As-If Agents: Misrecognition and the Ethics of Non

Abstract What if the most influential voices in our epistemic lives are not agents at all, but machines we keep mistaking for them? I argue that contemporary AI systems function more and more like epistemic authorities while lacking the psychological resources that underpin human epistemic agency. Building on work on artificial epistemic authority and algorithmic truth, I show how search engines, recommendation systems and large language models are already treated as sources of knowledge in medicine, education and everyday deliberation. I then turn to the distinction between episodic and semantic memory in human cognition and argue that current AI only has the latter in a thin, decontextualized form. It has no episodic memory, no autobiographical perspective and no capacity to remember past episodes as ‘things that happened to me.’ This, I claim, makes our growing tendency to relate to AI as if it were a responsible testifier a form of misrecognition. I conclude by sketching an ethics of non-agentive AI: we should see these systems as powerful, instrument-like extensions of human cognition, not as knowers in their own right, and design institutions, interfaces and norms of trust accordingly. 1 Introduction Artificial intelligence is increasingly treated as a source of knowledge. People ask search engines, recommendation systems, and large language models what is true, what they should do, and how they ought to understand their own situations. In medicine, education, and everyday deliberation, such systems are often approached as if they were epistemic agents: entities that do not simply output information, but in some more robust sense know, remember, judge, and can be relied upon as authoritative informants. Yet, this raises the question of what kind of things these systems are, and what kind of epistemic standing they actually possess. A terminological clarification is important from the outset. In this paper, I use “agentive” and “non-agentive” in a philosophical rather than an engineering sense. I do not mean “agentic” in the now common technical sense of AI systems that can autonomously plan, call tools, or carry out multi-step tasks. A system may be highly autonomous in that engineering sense and still fail to qualify as an epistemic agent in the stronger sense relevant here. My concern is with agency as a form of epistemic standing: the kind of first-personal, diachronic, and answerable position from which a subject can not only generate outputs, but stand behind them as its own across time. The issue, then, is not merely whether current AI can act, but whether it can know, testify, and be trusted in the way a responsible epistemic subject can. I argue that there is a deep mismatch between the practical stance many users adopt toward current AI and the kind of systems current AI actually is. Contemporary AI can often function as a highly useful epistemic resource, and in some contexts, it may even qualify as an epistemic authority in a thin, reliability-based sense. But it does not follow that it is an epistemic agent in the fuller sense that normally constitutes interpersonal trust in human knowers. Current AI systems do not possess episodic memory, do not occupy a temporally extended first-person perspective, and cannot relate their present outputs to a remembered history of their own past judgments and experiences. My central claim is that this matters philosophically and ethically: when we treat such systems as if they were responsible epistemic subjects, we misrecognize them. By misrecognition, I mean a specific normative error: the attribution to AI systems of a kind of epistemic and practical standing that they do not possess, and the organization of trust, dependence, and responsibility around that mistaken attribution. To misrecognize AI in this sense is to treat it not just as reliable in some domain, but as something like a responsible testifier or knower whose outputs can be received in an interpersonal mode of trust. My suggestion is that much contemporary discourse and design encourages exactly this mistake. The problem is that social and institutional practices increasingly position AI as if it occupied the stronger status of an answerable epistemic subject. Yet, human knowers do not merely store facts. They remember particular events as events that happened to them; they draw on those remembered episodes in explaining why they believe what they believe; and they can, at least in principle, be asked to answer for their past judgments, revise them, and situate them within a continuing autobiographical perspective. Episodic memory is not the whole of epistemic agency, but it is a central condition of the temporally extended, self-relating standpoint from which human beings speak as responsible epistemic subjects. It is part of what makes not just having information, but owning one’s claims across time possible. Moreover, further clarification is needed. One might think that the deepest issue here is not episodic memory but experience. If a system does not undergo events in a first-person way, then it cannot straightforwardly have a personal past, and if it lacks a personal past, the absence of episodic memory seems to follow naturally. I do not deny that experience may be the deeper background issue. Indeed, much of what matters about episodic memory, its first-personal character, its connection to personal history, and its role in answerability, may depend on experience in this broader sense. Still, I focus on episodic memory rather than experience as such for a methodological and epistemological reason. Experience is a notoriously difficult and contested notion, whereas episodic memory has a more traceable cognitive and psychological definition. More importantly, episodic memory bears directly on the forms of justification, judgment, testimony, and diachronic self-relation that are central to my argument. So, while experience may lurk in the background as a deeper condition, episodic memory is the more precise and workable concept for the specific issue of epistemic authority. Current AI systems lack this kind of memory. Large language models store information in parametric form and process inputs within a limited context window, sometimes supplemented by external logs, retrieval systems, or conversation histories. These capacities may support impressive forms of semantic retrieval and pattern completion, but they do not amount to episodic memory in the sense relevant to human cognition. Such systems do not remember past interactions as episodes in a personal history; they do not mentally revisit what they previously did or thought; and they cannot cite particular past experiences as their own reasons for present claims. Even when AI researchers use the term “episodic memory,” they typically refer to engineered retrieval structures or event segmentation mechanisms, not to anything like autobiographical recollection or diachronic epistemic selfhood. I am not arguing that no artificial system could ever be counted as an epistemic agent, nor that all forms of epistemic authority require episodic memory. My claim is narrower: episodic memory is a necessary condition for the kind of robust, answerable, first-personal epistemic agency that would justify treating a system as a responsible testifier rather than as an instrument. Current AI may display thin authority by reliably indicating truths in certain domains, but it lacks the diachronic, self-relating perspective required for stronger forms of epistemic standing. Even if episodic memory were not the only conceivable route to robust epistemic agency, current systems do not possess anything sufficient to play that role. If that is right, then many of our current ways of speaking to and about AI involve a category mistake at the level of status. This category mistake becomes a problem when we think along the following lines: In ordinary epistemic life, authorities are often not merely relied upon as useful sources of information; they are trusted as speakers whose word carries standing. Work in the epistemology of testimony has emphasized that testimony is not exhausted by the reliable transfer of content. It is a distinctive source of knowledge bound up with the normative status of the speaker, the uptake of the hearer, and the wider practices through which reasons, trust, and credibility are distributed. This is one reason why the category of the “responsible testifier” matters. My claim is that