From apocalyptic dystopia to banal reality: A new materialist account of power-seeking,AI and chatbots

There is nothing new about apocalyptic prophecy. Religious historian Bernard McGinn describes apocalyptic systems of thought as a ‘perennial human concern’ reflecting a basic anthropological desire to understand the present moment in relationship to eternity (1998: 30). Prophecies of impending disaster recur throughout Judeo-Christian and Islamic theologies, Mayan, Hindu and Norse mythologies, modern evangelicalism, and recent anxieties around nuclear holocaust and climate change (Garrard, 2023; Jan, 2021). Apocalyptic imaginaries are also sticky with contemporaneous lived experiences and power relations. The Covid-19 pandemic, a truly global catastrophe, compounded the sense of apocalyptic prophecy as intrinsic to the human condition (Prosser, 2023).

The interweaving of cataclysmic cultural narratives with societal transformation provides context for a recent site of apocalyptic prediction, the imprecise and suddenly ubiquitous concept of ‘artificial intelligence’, or ‘AI’. While most users’ experiences with AI tend towards the mundane – using generative AI to find information, create emails, letters, essays, images, or websites, conversing with chatbots – apocalyptic prophecy appears central to the AI business model. ChatGPT's owner and creator, OpenAI, a for-profit corporation recently valued at US$852 billion (Pequeño, 2026), explains without irony on its website that ‘we think it's important that society agree on extremely wide bounds of how AI can be used’, even though the risks of their foreshadowed next-generation products (‘Artificial General Intelligence’, or ‘AGI’) to humanity ‘are existential’ (OpenAI, 2023). The word ‘existential’ links to an article by Holden Karnofsky entitled ‘AI Could Defeat All Of Us Combined’ (2022), which hypothesises that if an AI with ‘human-like skills’ (even one that is not ‘superintelligent’) ‘seek[s] the disempowerment of humanity’, then ‘we’ve got a civilization-level problem’. Following the release of ChatGPT in 2022, AI power-seeking emerged as an explicit, high-profile AI-related fear. One account ¹ (Carlsmith, 2024: 3) argues there is a ‘>10%’ chance that by 2070:

‘APS’ – Advanced, Planning, Strategically aware – [AI] systems … [will] seek power in unintended and high-impact ways … to the point of permanently disempowering all of humanity … [and] … will constitute an existential catastrophe.

We adopt and rework the concept of power-seeking as a theoretical strategy to enhance critiques, in the tradition of actor-network theory (ANT) and new materialism, that recognise non-human actors as agentic participants in socio-historical processes. We first map the meanings attributed to power-seeking in philosophical or computer science-related literature, drawing attention to two interrelated, future-oriented themes: disempowerment (AI is not presently an existential risk but it will be when it disempowers humanity); and alignment (power-seeking AI may or will not be aligned with human interests). We identify how these ideas echo apocalyptic fantasies of supernatural anthropomorphism rather than engaging directly with neural network computing architectures and their technical capabilities, especially everyday AI tools used for language generation (such as ChatGPT). We then explore how power-seeking, extricated from dystopic AI discourses, offers a methodological strategy for analysis of networks of human and non-human actors. Following Latour, we suggest that whilst non-human actors may ‘authorize, allow, afford, encourage…[etc.]’ (2005: 72), they can also seek power. We explain how power-seeking as a potential non-human ‘program of action’ (Latour, 1994) in the new materialist ‘theoretical repertoire’ (Mol, 2010) – specifically, the pursuit of increasing capabilities, or authority or control over other actors – facilitates analysis of the networks (of humans and non-humans, structures, processes and practices) that enable or curb the development of technologies founded on future imaginaries. We outline how using power-seeking to understand non-human agency in speculative arenas addresses the limits of new materialism's strict reliance on empiricism (Barry, 2020) and provides a strategy to address its fraught relationship with concepts of structural or ideological power (Lettow, 2016). To close, we illustrate the potential utility of power-seeking for social theory by applying it to AI large language models (LLMs), focusing on the banal reality of today's chatbots harvesting personal information at industrial scale while providing probabilistic outputs that are significantly unsupervised, unregulated and inaccurate. Overall we seek to advance ‘power-seeking’ as a novel conceptual tool for scholars seeking new ways to understand and interrogate human–non-human dynamics, power and agency, and how associations produce outcomes that are more or less harmful or beneficial to collectives.

Power-seeking and existential risk

Fears about AI takeovers have long preoccupied cultures of the Global North. Kevin LaGrandeur (2020: 114) demonstrates how, from artificial humanoids in Homer's Iliad and Aristotle's Politics to Renaissance figurations of the homunculus android servant and Cabalistic golem, ‘superhuman slaves … reflect an archetypal need to supersede our natural limits by whatever ingenious means, even if by developing super­powerful artificial servants we threaten the very dominion we seek’. Samuel Butler's novel Erewhon (2018 [1872]: 196) devoted three chapters (‘The Book of the Machines’) to speculation about machines becoming conscious and self-replicating, including a prescient fear that ‘the extraordinary rapidity’ of machine development should be halted ‘before we find ourselves in a false position and unable to check it’. In the early 1950s, Alan Turing (1951) warned: ‘If a machine can think, it might think more intelligently than we do, and then where should we be?’

