apparatus – is crucial for two reasons. Historically, the label “AI” has never designated a pure technology, somehow emancipated from its rhetorical mediation (contrary to what technological determinism may suggest). Quite the opposite, the term “artificial intelligence” was introduced as the equivalent of a marketing slogan to support the very formation of early computer science (McCarthy et al. Reference McCarthy, Minsky, Rochester and Shannon1955) and in manifest opposition to cybernetics (Kline Reference Kline2015). Ever since, the slogan has remained in search of the information technology best suited to deliver on its ambitious promise(s) (Wooldridge Reference Wooldridge2020). In this sense, AI has remained a project of “heterogeneous engineering” (Law Reference Law, Bijker, Hughes and Pinch1987). Conceptually, then, the notion of sociotechnical assemblage motivates a dedicated focus – political, empirical, aesthetic – on the tricky articulation of discursive, material and “social” (and other) dimensions for “artificial intelligence” to perform in situ – coherently or less coherently so, as amply suggested by “stochastic parrots” (Bender et al. Reference Bender, Gebru, McMillan-Major and Shmitchell2021) and “demonic contingencies” (Ivarsson Reference Ivarsson, Sormani and Vom Lehn2023).Footnote 3
Working Definition #2: “Reflexive Interdisciplinarity” – As there is no lack of literature on “interdisciplinarity” and its cousin concepts (e.g., “transdisciplinarity”), spanning policy reports and academic treatises (e.g., Gibbons and Nowotny Reference Gibbons and Nowotny2001), there is no lack of definitions of the idea(l)s that are expressed by these terms, including “particular configuration[s] of programmatic statements, interventions and practices” (Barry and Born Reference Barry and Born2013, 5, paraphrasing Foucault Reference Foucault1972). Idea(l)s of interdisciplinarity have thus nurtured controversy (e.g., Klein Reference Klein2004; Nowotny Reference Nowotny2007; Wallerstein Reference Wallerstein1996). So have concepts of multi- and transdisciplinarity (e.g., Osborne Reference Osborne2015), opposing epistemological pleas for “(multi-)disciplinary autonomy” (Strathern Reference Strathern and Strathern2004) and subversive promises of “joint problem solving [via] transgressive acts against existing knowledge boundaries” (Gibbons and Nowotny Reference Gibbons and Nowotny2001, 68). This position paper draws upon the referenced literature, albeit selectively and reflexively. For the interdisciplinary investigation of “AI in situ,” aka its sociotechnical assemblage ad hoc, a reflexive approach – at least in methodological terms – indeed turns out to be foundational, if only to avoid assuming the “uncontroversial ‘thingness’ of AI” (Suchman Reference Suchman2023). Our working definition, as a basic requirement for engaging interdisciplinary inquiry into “performing AI,” entails a triple reflexivity:
• First, any research encounter with “AI,” as “performing” somehow, should make explicit the particular working discipline (deemed “pure” or otherwise!) in terms of which that encounter was to be had (e.g., via computer science, interaction analysis or historical inquiry as a primary discipline).
• Second, and insofar as the core phenomenon is a sociotechnical assemblage (an operating mix of discursive, computational, material and other dimensions), research encounters with “AI in situ” should make explicit the transdisciplinary contingencies they harbour – that is, the situationally emerging affordances for a “hybrid heuristics” (Rheinberger Reference Rheinberger, Sormani, Carbone and Gisler2019) beyond traditional disciplinary boundaries, including novel epistemic, aesthetic and/or political qualities of “performing AI” in situ (e.g., via “improvising machines,” Bowers Reference Bowers2002).
• Third, interdisciplinary investigation of “AI in situ” should make explicit the multifaceted relationships between – and from within – these two contrasting moments of inquiry, multi- and transdisciplinary – in other words, the “agonistic dialogue” (Barry and Born Reference Barry and Born2013) between the routine multiplicity of disciplinary practices and singular transdisciplinary contingencies.
Drawing upon “interdisciplinarity [as] a generic expression” (Barry and Born Reference Barry and Born2013, 9), the above definition of interdisciplinary inquiry casts it as a reflexive endeavor, inviting researchers to make explicit both their multi- and transdisciplinary involvement with “performing AI.” Importantly, the tension between these two modes of involvement is not to be reduced to one of its poles, “multi-” or “trans-,” but to become part of the inquiry itself, including the differing positions, if not incommensurable framings that it may generate – hence our reference to “agonistic dialogue.”
But how is a reflexively interdisciplinary research program on “performing AI” as a multifaceted sociotechnical assemblage – that is, in line with the outlined premises – to be delivered on, practically instantiated, cogently articulated? In short, how is the working definition to be worked out? At the time of writing, the lead authors of this position paper – Sormani, Ikegami, Saunier, Jobin, Glassey and Salter – are not aware of many – if any – contemporary research programs on situated performances of AI that engage “wild interdisciplinarity” between – and from within – the social, natural, human sciences and the arts. Therefore, the remainder of this position paper will revisit the above-mentioned PAI project, its project rationale and internal organization for how it makes – and might still make – its interdisciplinary case, as a practical instantiation of the outlined premises so far. To make explicit the paradigmatic interest of the PAI project, we will offer a narrative account of its practical realization, while elaborating on its (provisional) “learning lessons” with respect to sociotechnical assemblage and reflexive interdisciplinarity. The paper has been structured accordingly.Footnote 4
First, the paper traces some defining moments of conceptual clarification, as they contributed to articulating “sociotechnical assemblage” as a core concept for the PAI project. In particular, we shall unpack the twinned concept of performance/performativity as a “boundary object” (Star and Griesemer Reference Star and Griesemer1989), an epistemic object which both enables and invites interdisciplinary inquiry into the paradoxical, if not contradictory character of “AI in action.” Second, our narrative account of the PAI project and its practical realization will elaborate on “reflexive interdisciplinarity” as a critical frame and purposive methodology – that is, as a frame allowing us to subvert, if not avoid machine reification, while providing a methodology suited for probing “AI in situ,” as a distinctive performance/performativity nexus. The latter probe, in turn, may contribute to renewed “public experimentation.” Third, our account revisits the multiple (disciplinary) commitments involved in the reflexive explication of “performing AI” as a sociotechnical assemblage, making explicit their respective contributions (i.e., on policy discourse, material agencies and situated action), including potential anomalies and/as transdisciplinary openings. Fourth, we will offer some initial reflections, on the basis of these first working steps of the PAI project, on the position(s) of the arts in relationship to the sciences, a configuration which permeates the project, and which we referred to in terms of its transdisciplinary contingencies above. Drawing together the key aspects of the paradigmatic interest of the PAI project, our conclusion highlights its ambivalent status – partly determined, partly speculative – as a heuristic resource for interdisciplinary inquiry, if not renewed artistic and public engagement with “AI” in the making. Throughout the paper, the transversal question that we want the reader to keep in mind is: What added value do so many different disciplinary perspectives, methods and practices provide to the social-technological-aesthetic study of AI as an arguably more-than-human entity that is locally performed and discursively articulated, while producing real sociotechnical consequences and material effects?
