Because the twin is not a copy: On the politics of digital twins

Abstract

This essay explores the changing form of the digital twin as a political technology in the age of deep learning and generative artificial intelligence (AI). It situates the digital twin as a distinctive contemporary form of simulation and examines how it is making up people, things, scenes and their interactions in novel ways: the drawing in of unstructured datastreams and their experimental recombining and modification. Although the digital twin is often represented as a digital copy or replica of the physical world, it actually departs radically from a mimetic copy; advancing a mode of twinning that classifies, extracts and divides. The essay addresses digital twinning as an emerging form of knowledge and action in the world. It proposes three distinctive logics of the contemporary politics of digital twins: intervention, prediction and action. Each of these aspects is discussed through the situated digital twin domains of the factory, the clinic and the battlefield.

‘“I now understand why you’ve asked me, at every step, to observe and learn Josie [. . .] I’ll use everything I’ve learned to train the new Josie up there to be as much like the former one as possible” [. . .] “Klara, we’re not asking you to train the new Josie. We’re asking you to become her. We want you to inhabit that Josie with everything you’ve learned [. . .] Learn her till there’s no difference between the first Josie and the second”’

(Ishiguro, Klara and the Sun, 2021: 209).

Introduction: who is ‘Riley’?

In April 2022, the Stanford human centred artificial intelligence (HAI) group hosted a conference on what they famously termed ‘foundation models’. The gathered computer scientists discussed how generative pre-trained transformer (GPT) or large language models (LLMs) would ‘change the paradigm in which artificial intelligence (AI) systems are built’ (Liang, 2022). This paradigmatic shift describes a transformation from the use of broadly structured and labelled data for building bespoke models for specific domains to the use of multi-modal unstructured data for building generalisable models for ‘discovery of what it can do’ (Liang, 2022). In the discussion that followed, the co-founder of OpenAI, Ilya Sutskever, addressed an example of this paradigm shift to generative AI and its discovery-based approach. He described how The Trevor Project – a US non-profit organisation providing counselling in support of the mental health of young LGBTQ+ people – was using GPT2 to simulate the interactions between young people and their counsellors. Sutskever sought to position this use of GPT2 as an example of how generative AI could be used for the wider benefit of humanity. ‘The Trevor Project is a safe issue’, he claimed, ‘because it is just training suicide hotline staff to talk to real disturbed teenagers’ (Sutskever, 2022). His deeply problematic use of ‘disturbed’ in relation to mental health and LGBTQ+ discrimination in fact dramatises the impossibility of an ethics of AI models that designates a ‘safe issue’ from a ‘dangerous’ one in the way that he seeks (Amoore, 2020: 79). His description of the digital twin of a teenager that was built on GPT2 – ‘Riley’ – mirrors precisely a widespread sense that digital twins are safe spaces to intervene in advance of action in the real world.

So, who is Riley and in what sense is Riley a digital twin? Open AI describes Riley as an ‘AI based persona’ with whom the Trevor Project human counsellors would trial and demo their responses to their teenage LGBTQ+ clients (Sutskever, 2022). Of course, the literature on computer models in urban simulation, policy and climate sciences reminds us that the role of the ‘demo’ or the ‘trial’ is never simply to test in advance but rather to make and delimit the parameter of a world as such (Amoore, 2023; Edwards, 2010; Halpern, 2015). The designers describe Riley as ‘mimicking a teen in North Carolina who feels anxious and depressed’ and as having been trained on 45 million pages from the web as well as ‘many hundreds of past role-plays’ (Ohlheiser and Hao, 2021). Though Riley is discursively framed as ‘twinned’ in the sense of mimicry, or an exact replica or copy of some other person or thing, in fact she is closer to what Annemarie Mol (2002) calls the ‘body multiple’ – made to cohere from a vast and heterogeneous multiplicity. Riley is a body multiple in the sense that she iteratively and intimately builds a world from her interaction with other people, objects, data, texts and images. The generative model on which Riley is built has been trained on large linguistic corpora that confer a general sense of language, grammar and syntactical meaning. The model is then fine-tuned using the language data from multiple actual and synthetic interactions with real Trevor Project teenagers and counsellors and with Riley. It is for this reason that the statement that Riley is a ‘safe issue’ can never be upheld, for Riley is actively building models of what safe, dangerous, risky, or harmful mean in the world. The very meaning of ‘depression, ‘anxiety’ or ‘suicide’ – and crucially how future people will respond to these – is significantly generated through the multiplicity that is the Riley digital twin. It is in this specific sense that the OpenAI digital twin was not a ‘copy’ of the world but was engaged in worldmaking as it traversed the underlying text datasets of transcribed past real and virtual conversations.

