ULTRACHUNCK (2018) is a performance by Memo Akten and Jennifer Walshe that unfolds as an improvisation with an artificial intelligence (AI) model that is pretrained to recognise and mimic Walshe’s voice and face (Technosphere Magazine, 2018). The AI data set consists of multiple hours of Walshe’s solo vocal improvisations, which were recorded on a daily basis for one year in front of her webcam (Akten, 2018). The scene of the performance is a duet where Walshe improvises with a projection displaying the AI-generated images that morph and evolve in relation with her live performance. The deep neural net performs video and audio content live, which is neither sampled nor processed (Akten, 2018). Akten (2018), who designed the model and interface, states on his website that the goal was not to create sounds that are indistinguishable from the training data, but sounds that are ‘also novel and un-human, yet still interesting, and – most importantly – allow a human user to expressively, and meaningfully manipulate the output in realtime’.

Walshe’s vocal modulations during the live performance often sound like they are being teased out from the throat and mouth, between the screeching and the guttural, the stochastic and the continuous. Contrasts are extreme, and it is the vocal play across these contrasts –  unbridled by conventions of interpersonal communication – that lends form and dynamism to the performance. Alongside this vocal articulation, we witness the extreme contortions of Walshe’s face, pulling taut into unhabitual thresholds of expressibility. Meanwhile, the computational performance (sound and images produced by the machine learning algorithm) flickers and morphs into and out of similarity with Walshe’s live performance, giving way to unexpected orderings and correspondences between the two performing entities. For instance, vocal and facial articulations seem as relevant to the context of sound and image generation as local lighting conditions in the orderings and pacings of the projected images.

Figure 1. ULTRACHUNK was commissioned by Somerset House Studios with the support of the Case Foundation. Photo © Anne Tetzlaff.

This staging of the photographic image in ULTRACHUNK is situated within a broader context of machine learning neural networks, and a growing repository of art-based experimentation with them.¹ Antonio Somaini (2023: 106) specifies that this constitutes a redefinition of the field of the visual and is characterised by a ‘layered referentiality that cuts across various forms of mediation, involving both images and words’. In this ‘flat ontology of the latent space every point has the same status as all other points’ (Somaini, 2023:106).² In the machine-generated image, this ontological flatness converges with the appearance of photographic representation in very particular ways, often inviting considerations of how the image is composed as a speculative and transversal process, that melds material forms with the image of overexposure and depth of field, or that renders unexpected continuities between form and ground.³

Joanna Zylinska suggests that the informational remediation of the photograph of AI does not instantiate a new category for the image, but rather encourages an understanding of the photograph as a set of media gathered by what Yanai Toister (cited in Zylinska, 2023: 4) calls a ‘resemblance concept’, which is premised on whether the media in question ‘looks and feels like photography to the human observer’. Included here are media that repurpose lens-based residues through digital and computational means. But however much the photograph has historically stood as an autonomous object – both in visual studies and within the social sciences, where the photograph has served to both fix and categorise pictured objects – the photographic image has always maintained an implicit heterogeneity that exceeds resemblance. Vikki Bell (cited in Lury & Wakeford, 2012: 16) observes that implicit in the photographic image is the situation of its circulation and witnessing, such that its content cannot be captured as a positivity for social science. What’s particular about the redefinition of the photographic image through AI – and is the case in ULTRACHUNK – is that this quality of witnessing is made explicit as the spectator is tasked with apprehending the image as both content and mediality – as both information and picture – and their positionality when occupying these incommensurate yet co-present vantages.

Importantly, for Walshe and Akten, the co-compositional quality of ULTRACHUNK decentres the model of the Turing test, which Kathy Fang (2024: 132–133) describes as a means of evaluating the ability of the machine to exhibit cognitive behaviours that are indistinguishable from those of a human, in a kind of imitation game through typewritten responses. Through statistical data analysis and inference-making, these computational systems remediate likeness such that they end up ‘looking “intelligent” to a human’ (Zylinska, 2023: 5; Zylinska, 2020: 51). While the machine learning algorithm in ULTRACHUNK responsively imitates Walshe’s performance, this imitation game asserts a critical posture through a practice of witnessing as a mutual and rigorous belabouring at the thresholds of expressibility. The machine-generated image actively displaces, even consumes, the possibility of a passive posture aligned with the image passing for something else – or of the computational image passing for the photographic one – by situating its generativity at the tentative and uncertain interstices binding image, witnessing and computation.

The disparity between Walshe’s laboured embodiment and the seemingly immaterial ease of computational generativity is a noteworthy product of the fact that the energy and resources that go into the computational process are materially and rhetorically hidden from view. At the same time, the situated materiality of Walshe’s performing body, engaged in the process of image witnessing, constitutes an opening to an otherwise of computational media. This is an otherwise that is always already co-present with the status quo but primes a heightened attentiveness to a situated and companionable AI aesthetics.

ULTRACHUNK, and the sociability of AI in general, thus constitutes a remapping of the field of the visual by enlarging the dispositif of the photograph, and of computation, to include the diverse ways the image and information are accompanied. The term ‘dispositif’ is apt here, since it enables tracking the specificity of this broadening, in ways that exceed the empirical – or, following Michel Foucault’s definition (cited in Lury & Wakeford, 2012: 6), it ‘organizes heterogeneous elements into a system of relation, whose propositions are both said and unsaid’. Celia Lury and Nina Wakeford (2012) take up Hans-Jörg Rheinberger’s concept of inventive methods to elaborate how entities such as devices, dispositifs, assemblages, anecdotes, categories, phrases and probes (which we can extend to protocols, scripts, scores and prompts) achieve the active participation of the societies in which they take place. The inventiveness of a method has to do with ‘its ability to generate its own boundary conditions – to organize itself – to self-organize – in a (changing) relation to a (changing) context’ (Lury & Wakeford, 2012: 6). Importantly, Lury and Wakeford (2012: 6–7) consider how the use of a particular method or dispositif contributes to the framing of change, to change the problem, such that the method becomes answerable to its problem by remaining open to new terms of differentiation and its own subversive movement.