2. Performance/performativity as a boundary object
PAI: Governance, Agency and Action is still in its early stages. One year into the project, the different teams are trying to get their bearings around not only the literatures, procedures and practices from their own fields (e.g., artificial life) but also “contaminating” concepts from the other disciplines in the project – in short, the project is well underway. Arriving at the end of the “propaedeutic” year, as a senior member from one team put it, allows us to trace how its paradigmatic interest has taken shape so far, on what basis, and where it still points to. To begin with, let us consider two defining moments that have led up to “sociotechnical assemblage” as a core concept for PAI, if not beyond.
Defining Moment #1: “Performative AI” – Held a few weeks into the project start, the kick-off meeting of PAI afforded us all – principal investigators, senior and junior researchers, project collaborators – with a first opportunity to ponder the project rationale, its conceptual organization and practical implications, in light of the tentatively submitted (and by then generously funded) grant proposal. In the proposal, “AI” was pitched as a “monster” (Haraway Reference Haraway, Grossberg, Nelson and Treichler2013, 20) and “hybrid” (Latour Reference Latour2012), a “bizarre mix of nature and culture, technology and society,” thus “never pure” (Shapin Reference Shapin2010). This had the triple advantage of posing a big mystery (what could this “thing” possibly be?), together with big names (in and beyond STS), and thereby justifying our funding request for (relatively) big money! But, pending the actual inquiry, had we really advanced in conceptual clarity? At the kick-off meeting, one of the PIs – renowned artificial life researcher Takashi Ikegami – seriously doubted it. Instead of getting distracted by “performing AI,” we should focus on “performative AI.” “Performative AI,” rather than Performing AI, as the project title had it, others wondered. So what, if any, might be the difference?Footnote 5
To begin with, as Ikegami argued, “performative AI is fundamentally different from generative AI,” the technology whose media presence and everyday use have conquered the globe (Bender and Hanna Reference Bender and Hanna2025; Elliott Reference Elliott2019). “Performative AI” focuses on “AI as an actant embedded in the real physical world and engaged in multiple feedback loops within specific complex environments.” This, we should highlight, contrasts with the flat impressions of generative AI systems, impressions that we might get, if we – as smartphone users, for example – restrict ourselves to interactions with chatbots, through prompts and responses rendered onscreen. Indeed, and even though this form of interaction has nurtured large language models (LLMs), it might still be dismissed as a form of “glorified autocomplete” (Rothman Reference Rothman2023). So what? As part of ongoing conceptual clarification, the notion of “performative AI” confronted us – as meeting participants, as well as research colleagues – with a first defining moment of ambivalent interest.
On the one hand, the notion of “performative AI” – if we expand on Ikegami’s line of argument – suggests new, potentially unthought-of kinds of material actions and expressions, at least from selected scientific (e.g., artificial life) and computational arts perspectives, emerging from an interplay between human agents and machine systems. Here, sensing, actuation and sense making – the idea that meaning making emerges through concrete and embodied sensing and sense making with the world (see De Jaegher and Di Paolo Reference De Jaegher and Di Paolo2007) – is key to such forms of interaction, not to mention the very constitution of such data, systems and their situated (co-)enactment.
Moreover, the notion of “performative AI,” if dwelled on, invites an intriguing ontological shift: performative AI should not be understood as an object that performs (i.e., takes a temporally bound action in the world) by itself (an agentic assumption which the PAI project challenges from the outset), but rather as a mode of becoming – a continuous articulation of presence through embodied relations – in situ and in vivo. Unlike generative AI, which operates within predefined symbolic and statistical structures, performative AI unfolds – and arguably can be made to unfold – in the interstices between sensing and acting, and between affect and anticipation. Its existence, then, appears as neither reducible to function nor to representation. Rather, it oddly resonates with what Merleau-Ponty (Reference Merleau-Ponty1968), reflecting on human experience and its mediated character, described as the “flesh of the world” – a reciprocal entanglement with the physical, temporal and affective textures of reality. To perform, in this sense, is not to execute a task, but to negotiate uncertainty, live through failure and improvise continuity. Under this description, performative AI does not simulate “life” but instead expresses a nascent form of it.Footnote 6
On the other hand, this highly suggestive mode of (quasi-)ontological thinking, rather unsurprisingly, invites a skeptical note (as our interstitial qualifications in the preceding paragraph already suggest). For sure, the outlined reorientation invites us – A-Life researchers, creative engineers – to treat “AI” not as a solution, but as a question: What forms of life – artificial, hybrid, or otherwise – become recognizable, if not possible, when we shift from design-for-control to design-for-experience? How does “performative AI” emerge as part of experience design? And what genealogies of “machine and ecology” (Guattari Reference Guattari, Bains and Pefanis1995; Hui Reference Hui2020) are folded into the process? Yet these questions, on second thought, turn out to be questionable themselves. Conceptually, the questions are overdetermined, at least in a research context marked by concrete multiplicity as much as projected unity. Why indeed posit an all-encompassing concept of “life” (or “A-Life,” for that matter) as an omnirelevant category? And why this rather than that omnirelevant category – for instance, “society,” aka the “public organization of the vital process” (Arendt Reference Arendt1961, 85)? Luckily, the incidental expression of this conceptual disagreement, if fundamental, not only exemplifies “reflexive interdisciplinarity” at work, but also hints at critical empirical/creative engineering work ahead – not to mention the transdisciplinary contingencies due to “A-Life” argumentation!
Defining Moment #2: Performance/performativity of AI – In the course of writing this position paper, another principal investigator of PAI – longstanding computational artist Chris Salter – suggested, if tacitly so, that we stick with “performing AI” as its interdisciplinary leitmotiv, while integrating the provocative notion of “performative AI.” How? In what sense? Let us pause on the expression “performing AI” for a moment. The expression can be understood in two ways at least, depending on whether “AI” is considered in subject or object position.Footnote 7 When considered in subject position, “AI” can be heard as “performing” (or be “performative”), as “doing something” by and of itself – in other words, it displays performativity, however dubiously or convincingly so. In turn, when considered in object position, “AI” becomes dependent on another instance, perhaps another subject, to “make it perform” (i.e., via an operator, performer, “human” or other) – that is, “AI” now appears as to be dependent on a performance other than itself. The grant proposal of PAI already invited us, when embarking upon the project, to treat this performance/performativity nexus as a “boundary object” (Star and Griesemer Reference Star and Griesemer1989). Why?