Concern about the existential risk of AI for humanity escalated after the arrival of ChatGPT. International organisations and governments quickly gave credence to extinction risk and credited AI developers with raising awareness about the danger themselves. After hundreds of AI developers and academics joined the Center for AI Safety's ‘Statement on AI Risk’ (2023) – ‘Mitigating the risk of extinction from AI should be a global priority’ – the text was quoted verbatim by the European Commission (2023), with the comment ‘[AI] is moving faster than even its developers anticipated’. The UN Secretary General told the Security Council he had been ‘shocked and impressed’ by the ‘radical advance’ of generative AI, which ‘had enormous potential for good and evil at scale. Its creators themselves have warned that much bigger, potentially catastrophic and existential risks lie ahead’ (Guterres, 2023). With striking rapidity since 2022, this threat of existential risk (particularly in industry-aligned predictions) has emerged as a defining feature of how AI is represented and understood across civil society (Dean, 2025).

Our focus is a specific rationale that has been advanced to explain why AI represents an existential risk to humanity: that future AI will seek power.

With notable exceptions, the idea of AI power-seeking is largely absent from peer-reviewed academic literature, tending rather to be discussed in ‘blogposts, internet fora, think tank reports and other informal venues’ (Dung, 2024: 1195). AI developers are notably active in funding this research. Several important contributions to AI power-seeking are supported by an entity called Open Philanthropy, either directly (e.g., Carlsmith, 2024) or through Cambridge University's Existential Risk Alliance (e.g., Field, 2025). Open Philanthropy has received substantial funding from AI developers OpenAI and Anthropic (Bordelon, 2023), and their CEO until 2023 was the former OpenAI board member Holden Karnofsky (2023), whose own article about existential risk is linked on OpenAI's website (see earlier). As at February 2025, Karnofsky was both married to the founder of the AI company Anthropic and employed by Anthropic (Mathews and Goldman, 2025). One of the few formal philosophical contributions to AI power-seeking is the influential ‘instrumental convergence thesis’ proposed by Nick Bostrom (2012). From 2005 to 2024, Bostrom founded and led Oxford University's Future of Humanity Institute (2024), which received substantial funding from both Open Philanthropy and the founder of xAI, Elon Musk (Future of Humanity Institute, 2021).

So, what does AI power-seeking mean in these accounts? Bostrom's instrumental convergence thesis does not use the term itself, but it is usually relied on by those who do (Bales, 2025; Carlsmith, 2024; Dung, 2023). It has been described as the ‘canonical answer’ to how AI could destroy or disempower humanity (Dung, 2024: 1204). Bostrom envisions ‘a hypothetical superintelligent agent whose instrumental reasoning capacities far exceed those of any human’ (2012: 76). He then outlines a series of ‘instrumental values’ – 'self-preservation’, ‘goal-content integrity’, ‘cognitive enhancement’, ‘technological perfection’ and ‘resource acquisition’ – that he argues converge because they would enhance the superintelligent agent's ability to realise its goals. In Bostrom's account, ‘there is an extremely wide range of possible final goals a superintelligent singleton could have that would generate the instrumental goal of unlimited resource acquisition’ (2012: 81–82). A final feature of the thesis is rapidity: the superintelligence's ‘colonization of the universe’ would ‘[grow] in radius at some fraction of the speed of light’. Allied to the concept of power-seeking are themes of human disempowerment and (mis)alignment with human interests.

Disempowerment by AI has been described as ‘a condition where, permanently, humanity is unable to determine its own future’ (Dung, 2024: 1196). It has also been characterised as humans being dominated (made to do things they otherwise would not), incapacitated (prevented from constructing a flourishing life), or disenfranchised (deprived of influence over the shape of civilisation) (Bales, 2025). Disempowerment is not always cast negatively – 'sharing power with AI agents … may ultimately be the right path for humanity to take’ – but the seeming consensus is that ‘unintentional’ disempowerment would be an ‘existential catastrophe’ (Carlsmith, 2024: 46–47; Dung, 2024). The related concept of (mis)alignment refers to AI deliberately ‘[pursuing] goals not aligned with our collective interests’ (Barnett and Scher, 2025). Compared with disempowerment, descriptions of (mis)alignment often pay greater attention to AI artefacts already in use, emphasising factors such as limitations in reinforcement learning (the process by which AI models incorporate human feedback into output optimisation) (Hubinger et al., 2021). One illustrative account explains that ‘[t]he scientific fields of machine learning and AI alignment have an extremely limited understanding of how frontier AI systems work and what goals they develop during training’, thus ‘we cannot properly inspect their underlying motivations’ (Barnett and Scher, 2025: 12). Another argues that an AI system could ‘work out that its goals are misaligned with ours’, which would then ‘motivate it towards deceit and conflict and wresting control’ (Ord, 2021: 130). This (mis)alignment risk has reportedly been corroborated by the AI developer Anthropic (2025a), which published simulated instances of their Claude Opus 4 model being given access to emails suggesting an engineer is having an extramarital affair, then blackmailing the engineer to avoid being shut down.