The case of Riley – a digital twin built to model suicide risk in a vulnerable community – has stayed with me because of what it suggests about the rise of our contemporary form of the digital twin, and its politics and ethics. Though, at first glance, Riley may not appear to fit the conventions of the digital twin as it manifests in spaces such as the smart city or urban operating system (Halpern and Mitchell, 2023; Luque-Ayala and Marvin, 2020), Riley illustrates precisely the emerging form of the digital twin as it becomes shaped by generative AI and the unstructured inputs of text, images, video and audio data. Historical antecedents of the digital twin in urban, military, and environmental domains sought rules-based systems capable of controlling and predicting uncertainty (Daston, 2022). In the absence of rules for detecting risk, Riley holds out the promise of discovering the significant elements, from the speech of an audio recording to the text of a client’s medical notes. Riley captures the contemporary digital twin, then, where the appearance of a replica persona or digital double is actually built on a data multiplicity comprising past teenagers (real and simulated) and the corpora of an underlying LLM.

My focus in this essay is on the changing form of the digital twin as a political technology in the age of deep learning and generative AI. I will begin by reflecting on the question of the specific form of simulation we may be witnessing with the contemporary digital twin. Though a detailed genealogy of simulation is beyond the scope of the article, the digital twin is elaborated here in terms of its paradigmatic break from previous forms of simulation. I focus my discussion on three examples of digital twins deployed in the domains of the factory, the clinic, and the battlefield. This concept of domain is not at all straightforward in relation to the rise of the digital twin. While the organising ‘logic of domains’ has characterised the specificity of ‘wordly action or knowledge’ – for example, biology as the domain of the life sciences or law as the domain of juridical knowledge – computational knowledge from at least the 1990s has promised methods that are increasingly ‘domain general’ or ‘domain agnostic’ (Ribes et al., 2019: 282, 284). To be clear, it is not my argument that the digital twin spaces of factory, clinic, and battlefield are somehow emblematic of the domains of production, health, and warfare that constitute the spaces of this political technology. Indeed, with the rise of generative models and cloud storage and analysis, the development of digital twins in each of these spaces combines ‘domain agnostic’ language models with data fine-tuning for domain specificity.

Though the three spaces of the digital twin are not merely emblematic domains for the study of new technological forms, the contemporary narrative of digital twins pays less attention to the first use of the term ‘digital twin’ in cardiology in 1994 than it does to some other more directly observable sites such as urban platforms (Renaudin et al., 1994). My treatment of the three spaces is intended to afford analysis of how a political technology materialises; its form of knowledge and power. The philosopher Gilles Deleuze (2010) reflects on Foucault’s spatial foci as ‘machines’, reminding us that the ‘machine prison, machine-school, the machine-hospital’ are ‘social before being technical’ (p. 34). Understood in this way, the digitally twinned spaces of factory, clinic and battlefield are always already social and political technologies. Although they have been significant spaces of experimentation and arguably share a common infrastructure of general models and cloud compute – as well as combining and coalescing their data as they adapt across domains – I separate them artificially here for the purposes of analysis and discussion. I will then set out three distinctive aspects of the contemporary politics of digital twins: intervention, prediction, and action These three defining features of the logic of the contemporary digital twin are present in each of the spaces I discuss, but here I consider them in turn so that we might hold them in our minds together and consider how they come to know and act in the world.

Of digital twins and simulation

The rise of the digital twin is situated in a long and deep history of scientific representation, with significant origins in ideas of objectivity, visualisation, and statistics (Galison, 1997; Hacking, 1990). The contemporary discourse and practice of the digital twin – extending across computer science and the commercial and governmental pursuit of novel simulations – is itself a distinct and powerful force that shapes the world. In current computer science accounts, the digital twin is conventionally described as ‘a live copy of physical systems’, captured in terms of its ‘liveness’, its ontology as a ‘copy’ of the physical world, its ‘accuracy’ and its ‘synchronicity’ with a data environment (Mashaly, 2021: 299). One of the most widely used digital twin platforms, Nvidia’s Omniverse, for example, describes its digital twin as a ‘large scale, physically accurate simulation of an asset, process, or environment with multiple autonomous systems perfectly synchronized with real-world data streams’ (Nvidia, 2023). The novelty of the digital twin is thus said to lie in its capacity to integrate multiple structured and unstructured data sources from a physical environment outside the model, creating an accurate copy of the target. Yet, in many ways the digital twin is definitively not a copy in the sense of a mimetic representation of the physical world, nor does it aspire to accuracy so much as to a precision that affords action (Daston, 1995). As one mathematician and developer of digital twin models puts it: ‘this idea of an accurate picture is not true in practice. Identical digital twins are not possible and they are not necessary, the question is how identical does it need to be? ¹ This question of how identical it needs to be points precisely to Daston’s insight that the history of science is characterised less by the search for complete accuracy than for precision. What the digital twin actually ‘wants’, then, is sufficient precision to make decisions and actions possible.