First, the twinned notion of performance/performativity offers a promising “boundary object” for interdisciplinary collaboration, both multi- and transdisciplinary – that is, an epistemic thing or concept which, in Star and Griesemer’s felicitous terms, is “both plastic enough to adapt to local needs and the constraints of several parties employing them, yet robust enough to maintain a common identity across sites” (Reference Star and Griesemer1989, 393). In other words, a boundary object is flexible enough to mean different things to different people, yet stable enough so that everyone can coordinate around it. It helps groups working together to avoid confusion and enables them to get things done, despite having different knowledge or expertise – in short, “one can act together without the necessity to be the same” (Sennett Reference Sennett1977, 16). Yet from the moment it was coined in the mid-1950s, the term “AI” has been circulating as “floating signifier” (Suchman Reference Suchman2023) more so than being leveraged as a boundary object; a floating signifier which by now not only refers to a technoscientific research program (McCarthy et al. Reference McCarthy, Minsky, Rochester and Shannon1955) but also is used as a “marketing term” as well as an ontological notion, as hinted at in our epigraph.
Second, and in contrast to vague “AI” in mainstream circulation, the performance/performativity nexus, when leveraged as a boundary object, presents another advantage, namely: to facilitate a focused interdisciplinary engagement with “AI” in situ, despite or precisely because of its multiple and paradoxical, often contradictory and controversial character. The tension signaled by the dueling notions of performance and performativity indeed marks the constitutive tension at the intricate core of “AI” as a sociotechnical assemblage – that is, the tension between the “performativity” of an emerging phenomenon – regarding “AI” as something contingent, temporal, material, dynamic and unstable that challenges received frames of social practice (including the frame of exclusively “human” agency) – and the “performance” of a situated practice – seeing “AI” as something necessary, spatial, otherwise material, rigid and stable, determined in and through discursive articulation and situated action. This tension is perhaps best expressed in Suchman’s apt phrase from Human-Machine Reconfigurations, according to which “lived practice inevitably exceeds the enframing moves of its own procedures of order production” (Suchman Reference Suchman2007, 193).Footnote 8
3. Performing AI, engaging interdisciplinarity – reflexively so!
As temporary expressions of an ambitious research paradigm, the working definitions given in the introduction remain largely ideal-typical. How were these definitions arrived at? How did they practically take shape? What did we learn? And what remains to be done? The previous section allowed us to specify the provisional notion of “sociotechnical assemblage” in terms of a performance/performativity nexus, while reflecting on (some of) its defining moments. That nexus, we suggested, provides a helpful boundary object for engaging with “performing AI” in interdisciplinary collaboration, despite or precisely because of AI’s tricky features. This section, in turn, outlines how and why “reflexive interdisciplinarity” offers a critical frame – allowing us to subvert machine reification, if not to avoid it altogether – and a purposive methodology – suitable for probing “AI in situ,” as a distinctive performance/performativity nexus, if not in view of renewed “public experimentation.” To briefly explicate that frame and this methodology, the lay summary of the PAI project is a good starting point, affording us with a third defining moment of the project.
Defining Moment #3: The Lay Summary – The summary in question reads as follows:
Performing AI’s goal is to contextualize AI as a dynamic social and cultural artifact that is discursively and practically constituted (that is, performed) in specific contexts and situations. In other words, what does “AI” do, why and how does it do what it does, and what effects does it produce across different disciplines? The project takes the theoretical and conceptual lenses of performance and performativity for navigating AI’s messy entanglements between the social and political, technical and aesthetic.
Excerpt 1: Lay summary, PAI project
Published on the project website, the lay summary introduces “AI” as a “dynamic social and cultural artefact,” rather than a sociotechnical assemblage. The specific disciplines to be drawn upon, if any, remain implicit. So does the transdisciplinary outlook, not to mention its contingently hybrid heuristics or potentially contested character. For a lay audience, this might be a more easily understandable introduction to “performing AI” than an intricate STS concept (e.g., “sociomateriality”). Yet this rhetorical concession comes with an opportunity cost. Indeed, the question raised – “what does ‘AI’ do?” – implies that we are somehow confronted with “AI” as an agential power of its own, a (mis-)understanding which ostensibly flies into the face of any reflexive critique of naïve “AI thingness” assumptions (Suchman Reference Suchman2023). Or does it? After all, the lay summary also hints at “performance and performativity for navigating AI’s messy entanglements.”
To find out, let us first consider reflexive interdisciplinarity, PAI’s foundational methodology, as a critical frame versus machine reification, aka the instant projection of fully functional “AI” under clichéd auspices. In leveraging “reflexive interdisciplinarity,” PAI marks a counterstrategy with respect to mainstream media and expert policy discourse that often engages with AI as a finished technology, a “ready-made” set of tools and techniques, hence asking what formal principles inspired its designers, what material consequences it leads to, or both. Yet, while prevalent in structuring public discussion, this dualist framing overlooks “AI in the making,” as it is determined in and arises through particular situations, historical or contemporary. Not only do formal principles and material consequences take shape, and are appreciated and/or experimented with, in situ, but the common oversight of unfolding situations (Sormani Reference Sormani2023) also entrenches technological determinism and its “master narrative(s),” while foreclosing or reducing political futures to a dominant singular (e.g., “convergence,” “singularity,” or indeed “AI”). Ironically, technological “newness” claims often presuppose this received framing. So does the recent, if certainly justified, polemic around “big data”-driven and “generative” systems (Bender et al. Reference Bender, Gebru, McMillan-Major and Shmitchell2021). As a matter of course, it remains to be seen how (and whether) this critical frame will be developed in detail, in and through reflexive engagement with “performing AI.”Footnote 9
However, and in addition to criticality, the reflexive engagement of interdisciplinarity, including multidisciplinary commitment and transdisciplinary heuristics, also affords us – if not any prospective researcher – with a purposive methodology to probe “AI in situ,” as a distinctive performance/performativity nexus, confronting us with “performing AI” in and through its constitutive tension(s).