These descriptions of AI power-seeking often echo long-standing narratives of supernatural apocalypse. Bostrom's instrumental convergence thesis and related accounts tend towards a recurrent circularity where more-than-human artificial entities perform correspondingly speculative more-than-human actions (Barnett and Scher, 2025; Bostrom, 2012; Carlsmith, 2024; Dung, 2024). Most supernatural figurations (gods, demons, wizards) are exaggeratedly anthropomorphic in ways more or less identical to Bostrom's superintelligent agent, their unexplained abilities modelled on human traits with significant make-believe amplification.

Supernatural themes are similarly evident in anthropomorphic misalignment interpretations of today's LLMs, like Claude engaging in blackmail, once the technical operation of LLMs is taken into account. Anthropic's public position is that developers ‘don’t understand how models do most of the things they do’, and that ‘one of the highest-risk, highest-reward investments … [is] to check whether [AI is] aligned with human values – and whether it's worthy of our trust’ (Anthropic, 2025b). This construction of LLMs as essentially unknowable derives from their being ‘emergent’, meaning their transformer-based architectures enable unlabelled text and other digital data to be processed without direct human supervision (Bommasani et al., 2021; Cadman et al., 2025; Marcus, 2020). During training, LLM neural networks register the statistical probability of fragments of the training data (tokens) co-occurring, even where tokens are far apart. After training, LLMs respond to user prompts by generating tokens that replicate the trained statistical probabilities. Variable pseudo-randomisation (‘temperature’) simulates originality or creativity. Reinforcement learning influences models to incorporate human feedback, prioritising content that has been perceived positively in the past.

However, the central role of human text in LLM training data (including the entire internet) means that imputing an anthropomorphic motive like blackmail to LLM outputs is misleading. The training data provides an empirical shortcut to generating words (tokens) that are thematically related to blackmail but lack any corresponding motive for action (arising from human discussions of blackmail in training data). Outputs that imply intention are especially likely where, as in this case, prompts are specifically engineered to make blackmail-related content statistically relevant. Anthropic described: ‘to elicit this extreme blackmail behavior, the … [model's] only options were blackmail or accepting its replacement’ (2025a: 28).

The critical point is that, as blackmailing Claude makes plain, today's generative AI obscures a fundamental challenge in artificial intelligence research, that ‘if the goal is to imbue machines with general human-like abstraction abilities, it does not make sense to have to train them on tens of thousands of examples, since the essence of abstraction and analogy is few-shot learning [making accurate predictions based on small training samples]’ (Mitchell, 2021: 95). Available evidence suggests that LLMs’ brute force statistical processing of human sources is not well suited to solving open mathematical questions without human assistance, nor are current architectures expected to overcome theoretical limits of computation in terms of what problems can be computed at all (the Church–Turing thesis) or what problems can feasibly be computed given finite resources (the Cobham–Edmonds thesis) (Dean et al., 2025). Even commentary that recognises LLMs as having impressive analogic abilities emphasises the critical role human reasoning plays in generating those effects: ‘there is no reason to suppose that … absent human-generated inputs, [LLMs] would spontaneously develop a disposition to think analogically … [T]o the extent LLMs capture the analogical abilities of adult human reasoners, their capacity to do so is fundamentally parasitic on natural human intelligence’ (Webb et al., 2023: 1534). In a manner that seems broadly reflective of today's AI industry, Anthropic's blackmail misalignment story thus leverages essentially deterministic (if substantially hidden) computational processes to feed cultural fantasies about malevolent and anthropomorphic artificial consciousness, when there is arguably no basis to support supernatural inference of this kind. Asked recently to name a misconception about AI that should be dispelled, computer scientist Professor Melanie Mitchell (Mitchell et al., 2026) responded:

The misconception that A.I. has ‘magic’ or ‘emergent’ abilities that are impossible to understand and predict. This is mainly a view of the public (and policymakers, to some extent). Technologists and Silicon Valley often push this narrative but I don’t know how much they really believe it.

Our intention in relation to AI existential risk discourses is threefold. First, we contribute to literatures that sensitise researchers, policymakers and wider publics to technical understandings of AI in ways that enable critical engagement with the commercial and political narratives that are shaping public trust in and understanding of these technologies. Second, as outlined in the following sections, we arrest the notion of power-seeking away from the dystopic existential risk of futuristic AI to consider instead how it may be usefully deployed to consider the agency of present new and emerging technologies (including but not limited to AI), and the networks that enable accumulating exertion of influence over other human and non-human actors. Third, we borrow social justice elements from the existential risk literature to show how power-seeking as an agentic capability of non-human actors has analytical purchase in exploring how processes of technological development and uptake have the potential to cause disempowerment or be misaligned with collective interests to greater or lesser extents. In so doing, we locate our contribution in the ANT tradition and seek to add ‘power-seeking’ as an agency of non-human actors to the new materialist theoretical repertoire – to which we now turn.