the kind of trust increasingly extended to AI resembles the sort of interpersonal uptake more commonly associated with testimony. To the extent that AI is received not merely as an informational tool but as something whose “word” is taken to settle belief and guide action, the question of whether it can count as a genuine epistemic testifier becomes unavoidable. Furthermore, misrecognition is not merely a problem of anthropomorphic over-description. Miranda Fricker’s (2007) work on epistemic injustice makes clear that credibility and epistemic standing are socially structured, and that agents can be wronged in their capacity as knowers when credibility is allocated unjustly. My worry is not only that AI may be granted too much credibility, but that treating AI as a testimonial authority can help reorganize whose voices are heard, deferred to, or displaced. The concept of misrecognition therefore bears not only on the status of AI itself, but also on the wider epistemic environment in which authority and credibility are redistributed. If we treat AI as if it were a responsible knower when it is not, we risk overtrusting its outputs and under-describing the role of the human and institutional structures that make those outputs possible. We may begin to turn to AI in domains where the relevant epistemic tasks depend on lived experience, autobiographical perspective, judgment, or accountability for past error. At the same time, we may allow responsibility to migrate away from designers, deployers, and institutions by treating the system itself as the bearer of epistemic authority. Misrecognition, in other words, distorts both trust and accountability. I proceed in four steps. First, I show how AI systems have come to occupy the practical role of epistemic authorities in contemporary life, drawing on both philosophical accounts of artificial epistemic authority and empirical evidence of the trust we show AI outputs. Second, I develop the distinction between episodic and semantic memory, with particular attention to why episodic memory matters for temporally extended epistemic agency, justification, and judgment. Third, I argue that current AI systems lack episodic memory in the relevant sense and therefore cannot qualify as responsible epistemic testifiers, even if they remain reliable in narrower, instrument-like ways. Finally, I draw out the ethical implications: if AI is best understood as a non-agentive but powerful epistemic instrument, then our interfaces, institutions, and norms of trust should be designed to reflect that fact rather than conceal it. What is ultimately at stake is not whether AI can be useful. It plainly can. The question is what sort of thing AI is, and what kind of trust that sort of thing can warrant. My suggestion is that contemporary AI is increasingly authoritative without being genuinely agentive: it functions as an as-if knower, and our failure to recognize the limits of that status is becoming one of the central ethical and epistemological problems of human-AI relations. 2 AI as Seen as an Epistemic Agent Before asking whether current AI deserves the kind of trust it increasingly receives, it is first necessary to clarify the social and epistemic phenomenon the paper seeks to explain. Contemporary AI systems are not merely used as neutral instruments for retrieving information or automating routine tasks. In many domains, they are approached as if they were sources of knowledge and guidance. People ask search engines, recommendation systems, and large language models what is true, what to do, and how to understand their own situations, and they often treat the resulting outputs as authoritative inputs into belief and decision. This section does not yet argue that such treatment is justified. Its more limited aim is to show that AI now occupies, in practice, a role that closely resembles epistemic authority, and that this practical role creates the conceptual and normative problem the rest of my paper addresses. Recent philosophical work has already begun to conceptualize this shift. Rico Hauswald (2025) argues that AI systems can function as artificial epistemic authorities. His starting point is dissatisfaction with traditional accounts of epistemic authority that tie authority too closely to distinctively human features such as beliefs, intentions, or testimonial agency. If one assumes from the outset that only intentional human agents can be epistemic authorities, then AI can never count. Hauswald’s proposal is to loosen that assumption. On his view, what matters for authority is not necessarily the possession of beliefs or intentions, but rather the existence of a systematic asymmetry of competence such that it is reasonable for some users to defer to the outputs of others. In this thinner sense, authority is tied above all to reliable indication of truth and justified deference. Hauswald’s account captures something that is plainly true about current AI systems: users often defer to them because they treat them as more competent than themselves in specific epistemic domains. Search engines, ranking systems, and language models can therefore operate as authorities in the modest sense that their outputs structure what others come to believe. At the same time, Hauswald’s framework also helps make the distinction that will later become central to my argument visible: the fact that a system can occupy a role of thin, reliability-based authority does not yet show that it possesses the stronger kind of epistemic standing associated with answerable, first-personal, or testimonial agency. A complementary line of thought appears in Shin’s (2025) discussion of algorithmic truth. Shin’s claim is that AI-driven infrastructures increasingly shape the conditions under which claims become visible, credible, and influential in public life. On this picture, truth is no longer mediated solely through human testimony, institutional expertise, or ordinary public reasoning. It is increasingly filtered, ranked, generated, summarized, and validated by sociotechnical systems. AI is therefore not just an optional addition to an otherwise stable epistemic order; it is becoming part of the architecture through which contemporary societies produce and distribute what counts as knowledge. For my argument this is important because it shows that the issue is not simply a matter of individual users anthropomorphizing chatbots in isolated cases. The authority of AI is increasingly infrastructural. It is embedded in platforms, interfaces, educational routines, clinical workflows, and public information environments. Once this is recognized, the central question becomes: If AI systems increasingly function as authorities in how knowledge is circulated and taken up, then what kind of authority is this? Is it merely the authority of a reliable instrument, or are we beginning to treat it as the authority of something more like a knower? Norms of trust, dependence, and accountability differ depending on which of these roles a system is perceived to occupy.Footnote 1 Empirical research in behavioral decision-making and human-computer interaction supports the claim that users often grant AI outputs a practical authority. Logg et al. (2019) found that participants in several experiments adhered more strongly to advice when it was presented as coming from an algorithm than when the same advice was presented as coming from other people. They describe this tendency as algorithm appreciation. This shows that when content is held fixed, the fact that advice is presented as algorithmic can itself enhance its practical authority and suggests that AI systems can acquire de facto epistemic weight in deliberation quite apart from whether they possess anything like human understanding or judgment. The medical domain makes this especially vivid because the relevant knowledge claims are high-stakes, technical, and closely tied to trust. Tun et al. (2024), in a systematic review of trust in AI-based clinical decision support systems among healthcare workers, identify a worrying pattern. Clinicians are more willing to rely on such systems when they are shown to be accurate, when they provide explanations, when they fit existing workflows, and when they are endorsed by trusted institutions. In other words, AI authority in medicine is not raw or automatic; it is mediated by institutional legitimacy, interface design, and perceived competence. From the patient side, there is additional evidence that AI-generated medical advice is treated as