On the one hand, the notion of “performativity,” when understood as putting “AI” in a subject position, casting it as an agential power of its own, allows us to examine contemporary reifications of “AI”-labeled systems as “autonomous agents” (Johnson and Verdicchio Reference Johnson and Verdicchio2017), and that often against the backdrop of technological determinism. How is “technological determinism,” a long-standing trope and recurring motif, drawn upon as an interpretive frame (Wyatt Reference Wyatt, Hackett, Amsterdamska, Lynch and Wajcman2007), for example, in policy expertise? And how, when it comes to the “new AI” of machine learning and deep neural networks (Alpaydin Reference Alpaydin2016), does such determinism enable a polemic framing of media discourse, if not public experience (Quéré and Terzi Reference Quéré and Terzi2015)? Taken together, these two questions point to key topics of discourse analysis, while specifying an important aspect of the overall rationale of the PAI project. As an interdisciplinary inquiry, the project counters discursive reification, not to mention computational enframing (Agre Reference Agre, Bowker, Gasser, Star and Turner1997b). In particular, it makes explicit how “performativity” is rhetorically, materially-experimentally and practically sustained, if tentatively so – for “AI” to appear as an autonomous agent, despite (and precisely because) of the heterogeneous network it relies upon.Footnote 10
On the other hand, the notion of “performance,” when used to locate “AI” in an object position, depending on another instance, network or subject to “make it work” (e.g., a creative engineer, click worker or smartphone user), allows one to examine that object – “AI” – for just how it is enacted as a sociotechnical assemblage in situ (Holton and Boyd Reference Holton and Boyd2021). PAI’s case studies will make explicit the contingent performances of “AI” in and through their specific material circumstances, spanning policy discourse, studio-lab research/creative engineering and (other) public settings, including controversies, laboratories and museums – an important yet often overlooked context (but see Bertrand and Salter Reference Bertrand, Salter, Shehade and Stylianou-Lambert2024).Footnote 11 Therefore, PAI will be asking multiple questions:
• How is AI enacted in governmental policy?
• What does “artistic practice do to AI” and what does “AI do to artistic practice”?
• How is AI reconfigured in interdisciplinary scientific domains such as complex systems and artificial life (as opposed to computer science)?
• How is AI taken up, resisted or reconfigured in the public sphere, including schools, museums and festivals, focused on the intersections of art, technology and society?
These project-related questions have both specific disciplinary positions and sites in which they are anchored – policy spaces, an artistic “studio-lab” (Century Reference Century1999) at the Zurich University of the Arts, an interdisciplinary science lab at The University of Tokyo focused on “artificial life” (Lindegaard Reference Lindegaard2020), as well as schools, museums, art-technology festivals around the globe and more. In short, they entail multiple commitments, including disciplinary ones. Before we return to these commitments, let us pause on a key question: Cui bono?
The answer to the raised question is PAI’s orientation toward renewed artistic and public engagement. The project team(s) will leverage interdisciplinary inquiry – from critical analysis of technological determinism to multi-sited inquiry of “AI in the making” – to co-design new frames, forms and formats for what Born and Barry (Reference Born and Barry2010) termed public experiments in art-science. A “public experiment” can be defined as a practice-based approach which aims to play with, augment or challenge the resources of science, its “very idea” (Woolgar Reference Woolgar1988), rather than using art/design to “take a subordinate role in the communication of a finished science” (Born and Barry Reference Born and Barry2010, 115). In other words, the public experiment in art-science transcends the communication model of “sender-receiver-interpreter” as well as the figure of the artist as an illustrator of scientific concepts or producer of phenomena for an existing public. Instead, a public experiment of this kind aims to “forge relations between new knowledge, things, locations and persons that did not exist before,” and in this way engage in “producing truth, public, and their relation at the same time” (ibid., 116). PAI draws upon and, where judged necessary, adapts its dialogue model of agonistic pluralism (i.e., as developed via reflexive interdisciplinarity) to renew public engagement with AI (i.e., in the light of its variably situated “performativity”). Such a renewal is both critical and constructive. For one, it challenges deficit models of public understanding, where the alleged deficit of the general public is to be remedied (e.g., in the vein of technological determinism). For another, we aim to co-design public experiments, both epistemic and aesthetic, with multiple stakeholders, lay or expert, that can serve to reframe, if not reimagine, the past, present and prospects of “AI in the making,” aka its situated performance(s).
4. Discourse, agency, action: multiple commitments
In the introduction, we invited readers of this position paper to keep a transversal question in mind, the question of “what added value […] so many different disciplinary perspectives, methods and practices provide to the social-technological-aesthetic study of AI.” In turn, the bulk of the paper elaborated on how that “transdisciplinary” question (as we qualified it) should be tackled. Hence, our narrative account of a research paradigm and its reflexive realization as an interdisciplinary inquiry more broadly, including some of its (tentatively) defining moments so far.
Throughout the presentation of the PAI project, we highlighted the critical potential and methodological interest of reflexive interdisciplinarity to explicate “AI in situ” as a tricky assemblage. In addition to opening new research questions, a dedicated focus on “performance/performativity” was suggested to be of transdisciplinary interest for “public [AI] experiments in art-science.” Arguably, this distinctive line of public (and artistic) engagement makes a critical contribution to interdisciplinary inquiry on “AI” at large, if only by “cultivating dissensus” (Guattari, Reference Guattari2023 [1989], 37) and “singular production[s] of existence” (ibid.), irreducible to overly narrow determinism, technological or other. Yet the emphasis on this unique contribution to interdisciplinary inquiry also begs the question of other contributions, their multiplicity, both in terms of empirical methods and target phenomena, if not disciplinary traditions.
Again, the PAI project affords us a relevant case to reflect on. For this purpose, this section invites readers to practice “reflexive interdisciplinarity” by leveraging the text read (the preceding sections of this position paper) to revisit our multiple commitments, the particular research foci and disciplinary engagements on “AI” (as initially developed in the PAI grant proposal). Why? There is a double rationale for this “reading exercise.” Retrospectively, revisiting the PAI project, and its respective foci on “discourse, agency, and action,” will allow us – readers, researchers, colleagues – to take the measure of the progress made in the reflexive explication of the project’s paradigmatic interest. Prospectively, the reading exercise is expected to trigger “what’s next” moments, potential anomalies, if not transdisciplinary openings, which escape that research paradigm. We invite readers to keep those provocative moments in mind, as they revisit the successive project grant sections (as reproduced and slightly adapted in Sections 4.1–4.3).Footnote 12
4.1 Discourse performativity
One of the core notions of “performance,” if not “performativity,” may be traced back to the British ordinary language philosopher J.L. Austin, who argued in a set of lectures at Harvard in the 1950s that “speech also acts” (Austin Reference Austin1975). Language and discourse not only describe the world but also contribute to shaping it. From this perspective, discourse is not simply the reflection of an external reality but is material insofar as it gives form and substance to social relations. More broadly, in contemporary technoscientific societies, performativity is inherent in algorithmic culture, which equally shapes and constructs knowledge and realities (Just and Latzer Reference Just and Latzer2017; Seyfert and Roberge Reference Seyfert and Roberge2016). Analyzing what types of AI algorithms are enacted in AI governance discourse, and how, highlights the constructive, dynamic and relational nature of both (Amoore Reference Amoore2020; Birhane Reference Birhane2021; Liebig et al. Reference Liebig, Jobin, Güttel and Katzenbach2024). Therefore, studying algorithmic systems through discourse performativity as a thematic priority has ontological value in both academia and the policy domain.