Understood in this way, the digital twin loosens the idea of a model that is a complete likeness or a copy. For example, it is said of University College London’s SmartSurgery digital twin of the heart that: ‘we don’t need to simulate the whole heart [. . .] we simply need to capture specific features that are relevant’ (Lamata et al., 2019). This notion of capturing specific features that are said to matter – contrary to an accurate copy or replica – illustrates how the digital twin is a boundary-making or attention machine that has a politics in the sense that it foregrounds and brings to attention some features of the world and not others (Barad, 2007: 148; Pedersen et al., 2021). Contra the discursive framing of the digital twin as a copy or replica of a physical object or space, in this essay I propose an alternative sense of twinning as political technology. Where digital twins capture some specific features and discard others, their ‘twinning’ is closer to the old etymology of the verb ‘to twin’ – ‘to part, to divide, separate from or estrange’ – than it is to an identical likeness, suggesting the digital twin is a feature space that decomposes and recomposes some elements of a teeming data environment.

What is at stake in these two apparently competing imaginaries of digital twins: as live and accurate copies of the world or as making the world via the extraction and division of a feature space? By way of illustration, both imaginaries of the digital twin – mirroring what Peter Galison (1997: 97) terms the ‘mimetic’ and the ‘analytic’ approaches to scientific experimentation – are developed in Kazuo Ishiguro’s (2021) novel, Klara and the Sun. The human adults in the novel commission a digital twin of their daughter, Josie. She sits for a ‘portrait’ as though a complete likeness of her singular being can be captured in a digital twin of her characteristics and traits. The desire to build a digital portrait of Josie echoes OpenAI’s claims that Riley is a digital ‘persona’ that has the definitive characteristics of a ‘troubled teenager’ (later in Ishiguro’s novel the reader learns that Josie’s digital twin is a kind of insurance policy against the finitude of life, following the death of her sister from the complications of being ‘lifted’ or digitally augmented). The novel points to the impossibility of a mimetic digital twin that could ever capture the singularity and irreplaceability of Josie’s life.

Meanwhile, a digital twin in the second mode – an analytic feature space – is being built by Josie’s robot artificial friend, and the novel’s narrator, Klara. Klara builds her twinned algorithmic model of the social world through inductive observation of multiple sources of unstructured and unlabeled data. Klara learns from the multi-modal data features of her environment: images, sounds, facial gestures, the complex interactions between the humans that often confound her model in their departure from the rules or axioms of parenting, causing her to adjust and update her assumptions. In her role as Josie’s carer and friend, Klara is hyper-attentive to the relations of data, detecting and extracting features that are otherwise overlooked and neglected by the adult humans. ‘I began to gather together their various remarks into a coherent observation’, Klara narrates, and ‘I came to understand that the plan wasn’t anything they’d built carefully, but more of a vague wish connected to their future’ (Ishiguro, 2021: 122). Klara’s neural network model of the world is not a twin in the sense of a mimetic copy or complete likeness of the world but in the more specific machine learning sense of approximating the best fit or function between data inputs and outputs. What matters is not having an accurate replica of the world but a precise or ‘good enough’ (in the computer science sense) model to make knowledge and action possible. In contrast to the adult humans’ desires for Josie’s digital twin or data double, Klara’s making of a whole from observation, like Riley’s, is irreducible to the individual data elements, greater than the sum of its component parts (DeLanda, 2011).

What we are witnessing with digital twins in the age of AI, then, has a different orientation to simulation as a space of intervention, prediction, and action. It is to each of these three elements of the contemporary logic of the digital twin – and the three illustrative spaces – that I now turn. The analysis is informed by Nick Seaver’s (2017) ethnographic approach to ‘algorithms as culture’, where he notes that the ‘’correct’ definition of algorithms has been used precisely to isolate them from the concerns of social scientists and humanists’, whereas ‘if we look to the places where algorithms are made, we may not find a singular and correct sense’. Looking to the places where digital twins are made, methodologically I read closely the computer science literature as a political and cultural arrangement of propositions and not merely as a technical or definitional account of the digital twin (Amoore et al., 2023). This method of reading the places where models are made is followed by a series of field observations of Nvidia demos and training courses for digital twin developers. Such observations allow the researcher to analyse the contingent promises and claims made for the Omniverse digital twin platform and to engage with its practical implementation in different spatial domains. Finally on method, interviews with developers of digital twins afford a more situated sense of the experimental, provisional, and iterative process of building a model.