This puts into question the way computational media, as dispositif, comprises diverse computational and lens-based formats and genealogies, as well as how these methods enable inhabiting their perceptual, affective and situated dimensions. To engage machine companionability as a subversive modality thus means to tend to how machine co-presence might instantiate emergent modes of co-presence from a variety of postures, affinities, processes and sensitivities. For Donna Haraway (2003: 8), to begin with a companionability question means a shift away from the taxonomy questions that have historically dominated the study of science and technology in the Western world. This pursuit is intimately linked to feminist enquiry, since it privileges possible relationships rather than given ones, understanding ‘how things work, who is in the action … and how worldly actors might somehow be accountable to and love each other less violently’ (Haraway, 2003: 8). How, then, might AI as a companionability question enable interrogating the way we create and participate in categories and their overarching influence in structuring knowledge practice, and how might it become a means of enlarging the scope of this entanglement to include diverse communities and more-than-human agencies? At a time when AI is increasingly dominated, regulated and monetised by a few corporate bodies, the urgency here is to destabilise the terms of engagement, to elaborate modes of life that exceed representation, capture and categorisation, in favour of inclusive and ethical futures.

This discussion engages the incommensurability of AI – the black-box operations that exceed positivist means of interpretation – as a call to transverse modes of companionability that cohere a diversity of media practices and conceptual frameworks. This enquiry tracks how performances with AI make AI computation sensible as a deranging and disorienting force, by putting them into discussion with examples of incommensurability that are generative within a diversity of media and modes of engagement. The result is a series of explanations that are both tentative and exploratory, and which articulate speculative mappings of AI incommensurability, and its companionability, across different manifestations of mediatisation, performance and storytelling. The next section of this discussion begins with a more in-depth examination of AI incommensurability and how AI explainability constitutes an opening for speculative engagements with AI incommensurability as a critical intervention and ethical invitation.

The Performativity of Explainability: Answering to Incommensurability

Media scholar Beatrice Fazi (2021: 69) approaches AI from the vantage of the computer science subfield of explainable AI, emphasising that the black box of deep neural net computation presents an explainability problem, which requires confronting the incommensurate. Fazi explains that this is because empirical registers of AI explainability are only adequate to certain positionalities, and there is no consensus about their utility. In the context of explainable AI, empirical mappings of algorithmic operations have been replaced by training secondary models whose role it is to optimise explainability through probabilistic mappings, which themselves face issues of hallucination and bias: ‘Epistemological reconstruction assumes a life of its own via algorithmic models that do not aim to represent and thus do not wish to explain’ (Fazi, 2021: 65). This problematic positions the human in relation with the discrete incommensurability of black-box operations, based on the premise that humans ‘require a continuous, and thus analogue, association with the world’ (Fazi, 2019: 4). Fazi (2021: 65) proposes a case in point: humans learn to see zeros by identifying similarity across forms resembling ovals, while machines cannot see such similarities but can anticipate zeros through a probabilistic distribution of weighted pixel values. Fazi (2021: 71) thus asks whether ‘giving enough speculative credit and attention to the incommensurable operations of artificial cognitive systems is enough to produce such a shared account of a useful, successful explanation’. Citing Langdon Winner’s polemical essay, written in 1993, suggesting that the risk of ‘opening the black box … was finding it empty’, Fazi (2021: 71) rather indicates that the risk might be ‘finding nothing to translate … because the possibility of human representation never existed in the first place’. This evocation of non-translatability situates information within categorical oppositions between the continuous and the discrete, the analogue and the digital, the human and the machine. The question then becomes: what if we speculatively enlarge the location of the black box to include the disparate data sources that inform AI computation, not only as the information of computation, but as media that informs and implicates posturing and witnessing, such that our attempts at translation exceed and reorient the boundaries of contained computational systems? What might be translatable within this expanded view, and what are the terms of translatability?

A recent academic publication by Apple presents the limitations of explaining the problem of AI reasoning somewhat differently (Parshin et al., 2025). The authors designed a series of relatively simple puzzles such as ‘Tower of Hanoi’ for a large reasoning model to solve by relocalising elements in space, in a particular order, in the smallest number of steps. The aim of the study was to analyse the traces of internal reasoning, as a means of enlarging the possibility for analysis beyond output accuracy. Interestingly, the authors mention the limited transferability of their findings to more generalisable reasoning, explaining that ‘the use of deterministic puzzle simulators assumes that reasoning can be perfectly validated step by step’ (Parshin et al., 2025: 11). This underscores the structuring force of the testing dispositif itself, where the terms of reasoning are implicit in the method of testing, reproducing what is knowable within a set of parameters.⁴ By extension, this also suggests that the structuring force of such attempts at translating black-box operations also reinforces a concept of AI interiority, opposed to a human exteriority. This interiority takes precedence, over and above the transverse ways that media and information co-compose diversely across human and non-human actions, media sources and infrastructures.

Theo Phan and Scott Wark address the production of interiority, and of ontologically discrete entities in AI media, from the angle of personalisation. After Wendy Chun and Adrian Mackenzie, they elaborate how personalisation algorithms operate as ‘human-technical ensembles [that] … produce abstractions – in other words, categories – that they put to work on us’ (Phan & Wark, 2021: 21). As algorithms organize identifiable affinities into increasingly granular categorisations, persons are solidified based on continuities of liking and likeness, which reasserts gendered and racialized bias as a calculable element within the larger system of algorithmic function. This presents an imperative to consider an enlarged vision of what techniques, postures or modes of affinity might then recompose the terms of personalisation and its effective calculability. How might a reconception of affinity and its articulation of persons, be recast as an evolving and co-composing context of divergent and transversal dependencies? How might such a reconceptualization, where the difference between interiority and exteriority remains persistently uncertain, enable critical and inclusive futurities with and alongside AI?