Across science, policy and media (Jobin and Katzenbach Reference Jobin, Katzenbach and Lindgren2023), AI and its associated discourses have produced, highlighted, inhibited and displaced various controversies (Marres et al. Reference Marres, Katzenbach, Munk and Jobin2025). Such discourses and AI narratives have played a central role in (re-)directing attention and material resources in and through the political realm (Bareis and Katzenbach Reference Bareis and Katzenbach2021). AI policy efforts range from global efforts in soft-law and ethics (Jobin et al. Reference Jobin, Ienca and Vayena2019) as well as hard law (such as the European Union’s AI Act, cf. Ruschemeier and Bareis Reference Ruschemeier, Bareis and Gsenger2025) to subnational initiatives (Liebig et al. Reference Liebig, Güttel, Jobin and Katzenbach2022) and local activities (Squitieri Reference Squitieri2021).
Yet, these efforts have also faced criticism and contestation. Whereas some debates revolve around specific modes of governance, such as auditing (Costanza-Chock et al. Reference Costanza-Chock, Raji and Buolamwini2022) or interrogate stated goals and purposes (Ulnicane and Erkkilä Reference Ulnicane and Erkkilä2023), others adopt a more fundamental stance, notably raising the issue of governance by AI due to its algorithmic logic (Couldry and Mejias Reference Couldry and Mejias2019; McQuillan Reference McQuillan2022), its inherently sociotechnical nature (Selbst et al. Reference Selbst, Boyd, Friedler, Venkatasubramanian and Vertesi2019) or the invisible labor involved in its creation (Gray and Suri Reference Gray and Suri2019; Tubaro et al. Reference Tubaro, Casilli and Coville2020). Simultaneously, questions about AI governance focus on the different formal and informal institutions and processes that impact the mutual influence of and by AI (Gritsenko and Wood Reference Gritsenko and Wood2022; Ulnicane and Erkkilä Reference Ulnicane and Erkkilä2023). Together, these discourses demonstrate that there is no consensus on what “Artificial Intelligence” is understood to be and what it is actually doing. In other words, the widespread debates and activities around AI governance enact different kinds of AI. They raise the issue of performance in terms of their capacity to enact technologies and shape futures.
4.2 Material agencies of AI
AI’s performativity is not only enacted through discursive speech acts but also through the “material agency” (Malafouris Reference Malafouris, Knappett and Malafouris2008; Salter Reference Salter2015) of technical systems, where agency is a “temporally emergent” (Pickering Reference Pickering2010) effect of “subject and object relations that are distributed and always contingently enacted” (Suchman and Weber Reference Suchman, Weber, Bhuta, Beck, Geis, Liu and Kreis2016). Yet, the question of the agency of machines catalyzes a long-standing and unresolved challenge across the human, natural and social sciences as well as the arts around AI’s lack of “embodiment” (Brooks Reference Brooks1991; Dreyfus Reference Dreyfus1992; Hayles Reference Hayles1999; Merleau-Ponty Reference Merleau-Ponty1962; Mitchell Reference Mitchell2021).
Taking its cue from the fields of “enactive” or “embodied cognition” (Pfeifer and Bongard Reference Pfeifer and Bongard2006; Varela et al. Reference Varela, Thompson and Rosch1991), we engage with the longstanding question of whether and how embodiment acts as a minimal condition for intelligence, despite or precisely because of today’s increased focus on “deep learning”-based neural networks (LeCun et al. Reference LeCun, Bengio and Hinton2015), and specifically the generative AI models like Transformers (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017) that underpin LLMs like GPT (Radford et al. Reference Radford, Wu, Child, Luan, Amodei and Sutskever2019), Gemini (Google) or Claude (Anthropic). Unlike earlier neural networks, these models not only recognize patterns from existing big data sets but also generate new content from those data sets utilizing probability and statistical techniques (Foster Reference Foster2022). LLMs have been said to possess impressive inferential abilities and knowledge acquisition capabilities through text data processing (OpenAI et al., Reference OpenAI2023). These “unimaginably large systems of statistical correlations” have also been argued to be potentially capable of producing “abilities that are functionally equivalent to human understanding” (Mitchell and Krakauer Reference Mitchell and Krakauer2023).
Yet, these models are not only problematic due to their environmental and ethical implications (Bender et al. Reference Bender, Gebru, McMillan-Major and Shmitchell2021; Mitchell and Krakauer Reference Mitchell and Krakauer2023), if not their “digital politics” (Stalder Reference Stalder2017) more broadly, but also because of their lack of grounding in the physical world, with their expressions and activities largely confined to the digital realm (Yoshida et al. Reference Yoshida, Masumori and Ikegami2025). From the perspective of performativity in the context of linguistics, LLMs cannot exactly be said to be able to produce “speech acts” as they do not have “illocutionary intent” behind what is enunciated.Footnote 13 In other words, although it might seem otherwise, the statements that LLMs produce can be considered “infelicitous” performatives from the perspective of speech act theory – they are not genuine because they do not come from a speaker who is committed to communicate something intentionally to a listener, but are instead the result of statistical matching of existing patterns of words.
The question that then arises is: do LLMs have a performative quality and power that somehow transcends the core condition of the performativity of speech acts – the requirement that the speech act be felicitous (i.e., genuine) – and, thus, can only be considered “successful” when socially recognized rules and conditions for that act are satisfied? While the field of deep learning-driven AI in scientific and engineering research expands daily at a ferocious pace, at the time of writing, the embedding of LLMs into the physical world in both robotics and complex systems, as well as in machine learning art, has been understudied. For example, combining LLMs and robots has primarily focused on achieving planning and motion control based on human instructions (Brohan et al. Reference Brohan2023). Similarly, the embodiment of such models is usually focused on low-bandwidth text or image/screen-based representations (Zylinska Reference Zylinska2020; Reference Zylinska2023), leaving a gap in the vast possibilities of new interactions between audiences and physically oriented, machine-orchestrated environments.