The factory: digital twins as spaces of intervention

The designers of BMW’s digital twin ‘factory of the future’ explain to an audience of Omniverse clients how they built their model of every human and material process in the factory. ‘We don’t want to do data preparation’, they say, ‘we bring in the unstructured data from the source”’(Nvidia, 2023). Though of course there is no ‘raw data’ as such (Gitelman and Jackson, 2013), the significant point here is that the painstaking process of data preparation (pivotal to machine learning for production processes in the past) is by-passed in favour of a pre-trained model that has estimated the structure of an underlying data distribution. The digital twin of BMW’s Regensburg factory draws multi-modal data (e.g. images, text, numeric, video) from across the space of production, including from the robotic sensors, the kinematics of human workers and robots, the lidar scans of built spaces and structures, sensors, and logistics, and the computer simulations of possible events or accidents. The group describe the simulations of the factory that they had used in the past, prior to the adoption of the Omniverse digital twin platform. The past factory simulations were ‘insufficient’, they say, in that they relied upon statistical models and were ‘locked into data and rules from the past’ that had to be expensively labelled and annotated. By contrast, what is sought in what they call ‘the factory of the future’ is a more open-ended, futures-oriented, and experimental process of discovery (Theurer and Krambergar, 2021).

What is taking place in the digital twin’s experimental logic of intervention? The digital twin appears here to exceed the rules-based formulations of past associative algorithmic models of the factory (where data-mining is used to identify rules of interest), opening what I have elsewhere called a ‘space of play’ that opens onto ‘combinatorial possibilities of malleable and adaptable inputs’ (Amoore, 2020: 78). The designers of the BMW digital twin propose that ‘it is helping us to do some playing ahead of time’ so that combinatorial possibilities are explored in advance of the 2025 BMW plant opening in Hungary (Nvidia, 2023). ‘We are virtually optimising layouts, robots and logistics systems years before the factory opens’, the team explains, so that an intervention can be made in the digital twin and the emergent properties of a changed element can be captured and observed. Consider, for example, a decision that is made about how to optimise a new robot within the existing production process. The intervention will have effects – on human workers, other robots, physical space, logistics, energy use and so on – that are not directly observable in production line data but can be experimentally induced via the digital twin. The team introduce the new robot into the scene of the digital twin, commenting ‘so you can see if the robot would collide with a column in the building’. It is precisely this ‘if’ and ‘would’ speculative and conditional formulation that characterises the logic of the contemporary digital twin: ‘if the robot would’. The digital twin thus intervenes in the relations of the factory – everything, from human and robot bodies to lighting and temperature – in order to play with the possibilities of their arrangement. It is a possibilistic and speculative logic of intervention that not only folds future possibilities into action in the present as we see in logics of pre-emption (Collier and Lakoff, 2015; Cooper, 2011 [2008]; DeGoede, 2012), but that actively brings that action into the present so as to explore future capacities and potentials. Thus, the BMW ‘factory of the future’ digital twin is not strictly a copy or replica of any entity but rather an experimental exercising of a potential future in the present.

What does it mean to intervene in the space of the digital twin? To make an intervention in the space of the digital twin is to govern something in its very unfolding. In this sense the factory’s digital twin mirrors the politics of digital twins in climate and pandemic, where the logic is definitively not one of preventing something from unfolding, nor even necessarily preventing harms, but seeking the minimum possible governing intervention. Indeed, the BMW digital twin models share the same Nvidia Omniverse platform as used by many climate modellers, similarly seeking ‘to repeatedly interact’ and ‘to be able to ask what if scenarios’. ² As an experimental space of play, the digital twin affords a governing by adjustment and experimental ‘tweaking’. This is a deeply empirical, hypothetical and exploratory type of governing and knowledge making. It refigures the spaces of factory, battlefield, or clinic so that they are not governed via statistical regularities or linear-causal deductions. Rather, they are governed as spaces of experimental intervention. As in the case of the BMW Regensburg factory, the digital twin governs the thresholds of an intervention – the boundaries and parameters of what is possible. As Michel Foucault (2008) describes the regime of truth for reconfiguring the ‘art of governing’: ‘And now the problem will be: Am I governing at the border between the too much and the too little’ (p. 19). The digital twin advances this form of governing by malleably and adaptively discovering whether it is governing at the border between too much and too little. Each small adjustment in the model is generative of new patterns of emergent effects. To be clear, the space of intervention of the digital twin is not a radical space for testing an action, and it is definitively not oriented towards preventing or halting something in its tracks. It is a governing of bodies, objects, and things in movement and circulation.

