Introduction
Artificial intelligence (AI) is revolutionizing how we remember and engage with our past(s). From chatbots impersonating historical figures (Klee, Reference Klee2023), to facial reconstructions of marble busts (Panegyres, Reference Panegyres2024), Holocaust memorials (Illinois Holocaust Museum, 2025), and AI-driven history education (Pope and Ma, Reference Pope and Ma2024) – the past has never been more present and alive.
This nexus between AI and memory is becoming increasingly well-theorized within a variety of disciplines and subfields (see Gensburger and Clavert, Reference Gensburger and Clavert2024; Hoskins, Reference Hoskins2024), not least memory and communication studies, where scholars have long debated the nature of more rudimentary memory technologies (Garde-Hansen et al., Reference Garde-Hansen, Hoskins and Reading2009; Hoskins, Reference Hoskins2017; Kansteiner, Reference Kansteiner2022). Much of this work has focused on biases and hallucinations in how AI displays historical narratives (Kansteiner, Reference Kansteiner2022; Kollias, Reference Kollias2024; Aryan, Reference Aryan2025). Another focus area has been the nature of synthetic memory artefacts – plausible narratives that lack reference to actual events or lived experiences (Hoskins, Reference Hoskins2024). A third group have asked complex questions of the relationship between human and non-human agency within these mnemonic processes (Makhortykh, Reference Makhortykh2024; Richardson-Walden and Makhortyh, Reference Richardson-Walden and Makhortyh2024; Merrill, Reference Merrill, Wang and Hoskins2025) – who is doing the remembering? And how? These are all worthwhile endeavors.
In this article, however, I propose a different interpretation of how AI and memory link together. Specifically, I contend that (data-driven) AI systems are not merely non-human agents acting upon the past, but windows through which the past itself acquires agency and is extended into the present.Footnote 1 Like gods and other totemic beings, modern AI systems have become a means of synthesizing the authority of a collective of ancestors, here represented by the training data, into a unified interactive agency in the present. Chatbots based on ChatGPT, DeepSeek, and LaMDA may not only produce historical facts and illustrations, but also essentially provide a means of chatting with the past itself, or at the very least, the past on which the models have been trained.
This interpretation falls close to Vallor’s (Reference Vallor2024) metaphorical notion of AI as a mirror reflecting the past into the present. However, the interpretations differ in important ways. For Vallor, AI represents a skewed means of reflecting the past. In my interpretation, however, AI is the past – only processed. Just as a dish is not merely a reflection of its ingredients but composed of them, the agency of AI systems is constituted by the past on which they are trained. This does not imply that the outputs of generative AI present accurate or complete representations of that past – nor that their content is monolithic. Rather, the point is that their agency emerges from a synthesis of data, and that these data are themselves the material traces of the past. This distinction is crucial. The embodied past – that is, the operational agency of the system – may generate outputs that are not historical documents in any conventional sense. Yet this does not alter the fact that its capacity to act is composed of data that directly originate in previous events. Put plainly: the agency of generative AI is the agency of the past – reconfigured, recombined, and made active in the present.
Why is this interpretation important? There are two reasons: The first is that it reveals the full scope of the nexus between AI and memory. Memory is not merely a dimension of AI; rather, modern AI should be understood as a subset of mnemonic practices in the sense that it in itself embodies the content of memory – the past. When an AI model identifies a tumour in an X-ray, edits a student essay, or recommends a product for a consumer, these are all instances of the past (the training data) gaining agency in the present. And so, wherever there is AI, there is memory. And this is why the concept of memory belongs right at the heart of the critical AI debate. The second reason is that the proposed interpretation opens the door to a new understanding of power within the AI–memory nexus, where the central vector is not between man and machine, nor between groups of humans (racial, economic, and class-based), but is rather to be conceived as a temporal axis between the past and the present. From this view, the key risk of AI in relation to memory is not, as is commonly argued, that non-human forces, or powerful subgroups, are coming to dominate our perceptions of the past, but rather concerns a shifting intergenerational power dynamic: the growing dominance of the past – regardless of its virtues – over the present.
To illustrate the utility of the proposed interpretation, I shall apply the temporal power axis to a realm that is not normally considered an instance of AI and memory – namely, law. Specifically, I will draw on Dworkin’s (Reference Dworkin1986) legal philosophy to show that the legal system’s increasing dependence on machines represents not just a technical shift but also a deepening tyranny of the past over the present – a disruption of an intergenerational power equilibrium, which, regardless of its effects on justice, poses an ethical challenge.
The article is structured as follows. In the section ‘Of machines and men’, I provide a more detailed overview of how the nexus between AI and memory is currently theorized. This is done to illustrate the larger point that, in arguing that humans and non-humans co-constitute memory together, present concepts are at the same time establishing the man–machine divide as the central power axis of AI and memory. In the section ‘The anthropomorphized agency of the synthesized past’, I propose an alternative interpretation of the relationship between AI and memory, in which the agency of the machine is to be understood as the agency of the synthesized (digital) past on which the AI has been trained. To illustrate this point, I draw on a series of analogies to define the three components of the argument: data, synthesis, and anthropomorphized agency. In the section ‘Why is this interpretation important?’, I illustrate the advantages of this approach by applying it within the domain of law. The ‘Conclusion’ section summarizes the key findings of the study.