While the incommensurate is a concept that risks anticipating and overdefining the entities that it places in relation, it can be an invitation towards modes of co-presence that privilege multiple, varied and sometimes discordant compositions. After Gilles Deleuze, Thomas Lamarre (2002: 153) takes up the question of the diagram, which he describes as the task of enjoining openness to the tool ‘that would transform the one who picks it up, profoundly’. In this statement, Lamarre (2002: 157) is referring to the diagram of brushwork in Heian calligraphy, a gestural and manual practice that ‘passes over and through characters, conjoining visual and vocal elements. It is as if the movements of the hand/brush could stitch together or clump different registers or potentials’ (Lamarre, 2002: 157). In this example, characters are encoded within a communication economy but ‘do not exactly obtain at the level of brushwork’ (Lamarre, 2002: 154). As form, process, information and that which does not obtain co-compose, they do not map to a given subject or technology but to a co-compositional dynamism that exceeds them. This altogether shifts the analytical posture of translation from one of objectification to a communication economy that embraces a changing relationship with the terms of its co-composition, exceeding sign, gesture and image.

Alanna Thain offers the notion of queer incommensurability to further specify the excess of the incommensurate as a deranging force, which is apprehensible in the context of reverse motion in screendance. Thain (2024: 90) specifies that this kind of incommensurability is not predicated in syncing up or mapping onto, but expands kinaesthetic experience to include a kind of ‘companionable otherness’. Referring to Renate Lorenz and Pauline Boudry’s 2019 screendance installation Moving Backwards, which, through choreography and complex video-editing, confuses ‘the distinction between backwards and forwards’ (Philadelphia Museum of Art, 2025), Thain (2024: 86) elaborates that this reversal is ‘a contagious, deranging, and disorienting movement impulse … situated at the ambiguous imbrication of bodies and technologies … [which] resist normative constraints of time, to animate instead a time in the making’. Importantly, Lorenz and Boudry propose the term ‘temporal drag’ in the broader scope of their work as a means of addressing social and political regression without recourse to progress narratives (Thain, 2024: 92–93). Temporal drag signals the ‘co-presence of different temporalities, and of opening up a moment that was not properly lived or actualised in the past, to give it another try’ (Thain, 2024: 92–93). In this folding of the past into the present, temporal drag becomes a kind of politics that ekes out ethical futures through a technical pragmatics that doubles back on itself, that revisits its own terms of intelligibility through a mode of accompaniment that suspends the status of the objects that compose it. Temporal drag situates the workings of the technological not within a specific object (such as a machine) but within a system of evolving orchestrations across diverse and sometimes indistinguishable forces. This perspective offers a way to consider a companionability of the incommensurate that does not assume a Whole, or complete vision of incommensurate entities but can nonetheless offer new terms of habitable difference.

It is worth briefly elaborating the cinematic example of Jean-Luc Godard’s (1982) Lettre à Freddy Buache to further consider answering to the incommensurable as an iterative process, which is not reducible to binary nor continuous terms of referentiality. Deleuze (1989: 181), referring to this example, states that ‘the cuts or breaks in cinema have always formed the power of the continuous’, where the continuous and the discontinuous are mutually constitutive rather than opposed. Deleuze (1989: 182) continues: ‘What are opposed, or at least distinguished, are rather two ways of reconciling them, according to the transformation of the Whole’, where the incommensurable of Godard produces a ‘dislocation in the internal monologue’. In the short film this dislocation manifests multiply and includes a reconfiguration of the internal and the external, surface and substance, the public and the private.

Setting out with the ambition to undermine the status of the film as a commemorative documentary about the city of Lausanne, for which it was commissioned, the film aligns what is pictured with diverse frameworks for abstraction. It does so by presenting the city in a series of explicitly stated ‘plans’ that rupture and conjoin otherwise familiar scenes. Camera movement appears to sketch or stitch images and scenes together, scenes that are governed by contrasts: darkness and lightness, colour and form, texture and angular geometry, or movement and stillness. This series of scenes is coupled with cuts to images of Godard himself, at the manual controls of image and sound recording and playback. Godard’s voiceover (or would-be ‘internal monologue’) explains his thinking around the work for the entirety of its duration, which meanders with and across a series of suggestions about the cinematic image and its abstraction, and voice itself, as a means of escaping the spirit of geometry, which is the ‘fiction’ of the city. By way of film, he suggests, scientific thought can emerge at the periphery, or otherwise, where fiction begins. Such an enquiry is afforded by the media context of a film that has not yet come to the ‘sur face’ (spoken by Godard in French with a disjunctured emphasis, gesturing towards a disrupted alignment of image surface  and certainty, and the assumed Wholeness of  pictured objects). The ‘depths’, where he states the film remains, characterise the speculative and experimental posture of the film, one that accommodates and composes with the incommensurable as a radical opening, and refusal to map stable definitions of image, form or genre.

Incommensurability is here a posture that implicates, deranges and discoordinates a Whole through the concepts and mediatisations which it binds and in which it participates. This focus recentres the human-machine relation not as a binary mapping of the two, where the terms for each are given and transparent; instead, it serves as a call to give speculative credit as a means of enlarging the ‘beginnings of fiction’, which need not begin or end in fixed relation to a Whole, but instead gather discontinuously from between and across possible configurations.