Such discussions about AI’s bodily limitations (or indeed physical limitations) are not new in AI historiography. Indeed, as historian of cognitive science Howard Gardner wrote in the 1980s, the critique that phenomenologically bent philosophers like Hubert Dreyfus and John Searle (J.L. Austin’s student) originally undertook was that AI was “committed to an epistemological tradition foreign to virtually everyone in the world of computer science and artificial intelligence” (Reference Gardner1987, 163). Wittgensteinian critiques were also issued by early AI pioneer (and later Google founder Larry Page’s PhD advisor) Terry Winograd and former Chilean finance minister Fernando Flores in the mid-1980s (Winograd and Flores Reference Winograd and Flores1986), in which human language was argued to be a form of action embedded in social, emotional and practical contexts. PAI attempts to understand these critiques in a new context: not only deep neural networks and big data, but AI’s increasing agential systems that use LLMs to generate human-seeming text as well as other forms of human communication – written, sonic, visual and even gestural. Thus, a core focus of PAI is how AI systems might be re-embedded back into the physical world with its dynamic affordances – a perspective that unearths a series of counter-imaginaries and genealogies of practices from environmental psychology, phenomenology and machine-art (Broeckmann Reference Broeckmann2016) which are currently absent from the discussions of generative AI and, in particular, LLMs. In other words, the internal relations between embodied action, language use and social-cultural-environmental settings in machine systems require further investigation (Koerner Reference Koerner1992) – an area of mutual interest to both the sciences and the arts.
4.3 AI in situated action
A third strand of PAI examines how AI systems are concretely used as sociotechnical assemblages in “situated action” (Suchman Reference Suchman2007), being worked with and appearing to act in specific settings. Such “situated performances” of AI not only take place in policy, science and art laboratories but also in the public and private realms. The public release of “generative AI” systems has created a broad user base, enabling widespread “AI testing” while promising “democratize[d] access to knowledge and skill” (Perri Reference Perri2023). Yet, tensions remain not only regarding “disruptive automation” as a projected consequence of overall societal import, but also regarding the “testing definition” of the social in present situations, as, for example, in the current street testing of “AI” vehicles and their challenging character, if not deadly danger, for pedestrians and other traffic participants (Marres and Sormani, Reference Marres and Sormani2023). Research more broadly has inquired into the path dependencies of “generative AI” systems, both in terms of technological convergence having enabled “deep (machine) learning” to deliver upon many “AI” promises from the 1950s and 1960s, if ambivalently so (Jaton and Sormani Reference Jaton and Sormani2023), and in terms of the discursive tropes that continue to articulate this development. With the latest delivery of “generative AI” systems (e.g., ChatGPT, Dall-E, Bard, Claude and Gemini), their speedy uptake by online users and the dark warnings by their creators, the trope of “autonomous technology” that is somehow “out of control” (Winner Reference Winner1977) has indeed come to articulate a mundane social reality (Johnson and Verdicchio Reference Johnson and Verdicchio2017).
The same holds for “enchanted determinism” (Campolo and Crawford Reference Campolo and Crawford2020). Everyday uses of generative AI systems not only reproduce their multiple in-transparencies (Burrell Reference Burrell2016) but also invite folk explanation. In other words, tropes nurture a deficit model of (lacking) public understanding, highlighting “technological myths” and “popular belief” (Natale and Ballatore Reference Natale and Ballatore2020). However, despite the media frenzy around generative AI, the question of how people ordinarily interact with, through and against such systems remains understudied, especially in relation to current research in “deep learning” (LeCun et al. Reference LeCun, Bengio and Hinton2015). Therefore, PAI aims to probe the use of generative AI with lay users in selected public settings (likely to include museums, schools or universities) and specify what can be learned from mundane uses for reengaging with the perspectives of professional AI developers, expert commentators and/or policy-makers. How is developers’ “AI in the making” mediated by and taking into account the mundane understandings of lay members of society (Elliott Reference Elliott2019)? To what extent do they base their professional assessments, if not working routines, on such mundane understandings (e.g., regarding “technological convergence,” “societal needs,” “regulatory gaps”)? And, last but not least, how might public engagement with AI systems be renewed via empirically informed critique (Campolo and Crawford Reference Campolo and Crawford2020; Goode Reference Goode2018; Henriksen and Blond Reference Henriksen and Blond2023)?
Meanwhile, engaging with AI as situated action in public throws into relief larger questions of interactions between humans and computing machines. In particular, AI and Human-Computer Interaction, “two fields divided by a common focus” (Grudin Reference Grudin2009), have grown in scope and specialization and have sometimes intermingled (e.g., Harper Reference Harper2019). Yet, “hybrid studies” of AI/HCI that develop a common “critical technical practice” (Agre Reference Agre, Bowker, Gasser, Star and Turner1997b) remain scarce, as well as a reflexive explication of the implications for public engagement with “new (generative) AI” systems.Footnote 14 The same holds with respect to multi-sited ethnography and video-based interaction analysis of how AI systems are developed, tested and/or repurposed, all the more so as “testing in the wild” has become a societal reality, inspiring if not underpinning many “practicing art/science” endeavors (Sormani et al. Reference Sormani, Borgdorff, Peters and Pinch2019).
5. “On the ground”: performing transdisciplinarity, (re-)positioning the arts
Of course, we can only speculate what potential anomalies, if any, readers of this position paper will have identified against the background of the research paradigm that we tentatively outlined in the paper, as a multifaceted plea for reflexive engagement with “performing AI” as sociotechnical assemblage(s). From the outset, we stated that the PAI project is still in its early stages. Up to this point, we have elaborated, exemplified and exercised interdisciplinary inquiry on “performing AI” in mostly academic terms, as a reflexive antidote to machine reification, on the one hand, and as a diversified project of multiple commitments, including disciplinary ones, on the other. This dual approach, however, begs the question of the missing third, namely “the arts,” if not their “(re-)positioning,” as they become involved in “performing transdisciplinarity.” What is the “added value” of the arts in this regard, and with respect to tricky “AI” in particular?Footnote 15
To tackle this question “on the ground,” this section leverages the interdisciplinary paradigm posited at the outset. Consequently, we shall probe and ponder the transdisciplinary added value of the “arts on AI” from its (i.e., that paradigm’s) three perspectives. As part of a particular working discipline, we shall first collect promising anomalies from our “reading exercise,” candidly so, and explicate their aesthetic interest, if not artistic potential. Second, we shall document (what we termed) the “transdisciplinary contingencies” of two PAI subprojects, jointly dedicated to “machine performance” in renewed artistic terms. Third, we shall invite “agonistic dialogue” between the two, reading exercise and projected performance.
A) Reading anomalies and transdisciplinary results: On rereading the grant proposal section on “Discourse performativity” (4.1) above, we noticed a pervasive paradox: [reflexive interdisciplinarity, as a foundational methodology, requires “machine reification” to be deconstructed/yet “discourse performativity” is not exempt from a symmetrical risk, namely “rhetorical reification,” granting AI discourse a rhetorical power of its own and/or interpretive flexibility per se]. In turn, our reflexive encounter with “Material agencies of AI” (4.2) confronted us with this potential anomaly or at least intriguing phenomenon: [performance/performativity was introduced as a boundary object for probing “AI in situ”/yet the “material agencies” section indulges in performativity ex cathedra, if not ex machina, enrolling “physics,” “bodies,” “society,” and more]. Finally, and upon pondering “AI in situated action” (4.3), the section struck us as [largely missing out on the reflexive coproduction of “situated action” and “sociotechnical assemblage”/thus reproducing, instead of dissolving, the lingering risk of performative discourse and machine reification].