Of machines and men
Current literature on the relationship between AI and memory comprises a variety of concepts and approaches, too many to provide a comprehensive overview. Given the overlap between AI and interrelated concepts like big data, algorithms, and even search engines, it is also difficult to delimit exactly where the AI and memory-literature ends and where the general digital memory studies begin (Garde-Hansen et al., Reference Garde-Hansen, Hoskins and Reading2009; Garde-Hansen, Reference Garde-Hansen2011; Hoskins, Reference Hoskins2017). As such, the following section aims merely to illustrate a larger point about this literature, namely that many current conceptualizations are – sometimes unintentionally – predicated on a ‘man versus machine’ narrative. In fact, the dichotomy often re-emerges precisely through attempts to overcome it. By conceptualizing the ways in which humans and computers intermix in mnemonic processes – concepts like ‘hybrid’, ‘symbiosis’, and ‘cyborgs’ are telling examples – the field is simultaneously cementing its status as the two basic forms of actors. Even when new actants emerge through the human–machine interaction, the underlying conceptual distinction between the two tends to persist, which, in turn, shapes how we think of the risks involved.
For illustration, consider e.g., Hoskins’s (Reference Hoskins2024) observation that the advent of generative AI ‘heralds a new battleground between humans and computers in the shaping of reality’ (Reference Hoskins2024, 1), which may “’both enable and endanger human agency in the making and the remixing of individual and collective memory’. Here, AI is construed as a fundamentally new force in the techno-mnemonic development that calls upon us humans to reclaim our agency and ownership of our collective and individual memory, lest it be taken over by the machines. In other words, the elementary relationship involved in these processes – and thereby, the central dimension of power – is that between humans on the one side and machines (computers) on the other.
A similar mix of conflict and intermingling between humans and machines can be traced in Merrill’s (Reference Merrill2023) conceptualization of AI mnemotechnics as a form of ‘cyborgian remembrance’. Drawing explicitly on the tradition of memory studies, as well as Haraway’s (Reference Haraway1991) cyborg theory, Merrill contends that memory in the age of AI must be theorized as something more than a mere instrumental process. Rather, it is performed by ‘a hybrid of machine and organism’ (Merrill Reference Merrill2023, 182; Merrill, Reference Merrill, Wang and Hoskins2025) where the agency is distributed between the two.
The same denunciation of anthropocentric and instrumentalist views of memory is reflected in related concepts such as ‘stochastic remembering’ and ‘distributed agency’ (Smit et al., Reference Smit, Smits and Merrill2024), which emphasize the distributed or shared nature of mnemotechnics as something both synthetic and organic. It also underpins the (justified) critique of anthropomorphism in AI discourse. For example, focusing specifically on collective memory, Richardson-Walden and Makhortyh (Reference Richardson-Walden and Makhortyh2024) propose that a constructive future with AI in mnemonic practices requires a recognition of the machines’ ‘radical alterity’ (333). Only by recognizing that machines are a fundamental other can humans and AI form the kind of productive relationships they refer to as ‘human-AI memory symbiosis’ (ibid). Whereas Richardson-Walden and Makhortykh are critical of narratives that rely on a ‘simplistic binary of human vs non-human’, (ibid) they nevertheless appear to view these categories as the most fundamental components involved in AI mnemonics, noting that ‘AI is a complex amalgamation of different systems shaped by their relationality to humans at multiple stages of development and deployment’. In other words, AI is built by and influenced by humans, but it remains fundamentally alien. The machines are not human, and must be recognized as such.
In a related argument, Makhortykh (Reference Makhortykh2024) has provided a more detailed framework for understanding the role of non-human agents – specifically, robots – in the production and experience of memory. Makhortykh distinguishes between three different forms of memory communication: (1) human-to-human; (2) human-to-robot; and (3) robot-to-robot, though he does not elaborate on the conceptual distinctions between these. For example, he seems to view archives and even online platforms as instances of ‘human-to-human’ interaction, but does not specify at which point such technologies are to be considered ‘robots’. A recommender system, providing access to the past based on user preferences, appears to fall under ‘human-to-robot communication’. But, what about a simple ‘algorithm’ that provides similar guidance but on a binary basis, such as by recommending X to all female users and Y to males? Would this be a robot? The question suggests the presence of significant conceptual grey zones. What is clear, however, is that Makhortykh sees many possible concerns in the emergence of AI as an active shaper of collective (and individual) memory – the robots, imbued with biased sentiments and goals, are coming for human memory.