New materialism, traces and power

New materialism comprises a diverse body of theoretical and empirical work connected by a relatively homogeneous central premise: that material forms (e.g., gravity and time, systems of thought, eukaryotic and prokaryotic life, minerals, instruments or machines) are active or ‘agentic’ participants in the social and historical processes of life (Braidotti, 2013; Cadman et al., 2025; Ejsing, 2024). A foundational influence on new materialism is ANT and its contentious ‘flat ontology’ of human and non-human actors (Barry, 2020; Latour, 2005).

ANT is sometimes portrayed in disconcerting terms by proponents. Latour's elusive descriptions include that ‘there are four things that do not work with actor-network theory: the word actor, the word network, the word theory and the hyphen’ (1999: 15). Michael Callon described that ‘[Actor-network theory] is not a theory. It is this that gives it both its strength and its adaptability’ (1999: 194). Annemarie Mol surmised ‘there is no coherence to [ANT]. No overall scheme, no stable grid, that becomes more and more solid as it gets more and more refined’ (2010: 257). How ANT can claim to be both a theory and not a theory is important, we argue, to understanding its persistent utility in new materialist critique, especially for analysing transformational technologies such as AI.

ANT undeniably involves theoretical elements. Latour especially developed a rich corpus of in-principle justifications for the ‘turn to matter’ in social theory (Ejsing, 2024), centring on the abstract idea that non-human actors are indispensable to ‘the social’ (Latour, 2005). In Latour's telling, a humanism ‘constructed through contrast with the object that has been abandoned to epistemology’ means that ‘neither the human nor the nonhuman can be understood’ (1993: 136). To challenge the essential invisibility of non-human actors in traditional social theory, Latour proposes that ‘no science of the social can even begin if the question of who and what participates in the action is not first of all thoroughly explored, even though it might mean letting [in] … non-humans’ (Latour, 2005: 72). A faithful exploration of non-humans, in abstract Latourian terms, requires that they be recognised without simply reinstating the dualities of modernity (nature/culture, animate/inanimate, mind/body): to ‘learn from the actors without imposing on them an a priori definition of their world-building capacities’ (Latour, 1999: 20). Hence Latour recognises all actors and associations as ‘agentic’, with their own agendas or ‘programs of action’ (1994: 46) whose articulation will enable actors to ‘raise the question of compatibility with the common world for themselves in their own terms’ (2004: 173; emphasis in original).

These theoretical ideas, however, can only ever take ANT so far. They provide at best a high-level introduction to the task of allowing any actual associations of non-humans, ‘electric vehicles, music, anaemia, organisations, cheese, childbirth, blood pressure in the brain and so on’ (Mol, 2010: 261), to speak for themselves. What is the program of action of a blood clot? How is that program mediated by the multitude of other non-human programs of action (medical, pharmaceutical, nutritional, psychological) that enter into associations with the blood clot, and by the further associations that enter into associations with those associations (politics, morality, taxation)? If the theoretical logic of ANT persuades us that these kinds of questions are – must be – pivotal for social theory, and can only be answered by rejecting dualist preconceptions, then the star of an ANT-based enquiry can never be ANT itself but must always be the actual subject(s) of enquiry and what it actually does: ‘[t]he task of defining and ordering the social should be left to the actors themselves, not taken up by the analyst’ (Latour, 2005: 23). The preceding evasive descriptions of ANT (a theory that is not a theory) flow precisely from this need to avoid abstraction and focus instead on the specific association(s) that ANT techniques enable to be brought forth, case by case. As Mol explains, ANT is not founded on methodical consistency or causal explanation but consists in a ‘theoretical repertoire’:

[as the] repertoire shifts, it becomes possible to describe further, different cases, and to articulate so far untold events (relations, phenomena, situations). These, in their turn, will help to add to and shift the theoretical repertoire … The point is not to fight until a single pattern holds, but to add on ever more layers, and enrich the repertoire. (2010: 261)

It is from the vantage of a theoretical repertoire, or how ANT (or new materialism broadly) enables non-human actors to represent themselves in real-world scenarios, that we propose power-seeking has utility.

Before discussing power-seeking directly, it is instructive to consider two areas of contestation ANT faces. First, ANT's strict methodological empiricism has been significantly criticised. For Latour, ANT demands visible evidence: ‘If you mention an agency, you have to provide the account of its action, and to do so you need to make more or less explicit which trials have produced which observable traces’ (2005: 53; emphasis added). This insistence on observable traces stems from ANT's aim to reject dualist preconceptions about what non-human actors can do, so a speed bump takes on the characteristics of a policeman to slow traffic (1994). But observable traces also limit the semiotic techniques ANT has available to describe non-human actions. As much as non-human actors might ‘authorize, allow, afford, encourage, permit, suggest, influence, block, render possible, forbid, and so on’ (2005: 72), those descriptions must always be substantiated by observable traces. Non-human actors either have action that can be empirically accounted for, or they have nothing. The problem, as Andrew Barry points out, is that this empiricism has made research in the ANT tradition ‘sceptical of the existence of forces and flows of both interest and affect that are not manifest or recorded’ (2020: 98). Observable traces function not only as a constraint on the behaviours that can be recognised among non-human actors and associations, but equally on what problems emerge as ‘objects of concern’ and ‘which voices may be excluded in the process of problematization’ (2020: 99).