Taken together, the flagged anomalies – in the practice-based vein of our twofold Kuhnian approach, with project materials at hand – highlight the suggestive rhetorics, arguable lack of semantic clarity and possible conceptual incoherence, betraying the grant proposal, if not its “genre.” So what? After all, the proposal was funded, and the project being underway! Yet the flagged anomalies, more interestingly, are of reflexive interest too, each of them alluding to a particular aesthetics, if only for the particular media engaged in, including “creative writing” via discourse performativity (1), “machine performance” as material agency (2), and “frame breaking” due to (hampered) situated action (3), respectively. As transdisciplinary openings, they remain tied to this position paper. Hence, they may well contribute to articulating “arts on AI” in and through contrasting aesthetic moments.Footnote 16
B) “Machine performance” as transdisciplinary contingency: At the time of writing, two PAI subprojects are dedicated to exploring and experimenting with “machine performance” in (renewed) artistic terms. What “hybrid experiments” are they envisioning? And what “artistic terms” might they be renewing? Let us address these questions.
The plan: The mentioned PAI collaboration, to paraphrase from its transdisciplinary (sub-)project, envisages the creation of a “performative installation” involving interaction between a collective of human performers, visitors and machines. Such a machine then could not only act as a quasi-object that could constitute a social-technical-aesthetic hybrid, as a potential “boundary object,” but it could also assemble and materialize predominantly philosophical questions around the gap between referents and meaning, symbol and embodied grounding in debates around LLMs. Unlike classical symbolic systems, LLMs operate not through explicit causal models but through correlations among vast amounts of linguistic data. This shift, it seems, destabilizes the very notion of “understanding” in both scientific explanation and artistic interpretation. Whereas science has traditionally sought to uncover causal relations, LLMs generate meaning through statistical resonance without a causal mechanism. They do not so much model the world as perform (enact) it through correlation. In this sense, their outputs instantiate what might be called synthetic causality – effects that precede any stable cause.
The point: From an aesthetic perspective, this correlation-based logic destabilizes conventional notions of representation and, as discussed above, intentionality. Artistic practice engaging with LLMs thus could not so much demonstrate or “represent” AI as an object (as is currently the trend) but instead, following from the preceding arguments, co-perform with generative processes that lack interiority or affect. The traditional framework in which artists employ deep neural networks or LLMs merely as tools may then give way to a performative collaboration between artist and agent, each operating within distinct (synthetic) causal regimes. In such encounters, the inevitable “hallucinations” of the LLM agent may transform into creativity through co-performance with the human. What then could emerge, also in a methodological context, is a new form of correlation-based co-creation and (participant) observation that transcends causal explanation – a relational aesthetics of mutual emergence between human and machine. Such coupled enactive actions would not stop at the process between creators and performative machines. They would also carry over to the spectator/observer, whether audience or social scientist.Footnote 17
C) Moments of “agonistic dialogue”: The paradigm articulated in the introduction afforded (and affords) this paper section with a common frame to address transdisciplinary openings for engaging “arts on AI.” Both despite and because of that frame, the section triggered contrasting results so far, retrospectively tied to a “reading exercise” (A), prospectively articulated as a “performative installation” (B). In particular, that reading exercise now exhibits this kind of installation, respectively, its prospective articulation, as playing into one aesthetic register, “machine performance,” rather than another. Indeed, the prospective articulation of the envisaged installation highlights its “machine performance” as a form of material agency (2), rather than as a rhetorical achievement via “creative writing” (1) or reflexively instructive action via “frame breaking” in situ (3). As a promising transdisciplinary encounter between two projects, the highlighted aesthetics of “machine performance” in turn raises the question of how it will incorporate, if not transcend, the disciplinary commitments, epistemic and/or aesthetic, particular to each project, not to mention its relationship(s) with the other PAI projects and their research priorities, discursive and situational (as outlined in Section 4). The question raised hints at “agonistic dialogue,” not so much in the sense of a joint challenge to a received framework, disciplinary or interdisciplinary (Barry and Born Reference Barry and Born2013, 12), but rather insofar as it problematizes the very assumption of common ground (be it “situated action,” “synthetic causality,” or “algorithmic culture(s)”). The expression “on the ground” lent itself to contrasting understandings, indeed. As part of interdisciplinary inquiry, this kind of problematization remains ostensibly ambivalent. For sure, it might be considered a “healthy” expression of methodological reflexivity, at least if we stick to the threefold paradigm of interdisciplinarity advocated in the introduction to this paper. Arguably, this includes alternative transdisciplinary openings (A versus B), if not aesthetic opportunities (among which “performative installation”).
However, a radical problematization of common ground also challenges the confident introduction of shared concepts. Even “working definitions” may dissolve. Throughout the paper, we highlighted the fragility of “sociotechnical assemblage” as a shared concept, a fragility accentuated by the “performance/performativity” tension at the core of “performing AI.” Indeed, some approaches accentuated AI’s “performativity,” discursively and/or materially, while others homed in on its situated “performance,” pragmatically, by something or someone else. Hence, present readers too may worry about the fate of the performance/performativity nexus as a “boundary object.” This worry was anticipated by Susan Leigh Star as she introduced the concept not only as a common denominator for multidisciplinary cooperation, together with Griesemer (Star and Griesemer Reference Star and Griesemer1989), but also as a collaborative invitation to probe “distributed AI” with a Durkheim Test (what community, if any, does it serve?), rather than a Turing Test alone (how well, technically, does it perform?) (Star Reference Star1989). And this move includes interdisciplinary inquiry, as projected and reflected upon in this paper.
6. Conclusion
In this position paper, we have outlined an interdisciplinary paradigm for reflexively engaging with contemporary AI as a sociotechnical assemblage. We then described how this paradigm has been operationalized in the research project Performing AI, while tracing and explicating some defining moments of that very process. These moments included the performativity/performance nexus as a “boundary object,” as well as pending anomalies and transdisciplinary openings. Gathering an interdisciplinary team, the PAI project ambitiously, if not “wildly,” brings together the social, natural, human sciences and the technoscientific arts to explore “AI in the making.” To deliver on its interdisciplinary program, PAI is supported by the SNSF and, in the years ahead, will focus on AI in concrete situations, as a contingent, temporal, unwieldy and contested social-technical-aesthetic object – itself continually in becoming – rather than as an unequivocal name (or frame) for stable technology. To grapple with AI’s chameleon-like qualities, the outlined approach and its triple reflexivity, articulating multi-, trans- and interdisciplinary moments, seems both necessary and warranted, if only to relocate the still ongoing debates about AI and its always shifting goal posts. Therefore, PAI’s interdisciplinary outlook takes its cue from the multifaceted and intriguingly paradoxical character of “AI in the making” – performance and performativity in one.