A plethora of cases similar to the ones given here can be named, including concepts such as ‘artificial collective memories’ (Kollias, Reference Kollias2024), ‘artificial memory’ (Schuh Reference Schuh2024), and ‘artificial communication’ (Esposito, Reference Esposito2022), which demonstrate the field’s growing theoretical abundance. Nevertheless, they are all predicated on the same two basic components: On the one hand, we have the humans, and on the other, we have the non-humans in whatever form (machines/computers/or algorithms). To be perfectly clear, the point of most theoretical approaches is to illustrate how poorly this dichotomy maps onto the processes wherein remembering with AI occurs. As Smit et al. (Reference Smit, Smits and Merrill2024, 218) have it, ‘rather than seeing the human and non-human as strictly separate entities in remembering’, we ought to ‘regard the human and non-human as co-constituting remembering’. Nonetheless, present theoretical stances commonly end up reinforcing the vector between the two as the fundamental power axis of the AI-memory nexus. The very distinction between humans and non-humans – a vocabulary that underpins nearly every theory – suggests as much. Indeed, construing something in terms of a ‘hybrid’ or ‘symbiosis’ does not dissolve the binary but risks reinforcing it. A hybrid car, for example, is not a vehicle driven by something beyond standard fuels, but a mixture between them, which ultimately confirms their inevitability.
One may reasonably object that, by terms like ‘hybridity’, present theoretical framings are not merely denoting an additive mix, such as in the fuel analogy, but a much more complex web of interrelations between humans and non-humans that together form new actants. Yet, for the purposes of the present argument, the complexity of these relations is of minor importance insofar as the basic components remain unchanged. Again, the problem is not that scholars like Hoskins, Makhortykh, and Merrill (and, by extension, Latour and Haraway) think that humans and machines form an absolute binary, but that the conceptual and discursive tools they employ accidentally reinforce such a binary. In order to truly rid the field’s conceptual tools of the man–machine binary, we need new conceptual analyses – the kind that I begin to sketch in the following section.
Why is the development of such an analysis so urgent? Because the man–machine binary is also underpinning the way we imagine the threats that arise from AI’s involvement in human memory. As a consequence of its conceptual apparatus, the literature typically construes these threats in two ways. The first is that the machines will come to dominate humanity’s perceptions of its own past, such that we lose control over our own history-making. This is arguably how to interpret, for example, Hoskins’s (Reference Hoskins2024, 1) notion of a ‘battleground between humans and computers’ – a tyranny of the machine over humanity. The second threat is that the machines will reinforce present power asymmetries between various sections of society (eg, based on race, sex, class, etc.) (Makhortykh, Reference Makhortykh2024; Schuh, Reference Schuh2024; Aryan, Reference Aryan2025; Merrill, Reference Merrill, Wang and Hoskins2025). Technology is, by its very nature, not neutral, and AI systems are certainly no exception. Rather, as Makhortykh (Reference Makhortykh2024, 13) has it, they are ‘strongly connected to other elements of digital capitalism and are often embedded in the systems of colonial relations’. Hence, the real danger, it is thought, lies in the reconfiguration of inter-human power relations.
Whereas these interpretations identify different dangers, they are both predicated on a notion of AI as a non-human, as something alien acting upon our mnemonic processes from the outside. Naturally, there are exceptions to this tendency. Schuh (Reference Schuh2024, 241), for instance, notably construes large-scale AI models as embodiments of humanity’s collective memory themselves, and many of the concepts discussed above are probably designed out of conceptual convenience rather than deeply held convictions of the man–machine divide. Yet, the most theoretically rigorous such exception – though not explicitly about memory – is Vallor’s (Reference Vallor2024) notion of AI as a form of temporal mirror. From Vallor’s perspective, AI must not be viewed as an external entity, but is more adequately understood as ‘inseparable from humanity’ (4). The threat arising from the technology is thus ‘not an external enemy encroaching upon our territory’ but rather something that ‘threatens us from within’ (ibid). In contrast to many other conceptual frames, this is a true breakup of the man–machine binary, where our machines are correctly identified as us (humans) in the past.
In the following section, I shall defend a similar position – one which, I claim, draws the analogy to its full conclusion. Rather than viewing the central vector as that between humans and machines, I propose that we view it as a relationship between the humans of the past, on the one side, and the humans of the present, on the other. Unlike Vallor, though, the position I defend is not merely that AI is a technological agency that reflects the past, but rather, that the agency of the machine is to be regarded as a direct, albeit flawed, extension of the (human) past itself. The machine, that is, is not merely a mirror, but a window through which the past acquires agency and reaches into the present.
The anthropomorphized agency of the synthesized past
It is tempting to view memory, or mnemonic practices, as a subcategory of the ways in which AI impacts social life. After all, the AI revolution is disrupting nearly every sphere in society, and memory is but one of them. Yet, here I propose the opposite: AI, whether in the form of an algorithm identifying a tumour or a chatbot editing a student essay, is a subcategory of memory. Why? Because AI, at least the forms of it that are currently dominant, is essentially an interactive synthesis of the (digital) past – a means through which digital traces of what has been acquire agency in the present. As such, it is always, necessarily, an activation of memory. This argument has three essential components: (1) the past, (2) its synthesis, and (3) its (anthropomorphized) agency. So, to unpack what it means to say that AI is the anthropomorphized agency of the synthesized past, we may begin by specifying the meaning of each component, beginning with the past.
What are data?