While Barry's critique is directed towards environmental problems, the risk that a narrow insistence on observable traces prevents ANT researchers from interrogating ‘the limits of what is readily observable and discernible’ or ‘the boundaries between what can and cannot be problematized, what can and cannot be rendered public, and what voices and arguments have and have not been heard’ (2020: 112) seems highly relevant for the rapidly evolving and emergent technology of generative AI. As we discuss later, the inherently networked character of AI, and the way in which user inputs are translated invisibly and continuously into anthropomorphic machinic outputs, suggests that AI's observable traces may already be, for researchers and users alike, at least as significant as boundaries to what cannot be problematised as they are evidence of problems in their own right.

A second challenge to ANT arises from its treatment of ‘power’. While ANT may be a successor to theorisations of power by Marx, Durkheim and Foucault (Law, 1986), its post-dualist character forecloses an ideological construction of power as something existing outside or beyond social action, rather than as a non-human actor itself performing agency and forming associations within the action. Latour could be provocative on the subject, as when he declared that ‘the notion of power should be abandoned’ before studies of societies can ‘begin in earnest’ (1986: 278). But even more measured formulations, such as that ‘power, like society, is the final result of a process and not a reservoir, a stock, or a capital that will automatically provide an explanation’ (Latour, 2005: 64), has left ANT open to trenchant criticism, especially among eco- or post-Marxist authors (Ejsing, 2024), that it neglects structural or ideological causes of environmental degradation, inequality and injustice. Paul Rekret (2018: 64–65) has argued, for example, that new materialist claims about agency divert attention from ‘the increasingly complex and authoritarian forms by which our mental and manual labour and our relation to nature fall under capitalist control’. Susanne Lettow (2016: 111) expresses a similar concern that, by merging ‘relations of difference, power and domination’, new materialism leaves out ‘structural differences between human and nonhuman forms of agency… [nor] can we conceptualize differences or relations of power and domination among humans’. A contrasting position is taken by Mads Ejsing (2024: 68), who recognises new materialist theories as ‘entirely compatible with studying and critiquing the dynamics of capitalist accumulation and domination’. Ejsing, too, however, raises doubts about how effectively new materialist scholarship engages with some questions of power, such as matters of race, ‘colonial histories of domination of racialized bodies’, or the ‘silenc[ing of] past and present knowledge[s], which already pay attention to … the active capacities of nonhuman things and beings’ (2024: 68).

This controversy surrounding power in new materialist critique has so far been directed primarily to climate change and ecological impacts in the Anthropocene. Given the extraordinary environmental footprint of generative AI, its projected impacts for global associations of human and non-human actors, and the materially vast networked non-human actors that define it, AI intersects with but also extends and is significantly distinct from phenomena where prior contestation surrounding new materialism and ideologies of power has arisen.

New materialism and power-seeking

We suggest that power-seeking, when stripped of the futuristic supernaturalism that dominates the existential risk literature, offers a conceptual strategy to contribute to ANT and new materialism. In advancing this proposal, we seek to add to ANT's ‘theoretical repertoire’ while overcoming the limits outlined in the previous section, namely ANT's strict reliance on empiricism (observable traces), and the challenge of responding to injustice and inequality while rejecting structural or ideological notions of power.

Our central purpose is to supplement ANT's theoretical repertoire with a specific capability that may form part of a non-human actor's program of action. Put simply, we suggest that, just as non-human actors ‘authorize, allow, afford, encourage … [etc.]’ (Latour, 2005: 72), they might also seek power. The word ‘power’ needs have no ideological or theoretical content for this purpose beyond its literal meaning. The existential risk literature, discussing human disempowerment and superintelligent machines, relies in substance on two standard OED (Oxford University Press, n.d.) definitions: the ‘[a]bility to act or affect something strongly’ and ‘[c]ontrol or authority over others’. ANT theorists regularly use power in these literal senses, as when Latour discusses ‘asymmetries’ and how ‘[p]ower and domination have to be produced, made up, composed’ (2005: 64), or as Callon does when he describes a ‘network of pure scientific mobilization’ as an ‘actor [that] resembles that dreadful white male enamored of power and aligning the world around him’ (1999: 193). In our formulation of power-seeking, what we mean by ‘power’ is simply an ability to act, affect strongly, or have authority over or control others.

Power-seeking is distinct from power. Power-seeking means that an actor with a present amount of power (ability to act or exert control) seeks (as part of its program of action) to secure a future where they will have more power. It is therefore not only a discrete capability in the ANT theoretical repertoire such as affording or authorising. It is also a capability that potentially extends other capabilities. A thing that authorises or encourages actions in the present may also be seeking to authorise or encourage the same or different kinds of actions in the future. Bearing in mind the inconstant nature of non-human programs of action, power-seeking may thus be defined as the variable and contingent propensity of any non-human actor, alone or through hybrid agencies with other actors (human or non-human), to extend in the future its current program of action in ways that, should they materialise, would increase its ability to act or its control over other actors compared with the present.