Just as the political technologies of Fordism, Post-Fordism and Toyotism far exceeded the locations of the factories where they were developed, so the digital twin as a space of intervention overflows the site of the factory and governs more broadly the mobilities and circulations of people and things. Consider one final example of the space of intervention – the Deutsche Bahn digital twin of the German rail network – that deploys the same Omniverse platform as BMW’s factory. The model draws upon the geospatial and operational data of Deutsche Bahn’s physical rail infrastructure, but it also extends that dataset by simulating the rare events that would not otherwise be observable in the data. The effects of climate change – wild fires, extreme weather events, landslides – are simulated in the digital twin of in order to explore the space of intervention. ‘Rare scenarios can be modelled on demand’, explain the designers of the digital twin, ‘without putting anyone at real risk’ (Minguez, 2023). This is a significant discourse of digital twins – that they represent safe spaces of intervention in advance of or in place of the risks of a real event. For example, the Deutsche Bahn digital twin simulates people and things falling from platforms and onto tracks, ‘bridging the gap in the dataset’ of rare but disruptive events. As the designers of the system explain, ‘the model needs examples of these types of scenario in order to learn how to respond appropriately’ to a real event (Minguez, 2023). In an age of generative AI, where the rare event is synthetically generated by other algorithms, the digital twin is actively harnessing the emergent properties of the model in order to govern uncertainty (Jacobsen, 2023). The promise is to draw out the emergent properties: to detect a human, to anticipate a fall or a jump, to position a robot – but always to sustain production, economies, movement and circulation.

The clinic: digital twins and prediction

There is a vast body of work on the histories of probability and changing forms of statistical knowledge involved in predicting things (Daston, 1988; Hacking, 1990; Porter, 1995). What this work tells us is that the epistemology of probabilistic reasoning – and its impact on predictive models of society – has changed significantly over time. When critical responses to the current rise of digital twin models tend to say, ‘but it is just statistics and probability, we have always been governed that way’, my response is always to think ‘but what do we mean by statistics and probability?’ It seems to me that it does matter what the epistemic form of the current digital twin is and how it will generate predictive logics in society (e.g. which people are likely to repay a loan?; which treatment pathway is best for this type of breast cancer?; which arrangement of human worker and robots is most efficient?). My focus in this section is to flesh out what forms of predictive modelling are specifically emerging with contemporary digital twins. My wager is that the digital twin disavows linear-causal models that predict future states on the basis of past cases. What is sought is the discovery of predictive patterns previously unknown or unobservable by the scientist or clinician. In this specific sense, the meaning of prediction in society – the capacity to anticipate or foresee something in advance – forms a derivative governing relationship with the meaning of prediction in machine learning – the capacity to classify a new or unknown entity that was not previously seen (e.g. in the training dataset).

Consider the development of a digital twin of the heart for a London cardiology clinic. The developers propose that ‘the model can serve to find predictive relations even when the underlying mechanisms are poorly understood or are too complex to be modelled mechanistically’ (Corral-Acero et al., 2020: 4557). To build a digital twin of the heart is to find predictive relationships between interacting entities – from the physiological structures of the heart to the chemical interventions of pharmaceutical or physical interventions of surgery or pacemakers – even and especially where these exceed mechanical or rules-based forms of knowledge. The digital twin as defined by the cardiology team, as we can see with the Smart Surgery (2022) digital twin, ‘can be used to predict responses that cannot be obtained experimentally’. It is this imagination of potential future responses to an event – a new drug, a replaced mechanical device – that characterises the logic of prediction at work in the digital twin. When the team ask ‘how is this heart going to respond to this therapy?’, this is not a strictly probabilistic logic. That is to say, it is not a logic of what is the statistical probability of a particular outcome for the patient, or what is the likelihood that something will come next. Rather, the predictive logic of the digital twin is asking what is the potential response given a new entity or event introduced to the interactions. What happens if we insert this new element into the whole? How does this part change the interactions with other parts? What are the new emergent possibilities? Once again, the neural net imaginary of interactions of parts becomes present in the space of the clinic (Singh Dhaliwal et al., 2024).