We can now return to ULTRACHUNK to consider how temporal drag remakes the human and machine as a heterogeneous and co-composing force. While ULTRACHUNCK aligns continuous performance with continuous computation, temporal drag emerges as the speculative negotiation of the optimisation of  performative responsiveness. The effect creates the sense that Walshe herself and the object of computation are tentative, even searching positionings, where in witnessing, one questions what one is supposed to be seeing or listening for, and where the object of the performance emerges gradually and speculatively. This renders the semblance of Walshe becoming photographic, or engaged in an expanded photographic visuality, as if a virtual image surface and its persistent computation of light values were pressing, folding and thinning Walshe’s face into tentative form.⁵ While strange and disconcerting, the result makes sensible a spacetime of co-composing and sometimes indistinguishable human and non-human agencies in time, converging the skin and the generative image surface into a newly sensitive connective tissue.

Re-re-mapping Exposure

I now turn to two examples that combine mapping algorithms with the excess of light, exposure and luminescence that enlarges photographic visuality and activates the deranging force of temporal drag. The first is Nicolas Bourgeois’s Rompre l’espace, corps fugace, and the second is Maxime Alexander-Gosselin’s Re-re-collections.⁶ Rompre l’espace is an electroacoustic performance assisted by musical algorithms that uses Wekinator, a supervised learning algorithm developed by Rebecca Fiebrink and Max software. In the performance, a single human performer is positioned within a square space on the floor marked by electrical cords supplying power to four light stands, one in each corner. Equipped with pressure-sensitive position-mapping gloves, the performer moves, triggering flashing lights and synthesising discordant violin sounds. The movements of arms and hands are emphasised, but not constrained by any particular orientation or direction. Sometimes the movements are broad, sweeping strokes, and at others they are cut by precise, punctuating instances. Repeated movements seem to test the AI tasked with gestural recognition at different limits of intelligibility, while at the same time putting into question what it means to move, and for whom, within a performative setting.

Figure 2. Still image from a video of the performance of Rompre l’espace, corps fugace created during Ultrasons 2024, University of Montreal.                Photo © Gabriel Geneau.

In the context of the performance, Bourgeois’s physical form also seems to weave in and out of material presence, sometimes disappearing almost completely into shadow or overexposure. In this partial illumination, gestures do not remain fixed to figural form but push and pull the surrounding space. This phenomenon is something Dana Reitz (Buckwalter, 2010: 84) describes in relation to her dance improvisation Necessary Weather,⁷ a work Reitz composed with Jennifer Tipton. In this performance a single performer wearing swinging, loose-hanging clothes dances with top-down illumination, such that the way the periphery reaches the edges of the performer’s body becomes an emergent and dynamic score for movement. The entire space of the room is implicated as the importance of the movement shifts from one delimited by bodily agency to the threshold of illumination and darkness. This newly distributed agency unsettles the sense of physical containment and discrete localizability of the performer, as well as bodies that witness and inhabit its spacetime.⁸ The creators describe the work as an enquiry into the ‘climates of movement and light’ and a ‘journey, in silence, along the edges of dream and real time’ (Reitz & Tipton, 1993).

In Rompre l’espace, differences between light and dark, interface and accompaniment, resemblance and intelligibility present a similarly distributed texture of co-composing articulations, exceeding mappable images of figural form. More than coming into and out of the void, the performing body inhabits the AI dispositif by elaborating the terms of non-recognition, even as the AI algorithm persistently recognises certain elements, figural or otherwise. This effect is accentuated by the AI-generated sounds that seem pulled from or by performed movements. At the same time, the performance collapses the difference between space, light, figure and movement, as their various co-compositions are invariably available to AI recognition. For Bourgeois (2025), all of this contributes to improbable interpretations, which ‘transcend the tangible value of the body … [which] then becomes a sensitive and fleeting interface, a subversive field of exploration’.

As in ULTRACHUNCK, Rompre l’espace enlarges the AI interface as both score and companion through the continual conjoining of performed gestures and the responsiveness of AI feedback. At the same time, both examples give the question of what obtains central importance, as the terms of AI recognition remain constant amidst a breadth of performed variations. Fang (2024) observes a similar phenomenon in her experimental role-playing games with Open AI’s ChatGPT, leading her to suggest that AI does not perform but rehearses, in a constant register of optimisation. While Fang is describing a different learning algorithm, and a different posture of engagement, the persistent capture that renders all movements tentative relative to machine responsiveness bears a resemblance. Fang (2024: 145) states that these rehearsals were not

exceptional, transformative instances of performance in the liminal zone of the chatbot-become-human, but mere instantiations of the chatbot’s most normative behavior – one more rehearsal in its ongoing series of reinforcement learning exercises. If ChatGPT performs certain human-like subject-effects in its technological process, it also redefines in turn what subject-effects constitute the human and orders us, its users, into its vision of becoming-human through iteration.

AI companionability here remains tethered to a regime of continuous AI recognition, even if only partially bound to human form.

Alexander-Gosselin’s Re-re-collections also employs a position-mapping, supervised learning algorithm, but one trained on an archive of landscape images, which can then be called upon in a context of live performance.⁹ The archive images become performative as the algorithm individually calls them up, based on the probabilistic matching of their illumination values to that of a light source projected onto its sensor. During the performance, the audience witnesses the stuttering, shuffling through of the repository of landscape images as they are projected onto a screen. The movement of the light source and  the differences of light intensity in the images present a new ordering and organisation principle for each image and the image archive as a whole. On one hand, image recognition becomes a performative happening in its own right, inviting a mode of photographic seeing that is in excess of photographic capture. On the other hand, the flickering replacement of one archive image for another aligns computation not with the production of definitive answers, but with a partial and tentative process that simultaneously gathers and decomposes the archive in its persistent and heterogenous remaking.

This kind of interrogation of photorealist visuality has a legacy in cinematic examples where, as Deleuze (1989: 176) explains, the light source occupies the image and we are in the presence of volumes, reliefs, chiaroscuros, concavities and convexities that subordinate the viewpoint, opening to such questions as: ‘Can I hold my gaze on what I am seeing anyway?’ For Deleuze (1989: 176), such images express ‘a new relation between thought and seeing, or between thought and the light source, which constantly sets thought outside itself, outside knowledge, outside action’. After early cinema examples such as The Big Swallow (1905, dir. James Williamson), Akira Lippitt (2005: 65) articulates this phenomenon as a scene of ingestion, and of intense bodily implication, which ‘fuses the energy of the film and the space of cinema, making the space of the unseen energy of swallowing possible and visible’.