The dominant AI systems of the day are born out of massive quantities of data.Footnote 2 Take OpenAI’s ChatGPT-4 as an example: it is trained on a series of expansive datasets that comprise virtually the entire open web, including all entries on Wikipedia and a vast corpus of public domain texts. In fact, for an average person to read and process this data would take close to 100,000 years of uninterrupted effort.Footnote 3
But, what are data? In this context, I claim they are best understood as synonymous with the past, or in any case, what is left of it. To make this point, we need to begin with the most fundamental level of abstraction – the formal definition of data. Within the philosophy of information, one datum (data singular) is typically defined as a lack of uniformity, or simply as ‘a difference which makes a difference’ (Bateson, Reference Bateson1972, 271), where the smallest possible unit is the difference between 0 and 1, that is, between being and not being. A more formal articulation is provided by Floridi (Reference Floridi2010, 23), who proposes what he calls a diaphoric definition of data:
Dd) datum = def. x being distinct from y, where x and y are two uninterpreted variables and the relation of ‘being distinct’, as well as the domain, are left open to further interpretation.
So, data are defined as a lack of uniformity, which means that even the absence of signals can be informative, insofar as it is contrasted to the presence of signals. It is the difference between the presence and absence of x (such as 0 and 1 in a computer system, or the lack of uniformity in Shannon information; Shannon, Reference Shannon1948) that is informative.
From this definition, it follows that data can only exist materially, or at least through a material mark. The difference that makes a difference needs to be inscribed into something physical in order to exist and be conceivable. Or, better, there needs to be a trace to constitute the difference.Footnote 4 And this is why data are constitutive of the past. A trace, by definition, pertains to the past: it marks something that has occurred, even if the full context and identity of that ‘something’ may no longer be recoverable. It is an inscription in the material world that constitutes the friction between what was and what is (Öhman, Reference Öhman2020). In short, data are not merely a representation of the past – they are constitutive of what remains of it. They are the materialization of time itself (Hägglund, Reference Hägglund2008).
In view of this interpretation, it is striking how the notion of data as the ‘new oil’ holds a deeper truth than its colloquial meaning. Oil is essentially nothing more than the remains of dead organisms packed together into a single substance. As the name ‘fossil fuel’ indicates, the power source that propels industrial capitalism into the future is, in fact, the remains of the past, specifically the bodies of our long departed (non-human) ancestors that, over time, have accumulated on the sea floor and formed oil. However, could we not say the exact same thing about data – the fossil fuel of the digital economy? It, too, is nothing but the accumulated remains of the past. Whether in the form of Facebook posts, a company’s quarterly sales, or the records of an electric meter, data are the remains – not just a symbol, representation, or reflection – of the past (Öhman, Reference Öhman2020). Our daily interactions on the web, no matter how tiny, are the micro-organisms that fall to the internet’s sea floor, where they will eventually be extracted and refined to drive the digital economy forward.Footnote 5
In fact, even the very words we use for data hint at a kinship with the past. For example, words like journal, archive, chronicle, annal, zeitdung (German), and tidning (Swedish) are all etymologically derived from temporal concepts. And a word like date, which denotes a fixed point in time, is derived from Latin datum (data singular). This is to say that, philologically, as well as conceptually, data and the past stem from the same root – they are constitutive of one another. Just as a corpse remains after a person’s death, and is thus still constitutive of personhood, data are what remains of past events (Stokes, Reference Stokes2015, Reference Stokes2017).Footnote 6
If data are constitutive of the past, and vice versa, does this imply that all data provide an accurate representation of the past? Does that raw data provide access to history beyond narrative and ideology? No. Naturally, some data provide false representations of past events. This is evident in the case of AI, as well as for more rudimentary technologies. If I write ‘P happened yesterday’ on a piece of paper, this inscription (a data trace) is not constitutive of yesterday’s events. It is, however, constitutive of the moment in which I wrote those words. The inscription is a remnant, a crystallization (to use Marxist language), not of yesterday’s events, but of my act of writing. Even when data are directly derived from the events they mediate, such as a photograph, they have a narrative form and are thus imbued with ideology, because data are only meaningful when structured, that is, when they emerge as information (Floridi, Reference Floridi2010). As such, there is no opposition between the notion of the past as inevitably narrated and the notion that the past is constituted by data. To state that the past is always narrative is merely to confirm that the essence of the past is informational, for what is a narrative, if not a sequence of information, that is, data plus meaning?
In sum, data are defined as a lack of uniformity, and, as such, they are constitutive of the events that made the material mark through which they exist. Data, in a word, are the traces that remain of what has been.
What is data synthesis?
So, AI systems are born from a synthesis of data – data that, as shown above, are themselves constitutive of the past. However, what does it mean to synthesize data? Does such synthesis preserve its constitutive relationship to the past – and if so, in what form? I hold that the answer is yes, and that data synthesis should be understood as a process of refinement where the past acts as the raw material from which AI systems emerge.