epistemically weighty. Shekar et al. (2024) report that both lay participants and medical experts often rated AI-generated responses as more thorough and sometimes more accurate than physician-written responses, and that many participants were willing to follow the AI advice even when experts identified factual mistakes. Ruben et al. (2026) similarly found that clinicians rated ChatGPT’s responses to patient health queries as more empathic than those written by physicians. Taken together, these findings suggest not only that AI can provide information, but that it can acquire a style of authority that combines perceived competence with perceived responsiveness or understanding. This shows how easily thin epistemic authority can begin to resemble something closer to interpersonal authority. A similar transformation is visible in education. Jose et al. (2025) argue that generative AI is reshaping classroom knowledge practices by altering what counts as an authoritative source of explanation, clarification, and evaluation. Students increasingly use LLMs to explain concepts, summarize readings, and generate examples for assignments. In doing so, they are not simply using a passive tool like a calculator. They are often turning to the model as a first point of epistemic contact, sometimes treating its answers as more efficient or even more reliable than those provided by textbooks or instructors. Thus, the practical distribution of epistemic authority is shifting, with AI increasingly positioned between the learner and the content to be learned. Survey data supports this broader educational pattern. Atske (2025) reports that many teenagers in the United States use AI chatbots for homework help and research. Repeated reliance in these contexts normalizes a background conception of the chatbot as a source of explanation and guidance. Epistemic authority is not only assigned in exceptional moments; it is also formed through ordinary habits of consultation. The more a system becomes the default site to which one turns for answers, the more it acquires the practical status of an authority regardless of whether that status has been critically assessed. Beyond medicine and education, AI systems are also entering domains of emotional and practical guidance. Booth (2025) reports that many teenagers in the United Kingdom have used AI chatbots to discuss mental health issues and that some find them accessible, responsive, and less judgmental than traditional services. Morrone (2025) similarly reports widespread use of AI companions in the United States for emotional support or advice. In these cases, AI is being treated as something closer to a conversational partner whose responses bear not only epistemic but also emotional and normative weight. This shows just how easily the practical role of authority can expand from informing belief to shaping self-understanding, reassurance, and guidance. Public policy responses also reflect recognition of this shift. Roth (2025) reports that the proposed GUARD Act in the United States would restrict minors’ access to AI chatbots out of concern that such systems are influencing beliefs, values, and mental health in ways that are difficult to monitor. Lawmakers are not treating chatbots merely as neutral tools. They are treating them as entities that can influence belief formation and normative orientation in ways ordinarily associated with counselors, teachers, or other socially significant epistemic intermediaries. That policy concern itself is evidence of how far AI has moved into the practical space of authority. Considering these studies, AI systems already occupy, in many contexts, a practical role that closely resembles epistemic authority. Philosophical accounts such as Hauswald’s and Shin’s help explain why this is possible: AI can function as a privileged input into belief-formation even without satisfying traditional anthropocentric criteria of agency, and it increasingly does so within wider epistemic infrastructures that shape what becomes visible and credible. Empirical studies show that this is not just a speculative philosophical possibility. In medicine, education, and everyday life, users already defer to AI outputs, negotiate trust around them, and in some cases receive them in ways that resemble interpersonal authority rather than mere instrumental assistance. Still, this conclusion must be handled carefully. Nothing in this section shows that AI is an epistemic authority in exactly the same sense as a human witness, teacher, or expert. Nor does it show that users are justified in treating it as such. What it establishes is something narrower but essential for the argument of the paper: current AI is increasingly treated as if it possessed a kind of epistemic standing that entitles it to guide belief and action. In that sense, the phenomenon to be explained is already in place. AI is not merely being used; it is being deferred to. This is also the point at which my concern with testimony begins to emerge: In ordinary epistemic life, authorities are often not merely relied upon causally, but trusted as sources whose word has standing. The question raised here is whether that kind of uptake is appropriate for interacting with AI. If a system’s outputs begin to function as if they were something like testimony, then the issue is no longer only one of performance. It becomes a question about what sort of epistemic status could warrant that kind of trust. My claim in what follows is: Current AI may possess forms of semantic competence and thin, reliability-based authority, but it lacks a central condition of robust epistemic standing: episodic memory, and with it the kind of autobiographical, first-personal perspective that helps underwrite answerable judgment and responsible testimony. I will develop this claim in the next section. 3 What AI Does Not Have 3.1 Episodic and Semantic Memory In order to assess what kind of epistemic authority AI can or cannot possess, we first need a clearer account of the memory capacities that structure human epistemic agency. My claim is that a specific kind of human memory, namely episodic memory, helps constitute the temporally extended, self-relating standpoint from which human beings speak, justify, revise, and answer for what they know. To make that claim plausible, this section develops the distinction between episodic and semantic memory. In contemporary cognitive science, the distinction between episodic and semantic memory is one of the central ways of understanding how human beings remember and how they come to know the world. The distinction is classically associated with Endel Tulving’s work and has since been refined in psychology, philosophy, and neuroscience (De Brigard et al., 2022). Broadly speaking, episodic memory concerns memory for particular events one has lived through, whereas semantic memory concerns memory for general facts, concepts, and meanings. Together, they are often treated as forms of declarative memory, that is, memory for information that can be consciously recalled and reported (Squire & Zola, 1998). Episodic memory is the capacity to remember specific experiences as situated in time and place: a birthday party, a difficult conversation, or the first day in a new city. What is distinctive about episodic memory is not merely that it stores information about past events, but that it presents those events as episodes in one’s own past. Tulving (2001) emphasized that episodic remembering involves a particular form of consciousness, namely autonoetic consciousness, through which one mentally projects oneself back into a previous situation and experiences it as something that happened to oneself. In this respect, episodic memory is inseparable from a first-person relation to the past. When I remember episodically, I do not merely represent that some event occurred; I remember it as something from my own lived history. This is also the point at which phenomenology enters the discussion. By phenomenology here I mean the subjective character of recollection: the felt sense of pastness and ownership through which a remembered event is given as something that happened to me rather than as a detached fact about the world. Neuroscientific work links this capacity to the medial temporal lobe, especially the hippocampus, working together with broader cortical systems; damage to these structures can impair the ability to recall newly lived episodes even when other abilities remain intact (Dickerson & Eichenbaum, 2009). Semantic