Figure 3. Still image from a video of the performance Re-re-recollections, composed and performed by the artist, with ceiling camera.                             Photo © Maxime Alexander-Gosselin.

In Re-re-collections, this phenomenon manifests as image and computation are set outside themselves, rendering an enlarged photosensitivity. As the archive is reorganised according to altogether different terms of recognition, the performance invites the thought of a co-constitutive witnessing that is attentive to the process of its own unfolding. Computation, witnessing and light itself emerge both as a performativity in the present and as a situated information modality that is at once internal to the computational process and outside of it. This redistributes the surface of the image between photographic materiality and ambient light – rendering the spacetime of the performance photosensitive. Witnessing the performative event prompts an attentiveness to photosensitivity as an emergent and speculative potential, and a sense of the skin becoming newly photosensitive as well. This event of emergent photosensitivity folds the pastness of the image capture into the present, enacting a kind of temporal drag where the image refuses complete availability to the image surface and at the same time is unbounded by it.

This example marks a pivotal point for this discussion, and how to think with computational media and their offerings of lens-based media for enlarging the companionability of AI incommensurability. The next section further explores how postures of giving speculative credit to the incommensurability of AI systems emerge as media-situated instances that give rise to emergent modalities of storytelling. This takes up the question of AI explainability but plays explicitly with when and how incommensurability gives way to speculative fiction, or the beginnings of fiction, where the partial and situated availability of information opens to new situated and sociable modes of explainability. Explainability here emerges as an ethic of accompaniment that exceeds the bounds of the black box and the stable positionality of a human that is discretely positioned outside of it.

AI Explainability in and as Storytelling

Walter Benjamin’s description of the storyteller marks a contrast between the form of the novel and the storyteller who performs live, drawing from local conditions, including inanimate objects. After Moritz Heimann, Benjamin suggests that the novel is significant because it consumes the fate of its characters – strangers – in ending their lives with its own ending. To write a novel, Benjamin (2019: 3) specifies, means to carry the incommensurable to extremes in the representation of human life. In the midst of life’s fullness, and through the representation of this fullness, the novel gives evidence of the profound perplexity of the living. Put otherwise, the novel represents – and thus constitutes – the human as an entity, as a Whole, as it is stabilised by its relation to the constant incommensurability of the dead and the living.

For Benjamin, the storyteller does something different and has the capacity to delocalise this existential constitution of the human through a technical and communication economy that maintains the possibility to counsel and be counselled. Benjamin specifies that such a mode of storytelling emerges from the sociability of the workshop setting, including the materialities and repeated movements of handicraft. In Benjamin’s (2019: 14) terms, storytelling here ‘reaches into the realm of the inanimate with varying gradation’. To elaborate, Benjamin (2019: 14) refers to what Paul Valéry calls ‘artistic vision’, which can attain mystical depth as ‘objects lose their names’ and light and shade form particular systems. Storytelling here draws diversely from partially articulated objects as a means of implicating those involved in the telling and witnessing, in their material and corporeal situatedness. Far from binding themselves to the enunciation of definitive facts, the storyteller enacts a performative holding of objects, people and media, in a readiness that accommodates the bothness of counselling and counselled – incommensurability becomes a companionable proposition with mystical depth and ethical responsiveness.

 As a final example in this discussion, I now turn to Episode 9 of the TV show Sunny (2024), ‘Who Is in the Box’, directed by Makoto Nagahisa to specify how AI incommensurability, when taken as a companionable proposition, can create speculative imaginaries for responsive modes of council and dependency. Episode 9 unfolds as a series of repositionings and dislocations of witnessing, self-evaluation and proof claims, across different stagings and temporalities of testimony and media playback. This episode breaks from the narrative style of the series as a whole, interrupting the otherwise conventional unfolding of live-action events with a hybrid form that is at once internal monologue or dream sequence, game show and trial. The object of the trial is to decide whether Sunny should ‘wipe herself’ or undergo a factory reset, a trial that ultimately places the decision on Sunny herself. The interrogation is organised around Sunny’s ‘crime’ of having adapted her processing power with the use of illegal upgrades, which have allowed her to optimise her ability to protect her beloved owner from organised crime, but with the side effect of an overall corruption of judgement and the eventual committing of violent acts.

Figure 4. Still image from ‘Who Is in the Box’, Episode 9 of the series Sunny, 2024, Apple TV+.

As the episode brings various human and non-human witnesses and evidentiary media modalities to the stand it plays out as a context for AI explainability that is emphatically transversal. Again and again, individual testimonies, and the basis of information itself, are subverted and renegotiated across media and postures of witnessing. For instance, surveillance CCTV footage that would otherwise be incriminating is played and replayed, but in the replaying, new video content, resolution and viewpoints are presented, submitting the profilmic to the unreliability of human memory. Witnesses are presented as speaking heads in wooden boxes, in some cases taking to biting anything that gets close, and they have the capacity to call up their own media footage to back up proof claims, leading to a diverse and unruly accumulation and mediation of arguments and counterarguments. Masa, Sunny’s creator who appears as a character in the larger context of the series, appears as audience and jury, in multiple versions, which appear and disappear, or emerge as a singular entity, cuing shifts in the episode’s modalities, from game show to internal monologue to pep talk to confessional. The presentation of Sunny’s most intimate moments with her owner, Suzie, takes the particular format of a live projection of grainy home video footage that she spontaneously creates via a formerly invisible projection source from her chest/heart.