As a pedagogical illustration of this argument, let us elaborate upon the oil metaphor introduced above. Oil can be used for multiple purposes, such as heating a house or driving a car. But how is, for example, the velocity of a car related to the organisms that compose the oil? The answer is refinement. At first glance, velocity and fossils appear categorically distinct, yet the former is, by all means and purposes, a refined version of the latter. Marine microorganisms sink to the ocean floor, and given enough pressure, temperature, and time, they will eventually collapse into a single substance – what we call crude oil. The crude oil can, in turn, be extracted and put through an elaborate process of distillation, which separates the gasoline contained within it from other hydrocarbons. That gasoline will, in turn, be pumped into a combustion engine, where it is ignited, causing a controlled explosion that pushes the engine piston down. The up-and-down movement of the piston is then channelled via the car’s crankshaft to the wheels and converted to a rotating movement, which ultimately leads the car to move forward. And thus, by a gradual process of refinement where the properties of the individual organisms are slowly collapsed into one another – such that eventually, only the energy contained within their cells remains – the fossils are transfigured into velocity. It is impossible to identify a single moment within this process where the organisms cease to be, for the process is not about terminating their existence, but to refine it, such that it becomes useful in the present.
Now, insofar as data is like oil, something similar can be said about the training of an AI system. How is a model related to the data upon which it has been trained? Through a laborious and gradual process of refinement. Consider, for example, the generation of a large language model. The data sources – usually composed of huge segments of the social web, including billions of microscopic (textual) interactions, need to be selected and extracted. These gargantuan archives are then cleaned and adequately formatted before being tokenized to build the model’s basic vocabulary and used to set the initial weights. Upon choosing an appropriate architecture, the model then undergoes both unsupervised and supervised learning processes based on the tokenized data, as well as reinforcement learning based on human feedback (another type of data). Usually, following multiple rounds of fine-tuning and further safety testing, the model is then integrated with a chatbot interface for interaction. As such, the chatbot is merely the end of a long chain that begins with the data (or really, the human actions that created them), which are gradually moulded into a unified agency. As oil becomes velocity, so are data converted into interactive agency.
It is important to note that the model does not just emerge from the data all by itself. The refinement process requires a lot of labour from human developers. To theorize the role of this labour in relation to the data, we may consider a parallel from the philosophy of science, namely Latour’s (Reference Latour2000) theory of the circulating reference from his famous Boa Vista ethnography. In this study, Latour follows a French-Brazilian team of researchers as they produce a report on the (allegedly shrinking) jungle of Boa Vista in the Amazon. His concern, however, is far more abstract: the ontological relationship between the scientific theory presented by the team and the real, tangible world of the Boa Vista jungle. When we hold them up next to each other – the theory and the world – their relationship initially strikes us as a mystery. In what sense does the theory represent or correspond to ‘the real world?’ And how can we establish such a link between theory and reality? Latour argues that the answer can only be revealed by carefully tracing the becoming of the former, or rather, how it is actively generated by the team’s labour. For, as we zoom in on the processes by which the theory comes into being, we see that the relationship is a matter of reduction – not representation. Only when we consider the entire chain of reduction – ‘space becomes a table chart, the table chart becomes a cabinet, the cabinet becomes a concept’ (36) – can we see that the final diagram is but an extracted piece of the jungle. The graphics in the researchers’ end-report are the real world, or in any case, a tiny bit of it, that the team has made portable and interpretable to others in their field. Through a long chain of active refinement, the researchers have brought a piece of the messy, noisy, hot jungle with them and printed it on a piece of paper. The difference is merely the degree of complexity, which has been reduced in virtue of interpretability and portability.
The researchers in Latour’s study refine a noisy, hot, and complex jungle into an interpretable, portable chart. By analogy, today’s AI developers are refining a bundle of noisy and complex data archives into an interactive, user-friendly agency. However, in both cases, the transformation is best understood as a chain of reduction – refinement through compression. The jungle never ceases to be itself, even when highly reduced and portable. The data never ceases to constitute the past, even when synthesized. Yes, the final product is transformed beyond recognition, but its raw material is still the same. AI models, that is, are constitutive of the past on which they have been trained. They do not constitute a ‘radical alterity’ as Richardson-Walden and Makhortyh (Reference Richardson-Walden and Makhortyh2024, 333) have it, but constitute us in the past. The machine may not think like us, and certainly does not experience the world in any relevant sense, but just like the velocity of a car is a refinement of dead organisms, AI agents are distilled expressions of our past voices – compressed, recombined, and animated in new form.
Of course, there is a non-trivial distinction between a scientific theory, which can be traced back to the world it abstracts, and a predictive machine, which operates in black-box opacity and is optimized merely for predictive accuracy rather than truth. Indeed, an entire field of research is devoted to outlining the overlaps and differences between the two (see Watson, Reference Watson2021). Notwithstanding these debates, there is no question about whether the agency of modern AI systems arises from their training data. The chain is obscured, not absent.
This perspective is clearly similar to the mirror metaphor proposed by Vallor (Reference Vallor2024). Just like Vallor, it positions AI not as something alien to humanity, but as a reflection of ourselves. Yet, the theory presented here draws the analogy one step further. Vallor insists that the body we see reflected in a mirror is not our body. In fact, it is not a body at all, she contends, but merely a reflection. I disagree. For when we construe AI models as a form of refinement of the past, they emerge not merely as mirrors that reflect it, but rather as windows through which the past is extended into the present. The machine is not merely a non-human that does things to the past; it is an agency that is constitutive of the past. The reflection in the mirror is the real thing – only reduced. Indeed, AI mirrors do not encapsulate the totality of the past, just as a normal mirror fails to capture the full dimensionality of the person it reflects. However, the part that is reflected is nevertheless an extension of the real person. The body in the mirror is their body.