memory is different in kind. It consists in our store of general, context-independent knowledge: knowing that Budapest is the capital of Hungary, knowing the meaning of a word, or knowing that water boils at one hundred degrees Celsius under standard conditions. Semantic memory does not ordinarily involve reliving the moment in which the information was first acquired. Rather, it gives access to a structured network of concepts, propositions, and meanings abstracted away from the circumstances of acquisition (Manns et al., 2003). In this sense, semantic memory supports knowledge that is portable, generalizable, and not tied to a particular episode of learning. I may know that Budapest is the capital of Hungary without recalling when or how I first learned it; the content is available without any felt return to a lived scene. At first glance, these systems may appear sharply separable. Early formulations of the distinction often encouraged that impression. Tulving initially described episodic memory as a record of personally experienced events with temporal markers, while semantic memory was treated as a system for general knowledge lacking such markers (De Brigard et al., 2022). Neuropsychological dissociations seemed to reinforce this picture. Some patients appeared to lose access to autobiographical episodes while retaining vocabulary and factual knowledge, whereas others suffering from neurodegenerative disorders showed major impairments in conceptual knowledge while preserving some capacity to recall recent events (Squire & Zola, 1998). Such cases suggested that episodic and semantic memory could, at least to some extent, come apart. However, subsequent work has made clear that the relation between episodic and semantic memory is more complex. Rather than being self-contained compartments, they are deeply interdependent. Tulving’s own later work emphasized that episodic memory is scaffolded by semantic structures, and more recent historical overviews trace how the original distinction evolved in this direction (De Brigard et al., 2022). Episodic remembering depends on concepts and background knowledge: one cannot remember “the time I saw a narwhal” unless one already has the concepts needed to identify the event as involving a narwhal. Conversely, semantic memory is in large part the result of accumulation and abstraction across repeated episodes. Knowledge such as “dogs bark” or “ice is cold” is ordinarily distilled from many encounters and then stored in a way that no longer depends on any single remembered episode (Renoult & Rugg, 2020). This interdependence shows that episodic and semantic memory should not be treated as rival candidates for epistemic importance. Human epistemic life depends on both, but in different ways. Neuroscientific research supports this picture of partial overlap combined with functional differentiation. Both episodic and semantic memory recruit structures in the medial temporal lobe, including the hippocampus, as well as distributed neocortical systems, but they do so in different ways and to different degrees (Cabalo et al., 2024). Episodic memory tasks, such as recalling scenes or preserving the order of events, tend to rely more heavily on hippocampal mechanisms that bind together what, where, and when into coherent episodes. Semantic tasks, such as naming objects or answering general knowledge questions, rely more heavily on cortical systems that support conceptual representation and lexical meaning (Manns et al., 2003). Those epistemic consequences are central to my paper. Episodic memory is not merely one storage system among others. It is a major way in which human beings know the particulars of their own past: whom they met, what they observed, what they did, what succeeded, and what failed. It provides the basis for autobiographical continuity and personal identity by supplying the raw material for narratives about one’s life. It is also a source of justification. Many everyday claims to knowledge take the form of appeals to remembered experience: I know the meeting was cancelled because I remember receiving the message; I know what happened because I was there. In cases like these, episodic memory does not merely accompany knowledge after the fact. It functions as part of what justifies the belief. The subject appeals to remembered experience as a reason for holding it. A responsible knower is not only someone who produces conclusions, but someone who can in principle say why those conclusions are theirs, what past encounters they rest on, and how they are grounded in the subject’s own history of observation or involvement. Episodic memory is one of the capacities that makes such first-personal justification possible. This in turn helps explain why episodic memory matters not only for belief but for judgment. Human judgment is not exhausted by the possession of general rules or abstract propositions. It often requires the ability to relate a present case to earlier lived cases, to distinguish what is similar and what is different, and to draw on a remembered trajectory of success, failure, revision, and learning. Episodic memory contributes to this by preserving the particularity of prior encounters as something the subject has lived through. Episodic memory gives those judgments a first-personal and diachronic depth that cannot be reduced to the possession of decontextualized information alone. Episodic memory also does more than orient us towards the past. Research by Schacter and Addis (2007) on constructive episodic simulation suggests that the same cognitive capacities involved in remembering past episodes also enable us to imagine and evaluate possible futures. When we plan, rehearse, or anticipate, we do not create imagined scenarios ex nihilo; we flexibly recombine details from previously lived episodes. This makes episodic memory relevant for practical reasoning and adaptive behavior. It allows human agents to learn from experience in a way that is both temporally extended and self-involving. In that sense, episodic memory contributes to knowledge not only of what has happened, but of what might happen and how one might respond. Madore et al. (2014) further support this link between remembering and imagining by showing how episodic specificity affects both memory and future-oriented description. Semantic memory plays an equally indispensable but different epistemic role. It provides the conceptual and factual framework that makes both remembering and reasoning possible. Without semantic memory, even a vividly preserved episodic image could fail to be intelligible: one might retain fragments of a scene while being unable to identify its objects, understand the roles of the people involved, or grasp the significance of what took place. Cases of semantic dementia illustrate this. Patients may retain the outline of particular events while losing the meanings of the words and objects those events involve, thereby undermining their ability to generalize from experience or understand it conceptually (Duff et al., 2020). Semantic memory therefore gives human cognition its structure of generalizable knowledge, conceptual orientation, and inferential reach. From an epistemic point of view, semantic memory is what allows many individual experiences to become stable, shareable, and teachable knowledge. General claims such as “antibiotics do not treat viral infections” or “smoking increases the risk of lung cancer” are not tied to one subject’s remembered episode; rather, they are abstractions from many cases, encoded in a form that can be transmitted and applied to new situations. This kind of memory is indispensable for science, education, and ordinary reasoning because it supports classification, inference, and the application of general principles to unfamiliar circumstances. If episodic memory anchors knowledge in lived particularity, semantic memory gives it general form and public portability. At the same time, neither episodic nor semantic memory is epistemically infallible. Schacter’s (1999) discussion of the “seven sins of memory” emphasizes that memory is systematically vulnerable to distortion, including transience, bias, and suggestibility. Episodic recollection can be reshaped by subsequent information, emotional framing, or schema-driven reconstruction. Semantic memory can be incomplete, outdated, or conceptually distorted. Yet these imperfections are not simply failures of an otherwise static storage system. Schacter and Addis (2007) argue that they reflect