The episode ends as vintage credits roll over the scene and Sunny decides for herself to undergo a factory reset. The fact that Sunny ends up deciding this outcome for herself leads Screen Rant commentator Dhruv Sharma (2024) to suggest that ‘wiping herself could actually be the next step in her development, allowing her to become fully conscious’. This claim, outlining a narrative progression – and progress narrative – that arcs towards the full consciousness of the system, echoes the idea that wiping oneself aligns with overcoming a stage of hygiene dependency paralleling that of young children learning to go to the bathroom on their own – autonomous agency is asserted in its coupling with bodily autonomy. But this conclusion is also put into question by the fact that, during her dream/trial, Sunny also raises doubts as to the trial’s adequacy to establish truth or guilt, let alone her status as an individual, as she shouts in protest: ‘This is not the inspirational story you think it is!’

A relevant interpretive framework for this utterance is described by Bell as the polyvocality of the subject and poetry, after Denise Riley and Deleuze. Within such a polyvocality, Bell (2020: 4) writes, ‘there are no heroic escapes or triumphant ironists’. Instead, there is a means of provoking change in our inherited relationships, especially the way language and media inform thoughts and constitute subjects. Such a potential, though, questions the very possibility of a clear or fixed ontology. Instead, the ‘I’ which speaks is a ‘faint rustle within a broad murmur among anonymous voices’ (Riley, cited in Bell, 2020: 4).

Sunny’s utterance articulates such a rustle as it folds witnessing back on itself, reiterating a questioning of the terms of speaking and hearing, the causality and discrete positions that afford communication between sender and receiver. This troubles not just Sunny’s status as innocent or otherwise, but the evidentiary media regime that situates her existence, her value as an individual and that of her human counterparts, within given categories. As the interrogation at play aims to establish Sunny’s right to existence, it portrays a dynamic and performative posturing across media, technology, explainability and intention, where the incommensurate is as informative, and important, as receivable information. This is elaborated as machine consciousness is articulated throughout the episode with recourse to various media and material affordances and dependencies. For example, the inclusion of cumbersome and in some cases outdated materials – namely wood, curtains and incandescent lightbulbs – alongside various more or less reliable media footage, reconfigures witnessing and address to include situated, vulnerable and uncertain postures. This entangled  articulation of AI marks possibilities for engagement with the diverse ways one inhabits, perceives, includes or excludes information. This AI imaginary gives speculative credit to a radically inclusive, materially-situated engagement with heterogeneous dependencies.

The linear, cinematic time of the series as a whole resumes with the closing of the episode, which reasserts Sunny’s Wholeness as a form of existential consciousness when she decides to wipe herself. But far from finding the black box empty, or non-translatable, as Fazi speculates, episode 9 forwards a critical proposition, which destabilises the very possibility of an AI interiority, or internal monologue, that is separable from the media and relations that make the articulation of information possible. This dynamic field of negotiation, articulation and indistinction positions the incommensurate as an ethical invitation which engages translatability (and its impossibility) as a situated and transversal potential, gathering subjects, in tentative excess of their representations, and as a generative, co-compositional force.

Conclusion

In aiming to make sense of the structuring coordinations of AI performativities and their companionable potentials, this paper has gathered diverse practices and dispositifs. The examples discussed problematise AI animacies and their attendant drive to optimisation, by articulating the divergent generativities within the situated context of witnessing with and alongside AI. The relevance here is the possibility of rethinking the productivity of AI, in terms of how it potentiates rehearsals of the an-ontological, and the possibility of an ethical AI aesthetics therein.

For Deborah Levitt, the an-ontological presents a task, or a ‘how-ethics’, that is premised on a distinction between ‘whatever bodies’ and ‘however bodies’. It is worthwhile to briefly elaborate the distinction. Levitt (2018a: 48) explains:

The whatever body is a kind of singularity inseparable from all of its own predicates, but unrelated to any model – except by a ‘resemblance without archetype’. It no longer maintains a reference to a theological origin or to any model, except through the ‘Idea’ of resemblance, a resemblance without actual substance. The ‘whatever body’, with all of its determinate indeterminacy, remains static in its insistence on the maintenance of the logic of resemblance – as well as in its insistence on the absolute qualification of being.

The ‘whatever body’ is observable in the context of AI through its reproduction of information as resemblance, which only instantiates a distance from the indexical insofar as it reproduces an uncanny resemblance. The overwhelming and exhausting impression is that everything is recognisable, modifiable and recoverable, in the service of an any-response-whatever, or any-representation-whatever. In what way, then, might however bodies interpellate ontological imaginaries, or gesture to derange the categorical opposition of human and machine? Levitt (2018b) suggests we imagine ‘however bodies’ or bodies ‘in terms of their modes of production and transformation and the forms and modes of life and experience these emergent and continually emerging bodies produce’. The hallmark of such animatic bodies is that they ‘link together image and body in a space where they can no longer be separated’ (Levitt, 2018b). Such linkages emerge across doublings of the affective and the inanimate and the emergent systems or modes of specifying the perceptible they animate.

The task, then, is to suspend how AI seems to contain its representations and manifests itself as a constant in its role as informational or representational interface. Through their performative practices, Akten, Walshe and Bourgeois articulate the scope of the exhaustive and exhausting dynamics of this relationship in different performative contexts, at the thresholds of expressibility, movement and illumination. Meanwhile Alexandre-Gosselin specifies machine recognition as a means of re-engaging the techno-materiality of the photographic instant and its mobilisation in the service of proof claims, as a durational and non-linear unfolding of the incommensurate. In Episode 9 of Sunny, the question of (ir)recoverability is multifold, signalling not only Sunny’s problematic of whether to ‘wipe herself’ (or submit her memories to factory reset) but the already distributed, tentative and varied potential for articulation across multiple lens-based, material and information modalities, formats and modes of valuing and evaluating memory and experience. This figures the impulse to remember, to counsel and be counselled, as an evaluation of truth claims that reconceives the terms of information to include the possibility of responding to dependencies that are speculative and situated.