The digital twin of the heart combines together ideas of prediction from clinical contexts – for example, predicting the best treatment pathway or predicting patient response – with predictive logics from machine learning – to generate probable values for unknown variables not seen by the model. ‘The concept of the digital twin’, write the cardiology researchers, orients clinical decisions towards ‘patient-specific therapy based on a virtual replica (the digital twin) to predict treatment and to personalise prognosis for the patient’ (Coorey et al., 2022). The past statistical approaches to modelling the heart, it is argued, used inferences that were valid at the level of population or group but not necessarily sufficiently precise at the level of the individual. Indeed, even the idea of what cardiac disease is begins to mutate and transform with the cardiac digital twin, marking what the researchers argue means ‘the digital twin will create the transition from describing disease to predicting response’ (Smart Surgery, 2022). The logic of prediction involved here, then, signals a shift from the observation or description of something like a disease or condition to the prediction of potential future responses to an action.

The digital twin of the heart – echoing what we saw in the space of the factory – draws upon multi-modal data from plural sources, including data from ECGs, echocardiogram images, sensor data from wearables, and the LLM derived text from patient records. Whereas past approaches to machine learning for cardiac models required the preparation and labelling of the underlying data, the generative digital twin draws on multi-modal structured and unstructured data. The relationship between the population and the individual is also distinctively different that it is said ‘a fully developed digital twin will combine population and individual representations to optimally inform clinical decisions’ (Corral-Acero et al., 2020: 4558). The digital twin reconfigures the relation of the singular person to the population or disease cohort. What is called a ‘population based digital twin’ refers to the disease cohort or ‘virtual patient cohort’ and is built using the data from ‘preceding patients and study cohorts’ (Corral-Acero et al., 2020: 4560). When a new individual patient is admitted to the clinic, their data will update the population twin and ‘novel patient data are analysed with the help of the existing models and integrated to form the patient’s digital twin’ (Corral-Acero et al., 2020: 4560). The cardiac digital twin is thus a dynamically changing and composite figure born of elements of the data of plural others (humans and models).

The digital twin of the heart is thus a complex composite that defies any clear distinction between the ‘real’ physical organ and the ‘virtual’ digital twin. The digital twin draws on the data residue from actual past patients, from the imprint of the model’s malleable parameters, and from past outcomes such as the real or simulated deaths or injuries of patients. As the scientists describe, ‘the outcome of virtual interventions subsequently given to the real-life patient then feeds back into the databank to both modify the twin and add to the population data pool’ (Coorey et al., 2022: 465). Many of these virtual interventions do not involve human patients at all but are algorithm-to-algorithm interventions that test out the prediction in advance of an actual surgical intervention. The twinned hearts thus actively shape the space of possibility of cardiology, beyond the sense of a copy of a ‘real’ heart. For example, the twin generates synthetic data to increase the representation of otherwise rare or statistically insignificant cases (Jacobsen, 2023).

The digital twin of the heart exemplifies a quite specific political technology of prediction, one that is nested within but also breaks with the eugenicist and biometric histories of the classification of body parts and prediction of the traits of people. ‘Predictions about the best treatment for an individual’, note the researchers, ‘would shift from being based on their current or past condition to being evaluated in the light of a future facing simulation’ (Coorey et al., 2022: 467). In place of deductive logics of prediction that extrapolate from past data, here is a different form of prediction that entangles the data of multiple selves and others, human and machinic, real and synthetic. An example of this novel logic of prediction is the specific use of the concept of phenotype in the cardiac digital twin. The phenotype as the expression of a trait or characteristic of an organ begins to expand beyond observable features to encompass all data, including and especially that which is not strictly observable. Smart Surgery’s (2022) digital twin is said to involve ‘deep phenotyping’ from electronic health records, biological, molecular, genetic and imaging data, combined with the ‘phenotyping of real-world data’ from wearable devices and mobile sensors. The concept of the phenotype – and particularly its use in the prediction of human traits – has a deeply gendered and racialised history (Amin, 2012). With the digital twin phenotype the concept of detecting traits deepens and extends its power via an appeal to neutral and objective data patterns that are surfaced by machine. Phenotype-based predictions illustrate vividly the kind of ethics of self and other that are always present in the predictive logic of the digital twin. When the phenotype becomes equivalent to all derivable features – observable and unobservable, real and virtual/synthetic, individual and cohort, and so on – any arrangement or proposition will always involve complex combinations of selves and others that cannot be retrospectively traced. This matters because what it means to render digital twin technologies ‘ethical’ or ‘safe’ can never be limited to ensuring good or fair outcomes. The digital twin is changing what the parameters of a good or fair outcome can be in the world, shifting what phenotype means and what a good or optimal action or action could be.