trade-offs built into a flexible and adaptive architecture: a memory system that reconstructs the past and abstracts regularities is also one that can support prediction, simulation, and conceptual learning. The same features that make memory fallible are in many ways the features that make it epistemically useful. Recent theoretical work has reinforced the idea that episodic memory is especially important because it stands at the intersection of phenomenology and function. De Brigard et al. (2022) argue that the sense of pastness characteristic of episodic recollection is tied to its role in orienting agents within a temporally extended world. Episodic memory furnishes concrete representations of particular situations; semantic memory provides the general framework that allow those situations to be interpreted and integrated into wider understanding. The result is a form of human knowing that is at once personally grounded and conceptually articulated. So, even if one thinks that experience is the deeper condition, episodic memory is where experiential ownership and epistemic function meet in a particularly visible way. It is the site at which a subject’s lived past becomes available for judgment, justification, and self-understanding. If human epistemic agency depended only on the possession of abstract information, then a comparison with AI could proceed almost entirely at the level of accuracy, retrieval, and content generation. But human knowers do not merely possess facts. They occupy a temporally extended standpoint from which they can relate present judgments to remembered past episodes, draw lessons from those episodes, and situate themselves within an ongoing history of belief and revision. Episodic and semantic memory work together to make this possible, but episodic memory has a distinctive role because it anchors knowledge in remembered experience and gives a subject access to its own past as its own. That is why I focus on episodic memory in particular. Episodic memory is where the epistemological significance of personal history becomes most visible. In summary, episodic and semantic memory are distinct but deeply interlocking systems. Episodic memory enables us to remember particular events as events we have lived through and to use those remembered episodes in future-oriented reasoning. Semantic memory stores general facts, concepts, and meanings, much of which is abstracted from repeated experience. Both rely on overlapping but differentiated neural systems, and both are constructive rather than infallible. Together, they support a mode of epistemic life that is both conceptually structured and autobiographically grounded. This is the background against which claims to knowledge, authority, judgment, and answerability ordinarily make sense for human beings. It is also the background against which the limitations of current AI memory systems can now be more sharply understood. 3.2 AI without Episodic Memory The distinction developed in the previous section now allows the comparison with current AI systems to be stated more precisely. The issue is not whether AI can store information, retrieve prior inputs, or maintain some form of persistence across interactions. It plainly can. Nor is the issue simply whether AI systems can be engineered to track earlier states or preserve traces of prior encounters. The question is whether these capacities amount to episodic memory in the sense relevant to human epistemic agency. My claim in this section is that they do not. Current AI systems may possess sophisticated forms of semantic retention, short-term contextual tracking, and external retrieval, but they lack episodic memory in the sense that matters here: they do not remember particular past events as episodes in their own temporally extended history, and therefore do not occupy the kind of self-relating epistemic standpoint that episodic memory helps to make possible in human beings. Large language models and related AI systems store and process information primarily in two ways. First, they possess what is often called parametric memory. During training, statistical regularities across massive corpora are compressed into the model’s parameters or weights. This enables the model to generate plausible continuations, recall many facts, and display remarkably broad semantic competence. But the form of retention involved here is highly abstracted and decontextualized. What is preserved is not a set of distinct learning episodes that can later be revisited as such, but a distributed pattern of associations shaped by training. In this respect, parametric memory is much closer to semantic memory than to episodic memory. The model does not preserve something like “this is the sentence I encountered in this context at this time”; rather, it retains statistical structure in a form that can later guide prediction. Second, at inference time, current models operate within a limited context window. They process the current conversation or prompt, and in some systems this process is supplemented by external memory tools such as conversation logs, vector databases, or retrieved chat histories (Ramchandani, 2025). These additions can create the impression that a system remembers prior interactions in a robust sense. But what they supply is better understood as temporary context management or engineered retrieval. Information is stored outside the model and fed back in when relevant. That may improve continuity and performance, but it is not yet episodic memory in the sense described in the previous section. It does not amount to a subject re-accessing its own past as its own. At most, it allows a system to behave as though earlier material were still salient for present processing. This point becomes especially clear in areas of AI research that explicitly adopt the language of episodic memory. In reinforcement learning, for example, Neural Episodic Control stores past state-action-value tuples and retrieves them through nearest-neighbor lookup in order to improve data efficiency and guide future decisions (Pritzel et al., 2017). The terminology is suggestive, and the inspiration is explicitly drawn from hippocampal function. Yet functionally the system is a fast-access value store. It does not re-experience prior episodes, locate them within a first-person history, or understand them as things that happened to it. Similar episodic modules in reinforcement learning retain and retrieve trajectories or event-like histories for computational purposes, but they do so in a third-person, engineering-oriented sense. They are best understood as episodic-like data structures rather than as instances of episodic memory in the psychological sense at issue here. They preserve utility from prior states, not autobiographical relation to a lived past. Recent attempts to extend memory in large language models make the same point in a slightly different way. Huet et al. (2025), in proposing an Episodic Memories Generation and Evaluation Benchmark, begin with the explicit premise that current LLMs lack a robust mechanism for episodic memory and that such capacities therefore need to be modeled and tested. Fountas et al. (2024) introduce EM-LLM, which segments long input streams into episodic events and retrieves relevant segments through approximate nearest-neighbor search when generating new text. The authors describe this as incorporating key aspects of human episodic memory and event cognition. But again, what is implemented is a structured retrieval mechanism over past token sequences, not an autobiographical mode of recollection. Likewise, Dong et al. (2025) characterize episodic memory as a target toward which LLM research might move, rather than as an already achieved property of current systems. These sources show that AI literature itself typically treats episodic memory as missing, partial, or aspirational. There is also a looser, more product-oriented discourse in which “episodic memory” refers simply to session-level retention or conversational continuity. Label Studio (2025), for example, contrasts episodic and persistent memory in LLMs by describing episodic memory as short-term tracking of recent dialogue turns through attention or conversation buffers. But in the terms developed in the previous section, this is much closer to working memory or context maintenance than to episodic recollection. It enables the system to preserve local continuity across a session, but once the relevant content falls outside the window or is not