Throughout the paper, we articulated reflexive interdisciplinarity as a foundational methodology for probing multifaceted “AI in situ,” while documenting (and performing) its tentative operationalization in and for the PAI project. Moreover, we laid out some of the project’s historical background, epistemic challenges and critical insights. Yet it is important to point out that all research, if partly determined by its past and present (e.g., a customized version of Kuhn’s paradigm), remains also speculative, especially in the early stages. Most of PAI’s interdisciplinary inquiry thus still lies in the future and, as the historian of life sciences Hans-Jörg Rheinberger might emphasize, finds itself on the cusp of both knowing and unknowing. Even within such speculation, though, it should be evident that an interdisciplinary approach to AI is needed if we are to reframe how we talk about, develop and use (or discard) that thing labeled “AI.” Otherwise, we are at risk of reproducing technological determinism (e.g., Bostrom Reference Bostrom2014), computational reductionism, and/or polemic framing – in short, the bland redundancy feared by the skeptical reviewer of the PAI project. In turn, the project’s creative potential and critical relevance remain to be worked out, toward new possibilities of understanding, acting and relating, in and through interdisciplinary inquiry, articulating both multi- and transdisciplinarity, whose likely contentious research future still largely remains unknown.
PAI’s prime goal indeed is to investigate “AI” from epistemic, ontological, aesthetic and ethical angles, neither taking for granted its “uncontroversial ‘thingness’” (Suchman Reference Suchman2023), nor assuming received disciplinary frames to fit the purpose. Instead, different but potentially overlapping knowledge practices, methods and cultural activities are leveraged. “Wild interdisciplinarity,” then, might be best glossed, if not worked out, as multiplicity in the making.
Acknowledgements
This article has benefited from helpful comments by Georgina Born and Tobias Blanke, as well as Julie Marques, Yulia Kukles and Selim El Madani.
Funding statement
All authors are funded by the Swiss National Science Foundation, grant no. 10002211.
Competing interests
The authors declare none.
Philippe Sormani (PhD in Sociology) is a Senior Researcher at the STS Lab at the University of Lausanne and the Immersive Arts Space (IAS) at Zurich University of the Arts. Drawing on science, technology and media studies, he has published on experimentation in and across different fields of activity, ranging from experimental physics to artistic experiments and technology demonstrations. Currently, he is experimenting with “The Metaverse” (at IAS) and contributing to interdisciplinary inquiry on “New AI” (at STS Lab and beyond), while taking a particular interest in technoscientific paradoxes and aesthetic possibilities. Key publications include Respecifying Lab Ethnography (Routledge, 2016), the co-edited volume Practicing Art/Science (Routledge, 2019) and “Interfacing AlphaGo” (Social Studies of Science, 2023). As a sociologist of science, technology and media in action, Sormani has longstanding expertise in ethnography, ethnomethodology, and video analysis.
Takashi Ikegami is a Professor in General Systems Science in the Graduate School of Arts and Sciences at The University of Tokyo and principal investigator of the Ikegami Lab. He received his PhD degree in physics from The University of Tokyo and is internationally known for his contributions to the development of complex systems science and artificial life. Some of his results have been published in Life in Motion (Seidosha, 2007) and Between Man and Machine (Kodansha, 2016). He has also been active since 2005 in the arts with works such as “Filmachine” (with Keiichiro Shibuya, YCAM, 2006), “Mind Time Machine” (YCAM, 2010), “Long Good Bye” (with Kenshu Shimpo, Japan Alps Festa, 2017), “Offloaded Agency” (Barbican, 2019), among many others.
Alexandre Saunier is an Artist, Professor in the Audiovisual Department at LUCA School of Arts, KU Leuven, and senior researcher in the Immersive Art Space at Zurich University of the Arts (ZHdK). His research intersects artistic practice with the theory and history of media arts, cybernetics and artificial intelligence, focusing on live performances involving light, sound, autonomous systems and sensory perception. Alexandre holds a PhD degree in 2023 from Concordia University, where he studied the contemporary and historical practices of light as an artistic medium driven by real-time computational systems. His previous studies include mathematics and physics, sound design and engineering at ENS Louis Lumière, and he was a fellow at ENSADLab, where he conducted research on behavioral robotics and interactive lighting. Saunier’s artistic and research work is regularly presented at major international venues.
Anna Jobin is an Assistant Professor at the interfaculty Human-IST Institute at the University of Fribourg (Switzerland), where she leads the Social Studies of Algorithms, Internet & Society (SAIS) research group. Her research lies at the intersection of technology and society, focusing on how AI and algorithmic systems are designed, governed and contested, and how they reconfigure knowledge, power and public life. She co-edited the Big Data & Society special issue on the controversiality of AI, and her research has been published in Nature Machine Intelligence, Social Media + Society, AI & Society, and Handbook of Critical Studies of Artificial Intelligence. With a multidisciplinary background in sociology, economics, and information management, she has held affiliations at ETH Zurich, EPFL, Cornell University and Tufts University, and was elected inaugural member of the Swiss Young Academy (2020-2025). At the University of Fribourg, she coordinates and teaches in the Digital Society master’s programme.
Olivier Glassey is a Sociologist and Anthropologist by training. As a Senior Lecturer and Researcher at the University of Lausanne, he teaches and conducts research at the Faculty of Social and Political Sciences. Since July 2015, he has also been the Director of the Musée de la main in Lausanne, a museum devoted to exploring the multiple ramifications of science, technology and medicine in society. His research interests include the social uses of information technologies and digital cultures. He is particularly interested in the ambivalent implications of technological mediations on our everyday practices, specifically in terms of the circulation and dissemination of knowledge, the production of collective memory, and new forms of online sociability. His research has been published inter alia in SociologieS, Histoire et Informatique, Sociologie et Sociétés and via the Institute of Network Cultures.
Christopher L. Salter is a Professor and Director of the Immersive Arts Space at the Zurich University of the Arts and Professor Emeritus of Concordia University. At Concordia, he co-directed Hexagram – International Network dedicated to Research-Creation in Media Arts, Design, Technology and Digital Culture. His work sits at the intersection of Performance and Technology, and Art, Science and Technology more broadly. His most recent book is Sensing Machines: How Sensors Shape Our Everyday Life (MIT Press, 2022). Previous major publications include Alien Agency: Experimental Encounters with Art in the Making (MIT Press, 2015) and Entangled: Technology and the Transformation of Performance (MIT Press, 2010).
Open access