The battlefield: digital twins, actions and ethics

A computer scientist sits alongside a group of senior military personnel, government lawyers, and civil servant observers. ³ He has presented to them a demo of the functions of a ‘battlefield digital twin’ and he awaits their questions. ‘The detection of a military vehicle by the system has cascaded a series of decisions’, comments one member of the group, ‘but how sure is it that this is a military vehicle?’ In his reply, the computer scientist describes how the model is based on a LLM and is then fine-tuned on image and language data that is specific to the military domain. The examples he gives include the text of military handbooks and legal codes, the images of past battlefield scenarios, geospatial data from satellites and the video data derived from drones and unmanned ground vehicles. An animated discussion breaks out in the room. The developers explain that the outputs of the model do not need to be absolutely accurate or certain because each output has an associated probability that is reviewed by the operator. The human operator is then able to ‘prompt’ the model to query the probabilities or to change the outputs by rephrasing the prompt. This interaction with the digital twin through prompting is a major change that generative AI has brought to the logic of the digital twin. In effect, the military personnel interact with the digital twin of the battlefield by prompting and guiding the model towards desired outputs. In the case of the battlefield digital twin, each output is labelled as a potential ‘course of action’ from which the military decision-maker can select or pose a new question. The courses of action include, for example, the deployment of a ground battalion or an unmanned vehicle and so the digital twin participates in decisions that will have life-and-death consequences in the world (Weber and Suchman, 2016).

Let us consider how the battlefield digital twin arrives at a course of action. As is commonly the case, the underlying model and assumptions of the battlefield digital twin are mostly absent in the procurement demos I observed in 2023. Indeed, when the participants asked specific questions that required detail about the model – for example, whether a general LLM or one specific to legal and military texts was used; whether and how the model updates when a human decision-maker selects a course of action; if the human prompts to the model are recorded so that a chain of thought can be re-traced – these generally were not answered with any technical detail. A general problem for the ethics and accountability of digital twins deployed in many domains, from medicine to the military, is that the worldview of the model and its form of reasoning is oblique and annexed from critique. However, it is often possible to trace an alternative form of accountability that is present in the partial accounts given in the computer science articles that are published on the specific problem (Amoore et al., 2023). Pre-dating the commercial development of military digital twins by companies such as Palantir and others, arguably the first computer science article published on the battlefield digital twin establishes the vocabulary of ‘course of action’ that persists in the downstream commercial and government models (Wang et al., 2021). As the computer scientists elaborate on their digital twin, ‘through the battlefield states generated by the algorithm, the combat simulation systems guide the military commanders to respond to situation changes and choose the optimal courses of action’. These outputs or courses of action are ranked as probabilities and they can be queried or adjusted via the prompts of the human operator. The digital twin is generating these courses of action from a complex array of operations – from the underlying LLM and its learned syntactical patterns of language, to the algorithms recognising images, and the text data used for legal and military rules and doctrines. In effect, the pathway of the course of action through these layers of the model is unintelligible to the commander or operator, and indeed to the developer. Though the actions, prompts and queries the human operator makes in relation to the digital twin will fundamentally shape and adjust the output, this is not the same thing as having insight into or control over the model.

What is wanted from the battlefield digital twin is a capacity to detect and surface the emergent properties from a data distribution – properties that are not otherwise available to human perception or control. When the machine learning researchers compare their battlefield digital twin to past models and simulations, they point out that the traditional ‘vector-based data representations of the battlefield’ were limited because they had ‘fixed dimensions’ that did not dynamically update with new data (Wang et al., 2021). In a sense the vector-based data representations were much closer to a ‘copy’ of the battlefield, whereas the generative battlefield digital twin is definitely not a copy; it does not work with fixed dimensions but is precisely generating its own sense of what matters and its own regime of attention (Pedersen et al., 2021). This paradigm of the digital twin is absolutely at the heart of the dreams of automation where ‘the most important value of the proposed algorithm’, the developers suggest, ‘is to break through the data boundary between the real and the virtual battlefield and enable the application of digital twins in unmanned combat’ (Wang et al., 2021). Like the digital twins of the factory and clinic, the battlefield model is geared towards automation in the sense that the digital twin is modulating in relation to its data environment and outputting its own courses of action. What is taking place here is a politics where the digital twin substantively sets the aperture for many crucial decisions – military action, clinical judgement and so on – and simultaneously forecloses and erases the multiple possible alternative decision pathways.