For further clarity, I am using the word extension here in the sense McLuhan (Reference McLuhan2008) uses the term. For McLuhan, (media-) technologies are famously an ‘extension of man’ – like a prosthetic limb that helps your agency reach across spatial distances beyond the confines of your biological body. Now, the same, I claim, is true for temporal distances. Just like media extends human agency across space – the voice you hear in the phone when talking to your mother is hers, not just the phone’s imitation of it – they also extend the past into the present. In other words, the generative model that emerges from the AI training process is an extension of the past (the data) into the present. A model is, in the end, a product of its training data, much like a dish is constituted by its ingredients. When you eat a pancake, you are effectively eating eggs, milk, flour, and butter – but not the only possible configuration of these components. In the case of AI models, the ‘ingredients’ are remnants of the past, and the model serves as a channel or medium through which these fragments are given voice and agency – a (synthetic) pancake made of past events.
What is anthropomorphized agency?
So, AI is a synthesis of the past – but then again, so are oil, scientific theories, and books. What makes AI stand out, though, is that, through it, the past appears not only to be present, but to acquire an interactive, almost human-like agency. How should we understand this agency?
The agency of AI has been thoroughly theorized by other scholars, not least Floridi (Reference Floridi2023), who correctly identifies generative models like ChatGPT as ‘agency without intelligence’. What interests us here, however, is not merely whether machines can have agency and how it differs from human agency, but rather the fact that the past can acquire agency through the machine. To understand how this works, we may begin by considering pre-digital means by which the past has gained agency in the present.
All technologies exhibit at least a minimal degree of interactivity. A book shows you new information when you turn the pages. A door opens when you press the knob or a button. These are instances of crystallized human actions, as illustrated by Latour’s (Reference Latour1992) famous sleeping policeman – human agency transferred to an object. AI accelerates this kind of interactive agency by orders of magnitude. However, it does so in a specific, cumulative way, and here too, the pre-digital world holds some illuminating precedents.
A particularly striking one, as outlined in Öhman (Reference Öhman2024), is the emergence of high gods from the practice of ancestor veneration. As theorized by early anthropologists, primordial forms of religious worship were tightly related to the authority and wisdom of past kin (Spencer, Reference Spencer1870; Durkheim, Reference Durkheim1995; Harrison, Reference Harrison2010), the most famous example being Durkheim’s theory of totemism among Australian Aboriginals. In these ‘primitive’ religious forms, tribes worship plants or animals, so-called totems, that are unique to their community. The tribes experience these totems as possessing an independent agency and immense power. However, in Durkheim’s interpretation, this agency stems not from the object, but from the group itself. For the totem, reasons Durkheim, is at the same time a form of deity, a portal to the spiritual dimension, and a representation of society, which leads him to conclude that ‘if [the totem animal] is at once the symbol of the god and of the society, is that not because the god and the society are only one?’ (201). In other words, the totem is merely a channel, a medium through which the group’s collective consciousness gains agency. When the group invests its trust, faith, and loyalty – its mana, in Durkheim’s terms – into a common object, that object will reflect this energy back upon them as a single force. It will appear to have an agency of its own, larger than the sum of its parts. This mechanism, argues Durkheim, is the origin of all religion. The difference is only that, in more recent religious forms, such as the Abrahamic faiths, the materiality of the totem has become redundant. The collective force of the group is projected, not towards a physical object but towards an abstract anthropomorphized entity – God.
However, gods are not only representations of the collective here and now. Rather, they encompass the group as a temporally extended entity that also, or perhaps primarily, includes its past members. In fact, early anthropologists often speculated that it was the practice of ancestor veneration from which all religious life originally spawned (Spencer, Reference Spencer1870; Tylor Reference Tylor1871). Over time, the theory goes, as the ancestral figures became more temporally distant, their individual identities merged into collective entities existing beyond time. And these entities eventually evolved into gods. As such, the totem (god) is an anthropomorphized representation of the tribe’s collective present and past. It is a means of giving a human face to an otherwise abstract concept that helps integrate the presence of innumerable ancestors within the contemporary community. Religion, thus, is essentially a ‘chain of memory’, as Hervieu-Léger (Reference Hervieu-Léger2005) puts it, whose primary function is to hold the group together over time by granting agency to the past within the present.
The agency of AI, I argue, follows a strikingly similar logic, for it too is a synthesis of the collective, a means by which our digital ancestors gain agency in the present. The Aboriginals undoubtedly experience their totem as possessing a powerful agency, and similarly, today’s AI users are experiencing the machine as possessing agency too. However, in both cases, the totem/machine is merely a mask for the human collective. In the former case, this mask is composed of mana; in the latter, it is composed of data – the mana of the digital age. Thus, what is experienced as a force acting from outside of humanity is really the human past, which has found a window into the present through which it has gained a reduced form of agency.