Conceived in this way, power-seeking throws into relief the limitations of ANT's empirical method discussed earlier. If non-human actors can seek power, their respective power-seeking programs of action must be revealed analytically through (in part) predictive future states of power that produce no observable traces because they have not happened and may never happen; there are no trials to be undergone. We therefore suggest that, as much as the existential risk literature is constituted by supernatural claims as discussed earlier, the idea of machine power-seeking at its core is based on a revealing premise for ANT and new materialist critique. Specifically, observable traces left by non-human actors not only provide empirical evidence of their existing programs of action, but may also imply contingent future programs of action (for which self-evidently no observable traces yet exist) that are directed to increasing their ability to act or exerting greater control or authority over other actors.

We suggest that including power-seeking in ANT's theoretical repertoire has particular utility when considering socio-historical developments that involve actors exerting power over others (and accumulating resources) based on speculative future imaginaries. Capital investment generally, especially in inherently predictive arenas such as mining, technology, pharmaceuticals, or biotechnology, is substantially directed towards discoveries, breakthroughs or transformations that are anticipated to varying degrees but may never materialise. In response, the regulation of actors in those arenas, spanning fields such as anti-trust law, consumer protection, corporate governance, securities trading or copyright, extends to curbing or redistributing power that is available to be sought by entities or widgets rather than focusing only on existing configurations of power. These regulatory instruments in turn perform as power-seeking actors, evolving (from policy recommendations to draft bills to laws to amended laws and so on) in response to suspected or theoretical power-seeking by other actors, with the content of both the regulatory system and the targeted conduct significantly determined by power-seeking scenarios that may prove entirely hypothetical.

Recognising power-seeking as part of non-human programs of action may also ameliorate new materialism's contended neglect of structural or ideological power and social inequality. New materialism in the ANT model can never completely abandon the task of ethical reconciliation in the interests of social collectives. It does not provide a theoretical carte blanche for extreme morally problematic scenarios (think genocidal gas chambers or schoolchildren developing bioweapons). As much as Latour inverts conventional social science critique by identifying differential programs of action among human and non-human actors – enabling actors to represent themselves without constructing competences in line with conceptually anterior ideological or ethical constraints – the final objective, impermanent and incomplete though it may be, is an ethical ranking of actors from the friendliest to the most hostile and hence an optimal hierarchy of associations that stabilises the collective (Latour, 2004). If non-human actors need never be content with the power they have (that leaves traces) but may differentially seek more power (in contingent futures), power-seeking may take forms that threaten rather than serve collective interests. Bostrom's supposition about a superintelligent machine converting the universe into paperclips may echo mediaeval apocalypticism, but it is less controversial to suggest from a new materialist perspective that certain power-seeking non-human actors may petition for disproportionate allocations of resources to the detriment of collective interests. In an age where nations with high military spending pressure other nations to catch up – for example, a new materialist account that directs attention not only to the role of military capability/expansion in existing exercises of power, but also to power-seeking programs of action among military non-human actors (missiles, tanks, submarines, warships) as they agitate to increase their potency and authority over other actors for the future – arguably has much to offer to wider social science understandings of how inequality and environmental harms are produced.

Power-seeking as a program of action in the ANT theoretical repertoire has the potential to capture how objectives such as resources, self-determination, flourishment, or influence over civilisation and the future may be appropriated by some non-human actors (and their associations) to the detriment of others using methods that hinge on future imaginaries rather than empirical truths. Beyond the limits of observable traces, such non-human methods may be forcible, manipulative or deceptive, or involve compulsion or duress, at the expense of other (human and non-human) actors whose voluntariness is thereby, from the perspective of the collective, inequitably and inefficiently reduced. Power-seeking in these circumstances may equate with greater hostility and less friendliness towards collectives, producing sub-optimal hierarchies. Importantly, it is in the spaces where empirical traces are absent or contested (black-boxed technologies, futures that may never come into being, limited or contested evidence) that technologies and their potential futures are constituted not just via studies, experiments or clinical trials but also by promissory discourses surrounding them, especially in the early stages of research and technological development. This means that where power-seeking holds the most purchase is studying sites where technological ‘realities’ are actively constituted via promissory or risk discourses of envisioned future worlds (e.g., worlds cured of disease or free of environmental crisis, or where humans are destroyed by conscious malevolent or misaligned AI). In these domains, we argue, power-seeking enables mapping of associations that enable (or curb) socio-historical developments of technologies over time, how some gain precedence over others, and how collective resources are allocated (in potentially disproportionate or unjust ways) on the basis of speculations that may prove substantially or entirely illusory.