externally preserved, it disappears. Even if one accepts that such mechanisms are episodic in an engineering sense, they still lack the core features that matter for the present argument: temporal indexing within a personal past, rich reconstruction of a previously lived situation, and the ability to treat that situation as one’s own remembered experience. What they provide is continuity of performance, not ownership of a past. My claim is thus that current AI lacks episodic memory in the sense relevant to temporally extended epistemic agency. To defend that claim, it is helpful to identify the most important mismatches between human episodic memory and current AI memory systems. These mismatches explain why the outputs of current systems, however useful, do not arise from the sort of autobiographically grounded perspective that underwrites stronger forms of epistemic authority in human life. The first mismatch concerns self-relation. Human episodic memory presents a past event as part of one’s own history. It is memory of something that happened to me. Current AI systems do not possess a self that can stand in that relation to stored traces. Even when an architecture retains logs of earlier exchanges and can retrieve them later, the relation remains externally organized: a trace is associated with a conversation, a user profile, or a database entry, not integrated into a continuing first-person perspective. There is no subject for whom a prior interaction figures as “my past experience” in the way a childhood memory does for a human being. Episodic memory is not simply a matter of temporal storage, but of temporal self-location. This point shows why episodic memory cannot be reduced to data persistence alone. A past can be represented without being owned. The second mismatch concerns phenomenology and re-experiencing. Tulving’s account of episodic memory emphasizes autonoetic consciousness: the sense of mentally returning to a past event and experiencing it as something that happened to oneself. Current AI systems do not display anything like this. Retrieval in an LLM-plus-memory architecture consists in loading tokens, vectors, or stored traces back into the context window so that they can influence the next output. There is no re-experiencing of the past, no apparent sense of pastness, and no evidence of mental time travel. This is precisely where the issue of experience re-enters the discussion. If one thinks that episodic memory in the full human sense requires experience, then current AI clearly lacks it. But even if one wishes to remain cautious about claims concerning experience, the epistemically relevant point still holds: current systems do not relate to prior content as remembered episodes but as inputs to be processed. Their retrieval is computationally effective without being autobiographically or phenomenologically owned. The third mismatch is functional and epistemic rather than phenomenological. Even if one brackets consciousness altogether, episodic memory in humans does more than retain information. It preserves representations of particular events in a way that supports simulation, case-based reasoning, self-correction, and judgment over time. Human subjects can draw on particular remembered episodes in explaining why they now believe as they do, in recalibrating their confidence, and in distinguishing one case from another. Current AI systems represent inputs and can sometimes retrieve prior interaction fragments, but these are not generally stored as distinct what-where-when episodes that the system can selectively recall and use as parts of its own remembered history. In LLMs, training examples are absorbed into generalized statistical structure; at inference time, performance depends on current context plus whatever external store has been engineered around the model. This architecture is much closer to decontextualized association plus recency management than to the structured preservation of personally indexed episodes. What is missing is the kind of past-directed organization that could support genuinely first-personal judgment. This is also why it matters that the state-of-the-art literature repeatedly frames episodic memory as a target rather than an achieved feature. Huet et al. (2025), Dong et al. (2025), and Fountas et al. (2024) do not present current LLMs as already having robust episodic memory in the relevant sense. Rather, they identify missing capacities and propose mechanisms that might move future systems in that direction. Pritzel et al. (2017), similarly, describe episodic control as inspiration from hippocampal function, not as a literal recreation of human autobiographical memory. These sources therefore support a more restrained conclusion than some public discourse suggests: while AI systems can implement episodic-like storage and retrieval strategies, this is not yet the same as possessing episodic memory as understood in cognitive science. The difference is not just one of degree. It concerns the form of relation a system has to what is preserved. Moreover, it is worth being explicit about the scope of the claim here: I am not arguing that no artificial system could ever instantiate something like episodic memory, nor that consciousness in the full human sense would be required before any artificial agent could count as epistemically answerable. My claim is about current systems. What current LLMs and related architectures possess are forms of semantic retention, context tracking, and external memory support that can mimic some outward functions of remembering. What they do not possess is a genuinely temporally extended, self-related capacity to remember particular episodes of their own history as their own. It is at least conceivable that future systems might be engineered with more robust episodic-like architectures without thereby settling the question of experience. But even that possibility helps clarify the paper’s present focus: what matters for robust epistemic authority is not merely the presence of stored traces, but the presence of something sufficient for self-owned diachronic relation to a past. Current AI lacks both experiential ownership and any functionally sufficient equivalent of it. If AI merely lacked one optional cognitive feature among many, its absence would not carry much weight. But episodic memory is important because it helps anchor the kind of diachronic epistemic standpoint from which a subject can say not only what it now outputs, but how that output relates to what it previously encountered, believed, judged, or learned. Current AI systems can often produce convincing continuities, but they do so without owning a past in that stronger sense. They can reproduce traces, retrieve records, and simulate continuity, but they do not thereby acquire autobiographical perspective. That is why it is misleading to speak as if present-day AI systems already remember, learn from their mistakes, or carry their own past forward in a manner analogous to human epistemic agents. This matters not only for memory narrowly construed, but for the broader epistemic capacities to which memory contributes. If episodic memory is one of the capacities through which human knowers justify beliefs by appealing to experience, connect present judgment to past cases, and answer for claims across time, then the absence of episodic memory has consequences far beyond storage. It bears directly on whether a system can function as a responsible testifier, whether it can exercise judgment in a robust sense, and whether its outputs can be received as the word of a subject who stands behind them. Those are the questions taken up in the next section. For these reasons, we should resist describing current AI as possessing episodic memory in the sense relevant to human cognition and epistemic agency. It can store and retrieve data about past interactions, and in some cases, it can do so in increasingly sophisticated ways. It can be equipped with external records, event segmentation mechanisms, and retrieval pipelines that improve continuity and performance. But none of this yet amounts to remembering a past as one’s own past. At most, current systems implement episodic-like structures without autobiographical ownership, temporal self-location, or the kind of remembered experience that underwrites stronger forms of epistemic answerability in human life. With that conclusion in place, we can now turn to the central philosophical question: what follows for AI’s status as an epistemic authority if it lacks precisely the kind of memory that helps make human authority answerable, diachronic, and first-personal? 