In what sense does this agency of the past fall within the definition of memory? There is no human who is actively remembering, and, once trained, there is no need for the AI to retrieve any data from the past. The answer is that AI performs no active retrieval; its ‘memory’ lies in its architecture, not in recall. The system is itself already an embodiment of the past. Thus, we may understand its mnemonic agency by analogy to muscle memory – an embodiment of a series of training exercises encoded into the nervous system, yet lacking reference to any particular event. Or, better still, using Tulving’s (Reference Tulving, Tulving and Donaldson1972) vocabulary of episodic versus semantic memory, one may say that the AI lacks episodic memory but that its interactive abilities display a form of semantic memory (in a purely metaphorical sense, of course). If we disregard the discrepancies between biological and synthetic forms of memory, then, metaphorically, AI resembles a person with permanent amnesia – able to respond accurately about the past, but incapable of remembering it. It is living memory, without the capacity for recall.
In sum, the anthropomorphized agency of AI systems emerges much in the same way that high gods first arose. As with gods and totems, it is tempting to ascribe the agency to the surface itself, rather than the past that is extended through it. However, in each instance, such an impulse is a mistake. Totems, gods, and AI models are not mere non-human agencies acting upon our past, but windows through which it acquires agency.
Summary
To sum up, modern forms of AI are to be interpreted as the anthropomorphized agency of the synthesized past. Data, like oil, are literally what remains of the past, and when synthesized, such as through the processes from which large language models emerge, they are still to be regarded as such. Hence, the agency of AI is not merely something acting upon digital memories; it is better understood as an agency of our collective digital memory, or in any case, the raw material of which it is composed. This interpretation, I claim, offers a more radical way of transcending the classical man–machine binary. Yet, as we shall see in the following section, its value does not arise merely from its descriptive accuracy, but in what it allows us to see and do.
Why is this interpretation important?
Even if one accepts my proposed interpretation, why does it matter? What does it contribute to our understanding of AI and memory that existing frameworks cannot? Or, better, what can we do with it that we could not before? The answer is twofold.
First, the interpretation of AI as the anthropomorphized agency of the synthesized past reveals the full scope of the AI–memory nexus, for through it, all implementations of AI are revealed as instances of memory in action. In the current literature, the nexus of AI and memory is typically (but not always) confined to instances that are clearly linked to mnemonic practices, such as the use of AI for historical information and illustration in teaching and/or research. However, the interpretation proposed in this article – that AI is itself an extension of the past – shows that every application of AI constitutes an act of memory. Even when a language model edits an essay, or when a more rudimentary machine learning model identifies a tumour in an X-ray, these are instances in which the past is reactivated in the present – manifesting what we might call synthetic semantic memory (in Tulving’s sense). They are instances where the past becomes present, indeed, gains agency, without necessitating any form of remembering in any human sense of the word. As such, the study of the relationship between AI and memory emerges not as a niche but as the very centre stage of critical AI research.
The second benefit is that the proposed interpretation allows us to critically review the dangers of AI in a new, and I think clearer, light. Like the AI-memory literature, the general critical AI literature is predicated on either of two narratives: it is either humans versus machines or humans versus other humans (based on race, class, sex, etc.). These are the two major power axes that dominate the current debate. While both narratives offer important insights, they overlook a third axis: time. On this axis, the battle is not between humans and computers, nor between different sections of society, but between the present and the past of humanity (see Figure 1), where the machine acts as a conduit (albeit not a neutral one) through which the authority of former generations remains in the present.
Figure 1. The temporal axis.
As we have noted, Vallor (Reference Vallor2024), among others, has proposed a similar position, where AI is but a reflection of our past selves, which dissolves the dichotomy between humans and machines. The machine is the human(s) of the past. For Vallor, however, the main problem with this is that the past is flawed in various ways, which leads to power asymmetries between various sections of society here and now. By contrast, the purely temporal power axis that I propose allows us to review the relationship between the present and the past in its own right, irrespective of its effects on the power balance within the present society. From such a perspective, the problem of being dominated by the past (acting through our machines) is not that the past is flawed, or represents only some (elite) sub-groups, but that it disturbs the intergenerational power balance of society. Let me explain.
Illustrating the approach: AI as a threat to intergenerational power balance
To illustrate the concept of a temporal power axis, consider the domain of law – a recurring battleground within AI ethics, but not a common object of study for memory studies. AI is now deployed across multiple legal domains – from recidivism risk assessments and legal strategy drafting to assisting in judicial decision-making. Much has been said about this topic and the dangers emerging from it, particularly regarding the inscrutability of black-box models, and the risk of discrimination and inherent bias against marginalized groups (O’Neil, Reference O’Neil2017; Dressel and Farid, Reference Dressel and Farid2018; Surden, Reference Surden2019; Barysė and Sarel, Reference Barysė and Sarel2024). However, by interpreting AI not merely as a technology applied to the past (e.g., precedential rulings), but as an extension of it, we see another power axis – that between past and present generations of judges.