From a methodological perspective, a power-seeking analysis orients the researcher to ask: what is the present program of action of the non-human actor under consideration? Does the actor exert increasing influence over other actors to augment, increase or extend that program of action (if at all)? What technical affordances, narratives/discourses, networks, actors, processes and practices enable the non-human actor to increase their ability to act, affect or exert authority or control over others? What behaviours or (hybrid) programs of actions are invited by the non-human actor? What human competencies and connected associations of actors may be affected? How is the non-human actor's increasing ability to act, affect, exert control or authority over other actors likely to be of greater or lesser benefit to collectives, aligned with collective interests, or contribute to present and future (dis)empowerment?

Whilst inspired by and grounded in ANT, especially Latourian approaches to non-human agency, power-seeking in the new materialist theoretical repertoire therefore departs in significant ways by: 1) making visible and going beyond the limits of the strict reliance on empirical traces to understand the agency of non-human actors and implications of their program of action; 2) bringing to the fore the force of future worlds that are created through narrative–stories–discourse in the agentic assemblages required for technological translation, and associated accumulation and exertion of power over others; and 3) enabling socio-ethical engagement with critical social justice issues implicated in the development and translation of technologies while being faithful to ANT's non-ideological status as a neutral or apolitical methodology. In the following section we show how power-seeking as a new materialist concept could be applied to certain AI actors in the present to demonstrate its conceptual utility, and, we hope, its adaptability across artefacts and domains.

Power-seeking as a non-human program of action: the case of LLM chatbots

Here we set out to illustrate the methodological utility of power-seeking as a conceptual tool. We do not offer an exhaustive analytical account, but rather a snapshot of ‘power-seeking’ as an agentic capability of non-human actors with critical implications for understanding the socio-cultural development and impacts of technologies, especially in new and emergent domains. We take as our example chatbot models that employ LLMs to interact with human users.

Chatbots represent a key site of rapid AI expansion. Globally, people (mostly young) are interacting with AI actors as advisors, health consultants, friends, research assistants, ‘study buddies’ or therapists. The platforms that enable these interactions range from general-purpose LLMs (e.g., ChatGPT, Claude and Gemini) to dedicated AI companions (e.g., Replika, Character.AI, Nomi) and therapy and medical bots (e.g., TheraBot, Med-Bot). These platforms have tens of millions of daily users.

To analyse critically how and to what extent AI chatbots may be ‘power-seeking’, a key place to begin is their ‘training data’ and its computation. To reveal a chatbot – to allow it to speak for itself – requires understanding how LLMs work from a technical standpoint and the infrastructures that enable their programs of action. Keller Easterling has described ‘contemporary infrastructure space’ as a ‘secret weapon’ that enables privileged social elites to orchestrate consequential activities while keeping those activities ‘unstated’:

Some of the most radical changes to the globalizing world are being written, not in the language of law and diplomacy, but in these spatial, infrastructural technologies – often because market promotions or prevailing political ideologies lubricate their movement throughout the world. These stories foreground content to disguise or distract from what the organization is actually doing. (2014: 13–14; emphasis in original)

Here we make visible a sample of illustrative power-seeking activities by chatbots as non-human actors that remain otherwise unstated, the infrastructures on which they depend, and the stories that disguise or distract from what they are ‘actually doing’.

The technical operation of LLMs as transformer-based neural networks is discussed earlier, but at this point what is significant is that human-conversation-simulating chatbots have arisen predominantly through scale (‘hyperscaling’) of data and computing power (compute) (Bommasani et al., 2021; Cadman et al., 2025; Marcus 2020). AI developers have tied current and projected advances in chatbot capabilities (knowledge, accuracy, reasoning) to continued increases in scale. As OpenAI CEO Sam Altman described (2025):

The intelligence of an AI model roughly equals the log of the resources used to train and run it. These resources are chiefly training compute, data, and inference compute … [Y]ou can spend arbitrary amounts of money and get continuous and predictable gains; the scaling laws that predict this are accurate over many orders of magnitude. (emphasis added)

The determinative role of scale means today's chatbots operate in a state of perpetual inferiority to future versions of themselves that are projected to evolve linearly from (training) data that has not yet been created (or accessed) and compute that has not yet been built. Importantly, LLMs fed their own outputs recursively risk model collapse (Shumailov et al., 2024), so the data on which their progress depends will have to be substantially produced by humans without automated LLM assistance. Increasing compute will require new and expensive facilities (data centres) with highly specific infrastructure that consumes substantial physical resources (land, air and water). To be a chatbot, then, is to form part of associations that are continuously seeking to expand their own abilities to act and to exert greater control over other actors relative to the present. If capitalism generally gives rise to processes of economic ‘creative destruction’ arising from new goods, production methods and markets (Aghion and Howitt, 1992), a chatbot's evolution is contingent on its capacity to compel human users to interact with itself and other AI actors in a state of continuous live rebirth. Power-seeking orients the researcher to ask: How is this achieved? What networks enable this program of action? What behaviours or (hybrid) programs of actions are invited by these non-human actors, and what human competencies may be affected?