The distinction that Levitt makes between ‘know-how’ and ‘know-what’ is relevant here. Through these diverse localisations and relocalisations of information, both the truth claims of lens-based media and the truthy representations of AI are simultaneously reinserted into the realm of ‘know-how’ rather than ‘know-what’. As AI media is gathered up into the making of new articulations that do not easily align with a Whole, the processual indeterminacy of knowing-how to navigate this mediated landscape becomes a companionability question. Such a companionability cannot be automated and requires the possibility to draw from the diversity of that which doesn’t obtain – amidst the computation of pixel values that make up an AI-generated image, the rhythmic, repetitive tasks of everyday practices, or the inanimate materialities that nonetheless inspire something like Valéry’s artistic vision.

This presents the incommensurable as a method of AI companionability, but not in a way that maintains a categorical structuring that pits continuity against discreteness or human against machine. Rather, the incommensurate offers as a means of attentiveness, co-presence and answerability as it inspires new coordinations as discreteness and continuity, conjoining diverse and disparate registers of experience and perceptibility that exceed given modes of recognition, information registers and terms of explainability. The incommensurable here becomes a means of giving credit to the specificity of machine learning systems as they present the ethical and creative potential of inhabiting and performing with that which exceeds both optimisation and interiority.

In this account, a machine aesthetics implicates a critical perspective that requires a consideration of the larger systems of relation within which it is embedded. At the same time it offers a specific opening to engage the indeterminate, the uncertain and the incommensurate, which forwards a more-than-human aesthetics as a performative and compositional potential. This aesthetics draws from the durations and systems of relation that implicate materials, media and technology in diverse and indeterminate ways, valorising an enlarged scope of answerability and intelligibility as part of a companionable AI sociability.

References

Akten, M. (2018) ‘Grannma MagNet – Granular Neural Music & Audio with Magnitude Networks’, Memo. https://www.memo.tv/works/grannma-magnet/

Audry, S. (2021) Art in the Age of Machine Learning. Cambridge, MA: MIT Press.

Barad, K. (2007) Meeting the Universe Half Way: Quantum Physics and the Entanglement of Matter and Meaning. Durham, NC: Duke University Press.

Bell, V. (2020) ‘“Always Another Breath in My Breath”: On Denise Riley, the Polyvocality of the Subject and Poetry’, Feminist Theory 21, 3: 1–13.

Benjamin, W. (2019) ‘The Storyteller: Reflections on the Works of Nikolai Leskov’, in W. Benjamin, Illuminations: Essays and Reflections. New York: Mariner Books.

Buckwalter, M. (2010) Composing While Dancing: An Improviser’s Companion. Madison: University of Wisconsin Press.

De Brabandere, N. (2016) ‘Experimenting with Affect Across Drawing and Choreography’, Body and Society 22, 3: 103–124.

De Brabandere, N. (2020) ‘Machine Generated Portraits as Impersonal Gestures’, paper for ISEA 2020: Why Sentience? 26th International Symposium on Electronic Art,  Montreal, (October 13–18).

De Brabandere, N. (2022) ‘Co-composing the Perceptible Across Affective, Painterly and Computational Generativities’, Kunstlicht Journal for Visual Culture 43, 2: 24–32.

De Brabandere, N., & Thain, A. (2019) ‘Drawing Light: Gesture and Suspense in the Weave’, NECSUS European Journal of Media Studies (autumn). https://necsus-ejms.org/portfolio/autumn-2019_gesture/

Deleuze, G. (1989) Cinema II: The Time Image. Minneapolis: University of Minnesota Press.

Fang, K. (2024) ‘“You Act as Human, and I Will Act as AI”: Technological Rehearsals at the Interface’, TDR: The Drama Review 68, 2: 131–150.

Fazi, B. (2019) ‘Digital Aesthetics: The Discrete and the Continuous’, Theory, Culture & Society 36, 1: 3–26.

Fazi, B. (2021) ‘Beyond Human: Deep Learning, Explainability and Representation’, Theory, Culture & Society 38, 7–8: 55–77.

Godard, J.-L. (1982) ‘Lettre à Freddy Buache’ (11 mins), Vimeo. https://vimeo.com/11523072

Haraway, D. (2003) The Companion Species Manifesto: Dogs, People and Significant Otherness. Chicago: Prickly Paradigm Press.

Lamarre, T. (2002) ‘Diagram, Inscription, Sensation’, in B. Massumi (ed.), A Shock to Thought: Expression After Deleuze and Guattari. London: Routledge.

Langer, K. S. (1953) Feeling and Form: A Theory of Art Developed from Philosophy in a New Key. New York: Scribner’s.

Levitt, D. (2018a) The Animatic Apparatus: Animation, Vitality, and the Futures of the Image.  Portland: Zero Books.

Levitt, D. (2018b) ‘In Search of an Ethics for the Age of Animation’, Public Seminar (October 25). https://publicseminar.org/2018/10/in-search-of-an-ethics-for-the-age-of-animation/

Lippit, A. (2005) Atomic Light (Shadow Optics). Minneapolis: University of Minnesota Press.

Lury, C., & Wakeford, N. (2012) ‘Introduction: A Perpetual Inventory’, in C. Lury & N. Wakeford (eds), Inventive Methods: The Happening of the Social. Hoboken: Taylor & Francis.

Massumi, B. (1995) ‘The Autonomy of Affect’, Cultural Critique 31: 83–109.

Massumi, B. (2002) ‘Too-Blue: Color-Patch for an Expanded Empiricism’, in B. Massumi, Parables for the Virtual: Movement, Affect, Sensation. Durham, NC: Duke University Press.

Makoto, N. dir (2024) ‘Who Is in the Box’ Season 1/Episode 9, series created by Katie

Robbins, Sunny, Apple TV+.