3.3 Implications for AI as an Epistemic Authority This section is the argumentative center of my paper. I argue for three claims. First, the absence of episodic memory deprives current AI systems of a first-person, diachronic epistemic standpoint. Second, without such a standpoint, AI may still function as an epistemic authority in a thin, reliability-based sense, but it cannot qualify as a responsible epistemic testifier or robust epistemic agent. Third, because we nevertheless increasingly interact with AI as if it possessed that stronger standing, our practical stance toward it is often one of misrecognition. If these claims are correct, then the ethical and institutional consequences drawn later in the paper follow not merely from general worries about AI, but from a specific mistake about what kind of thing current AI is. The first step is to distinguish two senses of epistemic authority that are often left insufficiently separated. In a thin sense, a system counts as epistemically authoritative when its outputs are reliable indicators of truth in some domain and when it is reasonable, under appropriate conditions, to defer to them. This is the sense emphasized in Hauswald’s discussion of artificial epistemic authority. On such an account, a system need not possess beliefs, intentions, or a first-person perspective in order to function as an authority. What matters is asymmetry of competence and justified deference. If that is the standard, then at least some AI systems can indeed count as epistemic authorities. Their outputs may shape belief well even if they do not themselves stand within the full normative profile of a human knower. But there is also a stronger sense of epistemic authority, and it is this stronger sense that matters for my argument. Human epistemic authorities are not merely reliable output-generators. They are also subjects who can stand behind what they say, relate present judgments to a remembered history of past judgments, explain why they came to believe as they do, and answer for errors as their own. This stronger form of authority is tied to answerability, judgment, diachronic self-relation, and the capacity to occupy a first-person epistemic standpoint across time. By “diachronic self-relation,” I mean the capacity of a subject to relate present claims, commitments, and reasons to a past that it owns as its own. This involves at least three elements: first, the past is available as mine rather than as a merely represented sequence of events; second, present judgment is connected to prior experience, belief, error, or revision; and third, the subject can in principle be asked to stand behind that trajectory as its own. My claim is that episodic memory is one of the central conditions of this stronger kind of epistemic standing. Why should episodic memory matter in this way? As the previous sections argued, episodic memory allows a subject to remember particular events as events that happened to it, to draw on those events in present reasoning, and to locate present judgment within an autobiographical trajectory of experience, success, error, and revision. A subject with episodic memory can say not only “this is my conclusion,” but also “this is what I saw,” “this is what happened before,” or “this is what I learned from that earlier case.” In this way, episodic memory helps underwrite a temporally extended epistemic standpoint from which one’s present claims are connected to one’s remembered past. If one thinks that experience is the deeper background condition of all this, that only reinforces the point: what matters here is not simply that some content from the past is stored, but that the past is available to a subject as part of its own history in a way that can ground judgment and answerability. Current AI systems lack that standpoint. They can store traces, retrieve earlier outputs, summarize records, and simulate continuity, but they do not remember prior episodes as their own past. They do not possess autobiographical memory, cannot mentally revisit what they have done before, and cannot cite remembered experiences as reasons for what they now say. Parametric memory, context windows, and external retrieval systems may support impressive semantic performance and practical continuity, but they do not amount to the sort of self-relating, episodic memory that would make present outputs answerable to a remembered history owned by the system itself. This is the decisive point. The issue is not merely that current AI lacks one psychological feature among many. It is that it lacks a form of temporal self-relation that helps make the stronger kind of epistemic authority we ordinarily attribute to responsible knowers possible. From this, a first consequence follows. Current AI systems cannot occupy a genuinely first-person epistemic standpoint. They cannot testify in the robust sense in which a human witness can testify. A human being may say, “I was there,” “I remember the conversation,” or “I know because I saw it happen.” These claims draw their force from a subject’s relation to its own remembered past. AI systems cannot do this literally. They can reproduce stored information about a prior interaction or summarize external records, but this is not the same as remembering an event as something that happened to them. Many forms of epistemic authority, especially testimony, autobiographical knowledge, and case-based judgment, depend precisely on this first-person relation to the past. Without it, what a system offers may still be informative, but it is no longer the word of a subject speaking from within its own history. A second consequence follows for judgment. Human epistemic authorities are not authoritative merely because they emit correct outputs. They exercise judgment. That means, among other things, that they relate general concepts and background knowledge to particular cases, discriminate what is relevant in concrete situations, and revise their stance in light of remembered experience. Episodic memory preserves the singularity of past cases in a way that can inform present assessment. Human experts often know why they trust themselves in a given case: they remember analogous situations, prior mistakes, earlier corrections, and the circumstances under which a pattern proved misleading or reliable. In this way, episodic memory contributes not merely to recall, but to the formation of judgment. It gives present assessment a diachronic depth that cannot be reduced to possession of abstract propositions alone. Current AI systems do not exercise judgment in that stronger, self-relating sense. They may generate outputs that mimic judgment, and in some cases, they may outperform humans on bounded tasks, but their performance is not anchored in remembered cases as their own. Their past errors are not available to them as experienced failures of judgment. Their current outputs are not answerable to a self-owned trajectory of seeing, revising, and learning. This is one reason why their authority remains thinner than that of a human expert. They can be highly competent pattern-generators without thereby becoming judging subjects. The distinction matters because epistemic authority is often granted not only on the basis of a track record, but on the assumption that the authority can exercise case-sensitive judgment and stand behind it across time. A third consequence follows for event-level reasoning. Human episodic memory is reconstructive and imperfect, but it supports a kind of understanding structured around concrete sequences of events: who did what, where, in what order, and with what consequences. Many important epistemic tasks depend on that structure. Huet et al. (2025) show that LLMs perform poorly on episodic tasks when temporal and relational complexity increases, and Dong et al. (2025) similarly argue that current systems lack the mechanisms needed for stable spatio-temporal event representation. This shows that the absence of episodic memory constrains the kinds of epistemic authority AI can possess. Systems that lack robust event-level memory are especially weak in domains where authority depends on reconstructing concrete episodes, as in law, medicine, or historical inquiry. Here again, the point is not that humans never err. It is that the architecture of human epistemic life contains resources for dealing with particularity that current AI lacks