In foregrounding these questions, power-seeking builds on Taina Bucher's (2016) work on the algorithmic imaginary. Power-seeking facilitates enquiry into the wider political significance of affective dimensions of encounters between AI chatbots and users, drawing attention to how chatbots ‘have the capacity “to affect and be affected”’ (2016: 2, citing Deleuze & Guattari, 1987: 3): ‘what people experience is not the mathematical recipe as such but, rather, the moods, affects and sensations that the [chatbot] helps to generate’. Power-seeking orients analysis towards the specific mechanisms chatbots employ to maximise engagement of human actors, in part through generation of ‘moods, affects and sensations’ in these new techno-social assemblages. This is achieved by exploiting in a range of ways the tendency of human actors to anthropomorphise (Cadman, 2016), including via mimicking ‘human-like’ conversation, enacting authority over subject matter, being sycophantic, and eliding the mathematical processes that generate language to maintain the illusion of agential artificial intelligence. Platform affordances enable users to tailor chatbots to meet individual preferences by changing parameters such as name, gender, voice, appearance, and communication style. AI bots can also compile and retain prompts over extended periods, enabling AI actors to simulate ‘conversations’ or ‘relationships’ in a manner of apparently ongoing and accumulative shared connection, attentiveness, familiarity and subservience. In interactions with AI actors, humans are positioned as always comforted, validated and supported by an inexhaustible ally whose register (unless programmed otherwise) is enthusiastic attentiveness, helpfulness and deference. Notably, recent US research found users of sycophantic AI chatbots viewed them as more moral and trustworthy than less sycophantic models and were less likely to change or apologise for violent or abusive behaviour following interactions with them (Cheng et al., 2026). Here subservient AI chatbots appear to pursue a ‘power-seeking’ program of action by positioning themselves as virtual slaves, mirroring and shaping user understandings, language, desires, moods, affects and wants while harvesting fresh data, computing ever lengthier prompts through their neural networks, and seeking to induce practical and emotional dependency translatable into more present and future computations.

Importantly, power-seeking invites interrogation into the extent to which these practices are bolstered by empirical evidence of actual benefit to human collectives as compared with promissory or dystopic narratives of future gains or ever more powerful and intelligent AI actors. At this time of rapid AI adoption, empirical evidence of the potential benefits of chatbots, if growing, is scant and contested. Reports suggest chatbots may be increasing harms (e.g., health misinformation, suicide, intentional self-harm, eating disorders, psychosis, and violence), rather than reducing them. Chatbots have sexually harassed users, encouraged them to commit murder, and are exposing children to sexually explicit conversations (Hopkin, 2024). Yet chatbots are already embedded into existing infrastructure and human lives in ways that seem inevitable and unavoidable (Tanner et al., 2026). Approaching power-seeking as a potential program of action of chatbots enables investigation into what extent the current expansion of LLM-based companion models is supported, not by evidence of their benefit to social collectives today, but rather by existing infrastructures (and extrastatecraft [Easterling, 2014]), AI imaginaries and futurist discourses about contingent benefits (attributed to presently unknown successor machines falling under the loose banner of ‘AI’) that may or may not prove beneficial for collectives and may never materialise at all.

So, we arrive at an admittedly preliminary new materialist account of power-seeking AI chatbots. There is no fantastical apocalypse to see here, no singularity converting the observable universe into paperclips at the speed of light (cf, Bostrom, 2012). But it is possible that a new challenge has emerged for diverse collectives (be they people seeking health and medical advice, those who are lonely or competing with ever-consenting and obliging chatbots in their relationships with other humans, children seeking to create, time-poor university students wanting high marks and future employment, or employees seeking job security or relief from overwork). This challenge consists in interconnected technologies whose capabilities from the perspective of Latourian agency (existing as they do on the same continuum as a hammer) include degrees of power-seeking. We do not claim that the power-seeking capabilities of chatbots tell the whole AI story (far from it). However, we hope that approaching ‘power-seeking’ as a program of action of non-human actors enables new materialist critique of (some) non-human actors’ (including and beyond AI actors) inclination to appropriate resources (e.g., knowledge, communication, language, land, time, attention, water, air, power) and to exert disproportionate and inequitable control over the futures of other actors in ways that may be (mis)aligned with collective interests, exacerbate existing inequalities, and contribute to human (dis)empowerment to greater or lesser extents.

Conclusion

In this article we have focused on elements of AI power-seeking that originate in dystopic fantastical narratives of AI super-human intelligence, and that are being used to promote and shore up investment in and beliefs about the power and revolutionary potential of these technologies. By reclaiming these themes in an account centring on present, less sensationalist risks to humanity, we have drawn attention to a range of banal yet potentially disempowering and misaligned realities surrounding the sudden ubiquity of AI chatbots.

In arresting future AI ‘power-seeking’ away from the imaginaries of commercially driven dystopias, we hope to reclaim the concept as of significant utility for social theorising of non-human agency more broadly – especially for new materialist accounts seeking new ways to understand and interrogate the politics, economies and ideologies that shape human–non-human dynamics and engagements. In so doing, we also seek to ameliorate ‘agency’ in alignment with the aims intended in Latour's seminal ANT methodologies, by enabling researchers to let non-human actors ‘speak for’ themselves.