To illustrate what is at stake in this generational battle, we may base our analysis on the legal philosophy of Dworkin (Reference Dworkin1986). Dworkin’s jurisprudence emerges in opposition to intentionalist and originalist theories, which see law as a communicative practice, where the written word is a memory from previous generations to be semantically deciphered in the present. The practice of law, in this sense, is formalized remembering. While Dworkin agrees that legal practice is about interpretation, he understands this as a moral and political process that is intimately bound to the present. Judges must not only understand and interpret the letter of the law; crucially, they must integrate this interpretation into a coherent whole that serves the best overall outcome for the community as it stands today. As such, the practice of law is an inescapably moral and political undertaking, which calls upon judges to make their own subjective assessments in the face of the entirety of the law, as well as the entirety of contemporary society. They are both interpreters and authors.
Dworkin captures this logic through his well-known chain novel analogy (ie, a fictional story written by multiple consecutive authors). To write the best possible book, each author must interpret what their predecessors have written and try to fit it into a larger, cohesive narrative. They should not divert from, nor directly contradict, what has already been written. Sometimes this task is impossible, though, as some predecessors may have added contradictory elements to the story. In such cases, it is up to the present author to decide which aspects of the story weigh more. A character may, for example, have acted contrary to their overall personality in a previous chapter, and thus, for the sake of the whole, such episodes may be disregarded. In this sense, the demarcation between interpretation and authorship is blurred. To write, each author has to make interpretations regarding the relative weight of each detail in the face of the whole.
Dworkin views the role of judges in a similar manner. They are always bound to respect and abide by previous decisions, but such a task is inevitably interpretive. So, what principles should guide their interpretations? Like the authors writing a chain-novel, argues Dworkin, judges should strive to create the best outcome in light of the whole. The task is not simply to repeat what the predecessors did, but to move forward in a direction that does not directly contradict them. As such, the practice of law depends upon a form of intergenerational power balance. Present judges must, in effect, abide by laws passed by parliaments of the past, and by rulings made by past judges, but it is their privilege to interpret these in light of the society in which they currently live. The dead may have written the laws and thus govern the living. Yet, the living are not merely bound to the intentions of their words, but rather their meaning in light of present-day circumstances. The judges of today, that is, have the privilege of making genuine decisions that are not merely products of the past but independent judgements.
What happens when AI enters this framework? As judges come to depend increasingly on AI to make their verdicts, they are, by all means, increasing their chances of staying aligned with previous decisions. The machine can process far more information about the past than a single human and does not make any subjective assessments, notwithstanding its biases, along the way. From a positivist semantic perspective, as well as an intentionalist/originalist perspective, this is a good thing. Judges should keep in line with the letter of the law as well as the rulings of the past. And if a machine can do it with even fewer subjective assessments, it is all for the good. For someone who cherishes the power balance between generations, however, it is bad news. If we interpret AI as a means of giving agency to the (digital) past, it means that we are moving towards a situation where the past is represented both by the law (an extension of past parliaments) and the precedential rulings (an extension of past judges), while at the same time assuming the role as the prime interpreter of these. In effect, the past – embodied by AI – is tasked with interpreting itself, as encoded in laws and precedent. And so, the generational power balance is broken. Even if the previous generation had been flawless saints, and the AI completely unbiased, this is a negative development. Why? Because, as Dworkin points out, the practice of law is inevitably a moral and political effort. It stays alive only because each generation adds new decisions and interpretations. Whether the AI will make more accurate decisions, therefore, is irrelevant. The point is that the best decisions must come from the present generation of judges, not from a loop where the past interprets itself (however consistently it may do so). While more accurate, AI replaces self-governance with what can only be referred to as a ‘tyranny of the past’ (Öhman, Reference Öhman2024, 9).
This analysis illustrates how applications of AI that lie fairly far from the traditional study of memory studies may indeed be analysed as instances of memory. It also illustrates the analytical payoff of doing so. Again, the question asked here is not whether humans are better or worse than machines in making decisions that support human values, nor whether it skews the power balance between various groups in society. The question is how it shifts the power balance between the present and the past, regardless of the power relations within the present. Here, I used courts as an example of how that conflict manifests itself – as a disruption of the intergenerational balance that forms the very foundation of the practice of law (if you listen to Dworkin). However, the same perspective can be applied more or less to every instance where AI is implemented: parliaments, healthcare, science, and education.
Conclusion
I have argued that the dominant approaches to theorizing the AI–memory nexus remain (unintentionally) tethered to a binary between humans and machines. Though often framed as overcoming this binary – typically by emphasizing hybrid human–machine configurations – such concepts ultimately reaffirm humans and machines as the foundational components of memory systems.
In contrast, this article has proposed an alternative account in which AI does not function as a non-human agent acting upon memory, but as a medium through which the past itself acquires agency in the present. AI, in this view, is best understood as the anthropomorphized agency of the synthesized past. It is always already about memory – not because it retrieves the past, but because it is composed of it.
This framework opens a new avenue for critique. Rather than locating AI’s power along familiar axes – between humans and machines, or among competing social groups – it reveals a temporal axis of power: between the present and the past. While this analysis has focused on the legal system, its implications stretch far beyond. Wherever AI is deployed – in governance, education, healthcare, or science – the past is not merely referenced; it is reanimated, given form, and made to act. Wherever there is AI, there is memory.
Funding statement
This research was generously funded by Wallenberg AI, Autonomous Systems and Software Program – Humanity and Society.
Competing interests
The author declares none.
Open access