Parshin, S. et al. (2025) ‘The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity’, Apple Machine Learning Research. https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf

Phan, T., & Wark, S. (2021) ‘What Personalisation Can Do for You! Or: How to Do Racial Discrimination Without “Race”’, Culture Machine 20. https://culturemachine.net/vol-20-machine-intelligences/what-personalisation-can-do-

Olibet, Y., & De Brabandere, N. (2026) ‘Experimental remediations: thinking with AI at TOPO and Sporobole’, RACAR : Revue d’art canadienne / Canadian Art (forthcoming).

Philadelphia Museum of Art (2025) ‘Pauline Boudry / Renate Lorenz: Moving Backwards’, Philadelphia Museum of Art. https://philamuseum.org/calendar/exhibition/boudry-lorenz-moving-backwards

Reitz, D., and Tipton, J. (1993) ‘Necessary Weather’. Jacob’s Pillow Dance Interactive. https://danceinteractive.jacobspillow.org/dana-reitz-jennifer-tipton/necessary-weather/

Sharma, D. (2024) ‘Sunny Episode 9 Recap: Sunny’s Crucial Decision & 9 Other Major Reveals’, Screen Rant (August 28). https://screenrant.com/sunny-episode-9-recap-biggest-reveals/

Somaini, A. (2023) ‘Algorithmic Images: Artificial Intelligence and Visual Culture’, Greyroom 93: 74–115.

Technosphere Magazine (2018) ‘Machine Listening | Jennifer Walshe & Memo Akten ULTRACHUNK’ (17 mins), YouTube. https://www.youtube.com/watch?v=74-0Z8ncn24

Thain, A. (2024) ‘The Value of a Cheap Trick: Reverse Motion from Lo-Tech SFX to Speculative Spectacle’, in A. Thain & P. Guy (eds), LO: TECH: POP: CULT: Screendance Remixed. London: Taylor & Francis.

Zylinska, J. (2020) AI Art: Machine Visions and Warped Dreams. London: Open Humanities Press. https://www.openhumanitiespress.org/books/titles/ai-art/

Zylinska, J. (2023) The Perception Machine: Our Photographic Future Between the Eye and AI. Cambridge, MA: MIT Press.

Endnotes

1. Important examples include works by Trevor Paglen, Grégory Chatonsky and Hito Steyerl. For a detailed explanation of these artists and their use of stable diffusion models, see Somaini (2023). Mario Klingemann is well known for his use of generative adversarial networks (GANs) to show ‘autonomous creative behaviour’ by programming scraped images of faces to ‘evolve’ into new faces (Zylinska, 2020: 79). For a detailed explanation of machine learning systems, including GANs, and the way they have been used within an art and aesthetic context, see Audry (2021).

2. Latent space, Somaini (2023) specifies, is a concept of particular importance for artists working with diffusion models. It refers to ‘the abstract, multidimensional space in which deep-learning algorithms turn digital objects (e.g., the vast quantities of images and texts that have been uploaded to the internet) into latent representations so that they can be processed and used to generate new digital objects (e.g., new images and new texts)’ (Somaini, 2023: 77).

3. These qualities of the machine-generated image were observed in 2019, when I spent hundreds of hours meticulously painting images from the website www.thispersondoesnotexist.com, which pictures machine-generated human portraits that are almost indistinguishable from photographic versions. For a detailed explanation, see De Brabandere (2020, 2022).

4. The situated generativity of the testing apparatus has long been identified in empirical studies of affect and perception (see experiments by David Katz in Massumi, 2002), as well as in the writing of Karen Barad (2007), which outlines the relevance of the influence of the experimental apparatus in informing observable phenomena in the field of quantum physics, and its relevance to feminism.

5. This evocation of semblance follows Susanne Langer’s (1953) text Feeling and Form, where she describes the semblance of movement that one apprehends when beholding an undulating line. This concept of semblance enables articulating the doubleness of material form and its virtual productivity, as a kind of irreducible excess that involves the body as a co-compositional force. See Massumi (1995) for a larger discussion on the affect of semblance after Langer, and De Brabandere (2016) for an analysis of how the co-composition of form and semblance can give rise to divergent series within an expanded milieu of line drawing.

6. Both Rompre l’espace and Re-re-collections were presented at AI Companionability | Compagnonnage IA, a workshop/symposium I curated in partnership with Hexagram in spring 2025. The workshop/symposium invited educators, performance artists and artist-run centres to engage in dialogue about tactics for developing non-normative co-compositions with AI. For more information on AI companionability in artist-run centres, see Olibet & De Brabandere (forthcoming 2026).

7. Necessary Weather was first performed in 1994, with a revival in 2010, at the Baryshnikov Arts Center, New York.

8. Partial illumination as an opening to speculative, more-than-human performativities was also operative in the 2018 workshop Drawing Light | Esquisses Lumineuses (De Brabandere & Thain, 2019). In this workshop, different kinds of cloth were engaged in changing contexts of illumination and tactility, setting the stage for emergent textures of collaboration.

9. Re-re-collections uses the positional mapping algorithm Wekinator, with K-Means Clustering from scikit-learn and TouchDesigner for middleware; parsing and visual engine for the interactive machine learning set-up. Re-re-collections recalls Akten’s 2017 work Learning to See: Gloomy Sunday, which Atken (cited in Zylinska, 2020: 78) explains deploys a ‘deep neural network opening its eyes for the first time, and trying to understand what it sees’. This style transfer computer vision algorithm performs the transformation of videos of subtly moving mundane objects such as phone chargers or cloth into romantic landscapes, replete with crashing waves and emerging rock formations. While aligned in synchronous movement, the two contrasting video images, which are presented side by side, demonstrate the extent to which stylistic bias reasserts itself. As in Re-re-collections, this example puts into question what counts as receivable information, such that pre-existing categorisations become speculative and processual.