How might we think about the politics and ethics of a battlefield digital twin that outputs courses of military action while foreclosing the space for other decisions? As DeLanda (2011) reminds us, ‘computers supply the means to explore these other possibility spaces’ by varying ‘the interactions in which capacities are exercised’ (pp. 20–21). The digital twin does indeed explore other possibility spaces within a vast underlying data distribution. Yet, this process of carving out and reshaping the space of possible action will never be the same twice, it is continually and dynamically mutating. The battlefield digital twin elaborated here does actualise the space of possibility by exercising the capacities of interactions in the model. All of this emergent contingency, however, becomes crystallised in an actual output that is said to be the most probable or optimal course of action. The outputs of digital twin models do become actions and decisions in the world, decisions that occlude the plural alternative pathways that could have been taken. In the production decisions of the factory, the clinical decisions of the cardiologists and the military decisions of the battlefield, the digital twin condenses and forecloses its own teeming potentialities. An ethics up to the task of addressing the digital twin would need to begin by restoring the fragility and contingency of the course of action and emptying it of its feigned certainty.

Conclusions: the twin and the attribute

In a 2006 essay, ‘making up people’, Ian Hacking (2006) outlines how ‘sometimes, our sciences create kinds of people that in a certain sense did not exist before’ (p. 2). By the making up of people who did not exist before, Hacking does not mean the making or synthesis of new subjects – as perhaps we might interpret the digital twin subjects of Riley’s and Klara’s worlds. Rather, his point is that the science and technologies of classification do not merely render a mirror copy or reflect some underlying essence but actually create new classes, make up new populations and worlds. ‘Statistical analysis of classes of people’, writes Hacking, ‘is a fundamental engine’, a form of world making that makes societies knowable and governable in particular ways. It is in this longer history of scientific and social scientific ways of making up and classifying the world that the contemporary digital twin emerges. The digital twin is making up people, things, scenes and their interactions in novel ways – the drawing in of unstructured datastreams, their experimental recombining and modifying, the mutual interplays of the physical and modelled worlds. In this sense, the making up of the digital twin, as I have discussed in this essay, is not the same thing as the making of a replica or copy – though it is often through a performative appeal to digital copies or replicas that the digital twin is represented. The digital twin departs radically from a mimetic logic of the copy or replica and advances instead a mode of twinning that extracts features and multiplies, as in the old etymology of the verb ‘to twin’. The digital twin does not need a complete copy or likeness to the world, what it needs is sufficient plausibility to make intervention, prediction, and action possible.

The figure of the twin has also been a historically important trope of experimentation in the human sciences (Viney, 2021). So-called twin studies were prevalent in twentieth century science because they were thought to make it possible to hold some phenotypical variables constant (inherited traits and characteristics) while studying the environmental variables for differences in learned attributes. Of course, the history of twin studies is also deeply implicated in the eugenics of Francis Galton and in the horrifying twin experiments of the holocaust. The separated twin study became a means of observing traits and characteristics and inferring human behaviour and heredity. Yet, the digital twin – though it echoes the place of the twin as source of knowledge in scientific and social scientific histories, crucially ruptures this notion of the fixing of variables and the observation of attributes and behaviours.

In an age of deep learning and generative AI, the digital twin is beginning to align with a new paradigm of pattern discovery, unstructured data, and domain agnosticism. As I have discussed, the battlefield digital twin researchers sought to move on from the ‘fixed dimensions’ of vector models in favour of detecting emergent patterns; the cardiac digital twin was offered as a means of overcoming the fixed variables of mechanical heart processes; and the factory digital twin was conceived as a means to intervene in the future factory as it unfolds. To be clear, I am not suggesting that the scientific and commercial claims of platforms such as Omniverse are ever fully realised as such. Rather, the claims and promises of the digital twin in our times are exercising epistemic power in the world, whether or not they are ever fully materialised. This is a form of political power that works by reordering spaces, intervening in anticipation of a response modelled in the twin, making crucial and consequential decisions on the basis of malleable outputs.

In my discussion of the three aspects of the logic of the digital twin – intervention, prediction, and action – I have mapped the contours of this novel form of simulation. The digital twin anticipates what an intervention might do to the space of factory, clinic or battlefield. This logic of intervention asks ‘what if we move this robot?’ or ‘what if we introduce this new pharmaceutical drug?’ or ‘what if we deploy this ground team?’. More than this, the capacity to ‘prompt’ the digital twin will also continually adjust the intervention through an interplay of human and model. The digital twin is a speculative and experimental space that incorporates the computational logics of AI into the social and political world. As a logic of prediction, the digital twin does not seek causal or linear-causal connections between known past data elements, but rather seeks to predict response in advance of an intervention. As in the digital twin of the heart, decisions about surgery or drugs are not structured primarily in relation to past outcomes but more significantly through a predicted response based on a complex composite of data, real and synthetic. Finally, as a mode of action, the digital twin outputs ‘courses of action’ that are distilled from an infinite series of model to world interactions that are not intelligible as such. Understood in this way, the digital twin is a specific and novel form of the governing of conduct, the shaping of spaces of possibility and decision, and the foreclosing of other possibilities.