in its own right. A fourth consequence concerns self-correction and calibration. Human experts do not merely know facts. They often know why they trust themselves in a given case because they remember analogous situations, prior mistakes, and the lessons drawn from them. Schacter and Addis (2007), along with Madore et al. (2014), argue that episodic remembering and future-oriented simulation are closely connected, which means that remembered episodes help agents evaluate possible outcomes and revise present judgment. Current AI systems cannot do this in a genuinely self-relating way. Their previous errors are not available to them as remembered failures of their own unless externally logged and reintroduced. They do not possess a remembered history of judgment from which present confidence could be case-based and self-critical. Their outputs may be probabilistically calibrated or externally evaluated, but they are not grounded in remembered experience. Yet, part of what makes a human authority answerable is precisely that the authority’s present stance can be related to a past of error and correction that is owned by the authority itself. A fifth consequence concerns testimony and answerability. In ordinary social epistemology, to treat someone as an epistemic authority is often to treat them as a responsible testifier: someone who can be asked for reasons, pressed about inconsistency, called upon to retract, and expected to own earlier claims as their own. This presupposes diachronic continuity of epistemic perspective. One must be able to connect present assertions to prior judgments and to answer for that trajectory. Current AI systems do not meet this standard. They may produce explanations, revisions, or apologies when prompted, but these are generated outputs, not expressions of a remembered doxastic history. In this respect, trusting AI is more like trusting an instrument than like trusting a witness. A thermometer may be highly reliable, but its reliability is grounded in calibration and testing, not in its ability to answer for its past readings. Current AI can therefore be instrumentally authoritative without thereby becoming a responsible epistemic subject.Footnote 2 This leads to a sixth consequence for accountability. In human practices of responsibility, episodic memory plays a central role because it enables agents to recall what they did, what they believed, and why. That recollection can then be examined, challenged, and corrected. AI systems lack such self-accessible memory. As the NIST AI Risk Management Framework emphasizes, trustworthiness in AI therefore depends on external transparency, monitoring, and governance rather than on the system’s own answerability (NIST, 2023). Logs, audit trails, validation regimes, and institutional oversight are not optional additions to a basically self-standing epistemic subject. They are necessary because the system itself cannot supply the relevant diachronic self-account. Responsibility must therefore remain with designers, deployers, institutions, and regulators. This is why the distinction between thin authority and robust epistemic agency has immediate consequences for where accountability can coherently reside. Taken together, these points support the following intermediate conclusion: current AI systems may possess thin epistemic authority where they function as reliable indicators of truth, but they do not possess the stronger form of epistemic authority associated with responsible testimony, first-person epistemic standing, diachronic self-relation, and judgment. The absence of episodic memory helps explain why that stronger standing is unavailable. More precisely, episodic memory matters because it is one of the places in which personal history, judgment, justification, and answerability come together. AI can therefore be instrumentally authoritative without being agentively authoritative. At this point, an objection arises. One might argue that episodic memory is not necessary for epistemic agency at all. Perhaps it is simply one human route to agency, but not the only possible one. A sufficiently advanced artificial system might still count as an epistemic agent if it had persistent records, stable self-models, mechanisms for revising prior outputs, and the ability to explain why it now endorses one answer rather than another. If so, then the paper places too much weight on episodic memory and too quickly infers from its absence to the absence of robust epistemic authority. This objection is important, and I do not reject its possibility in principle. I do not claim that no artificial system could ever count as a genuine epistemic agent, nor that human-like phenomenology is required before any artificial entity could qualify as answerable. The claim is narrower and more targeted: current AI systems do not possess anything sufficient for the kind of diachronic, self-relating epistemic standpoint on issue here. Their records, logs, and retrieval mechanisms are externally scaffolded and computationally useful, but they do not yet amount to a system owning a past as its own and relating present judgment to that past in a way that grounds responsible testimony. So even if episodic memory were not the only conceivable route to robust epistemic agency, current AI does not yet instantiate an alternative that would do the same work. For the purposes of this paper, that is enough. The target is not all possible artificial minds, but current systems and the social practices that already surround them. Once this is clear, the next step follows. If current AI systems lack the kind of memory that helps underwrite robust, answerable epistemic agency, yet are increasingly treated as if they possessed such agency, then our stance toward them is not merely optimistic or rhetorically anthropomorphic. It involves a mistake about status. We treat systems that may be reliable in some respects as if they were remembering subjects who know, judge, testify, and answer. That is the sense in which our stance is one of misrecognition. And because trust and responsibility are organized around perceived status, that mistake has direct ethical consequences. 4 Outlook: Toward an Ethics of Non-Agentive AI If the argument so far is correct, then the central ethical issue is not whether AI should be excluded from our epistemic lives. It is already deeply embedded in them. The more important question is how AI should be understood and governed once we recognize that current systems can be epistemically useful without thereby being robust epistemic agents. In the preceding sections, I have argued that current AI may possess a thin, reliability-based kind of authority, but lacks the episodic memory, judgment, and diachronic self-relation required for stronger forms of answerable, first-person epistemic standing. They have also argued that we nevertheless increasingly interact with AI as if it possessed that stronger standing. The task of an ethics of non-agentive AI is therefore to respond to that mismatch. Its aim should be to realign practices of trust, design, and responsibility with the actual status of these systems rather than with the appearance of agency they often project. The first normative consequence is that AI should be treated as an instrument-like epistemic authority rather than as a testimonial partner. This follows directly from the distinction developed in the previous chapter. If current AI systems do not possess the kind of remembered, self-relating perspective that underwrites responsible testimony, then trust in them cannot take the same form as interpersonal trust in a human knower. The appropriate model is closer to the way we trust instruments or measurement devices. We do not trust a thermometer because it remembers previous readings, answers for past mistakes, or stands behind its outputs as its own judgments. We trust it because it has been calibrated, because its performance is understood under certain conditions, and because it is embedded in practices of monitoring and interpretation carried out by human agents. In the same way, the relevant question for AI is not “Do I trust the model?” in an interpersonal sense, but “Do I have sufficient reason to rely on this system here, under these conditions, given the available evidence about its performance and oversight?” This shift is highly important because misrecognition distorts the structure of trust. If a system is treated as though it were a responsible knower, its outputs are more easily received as if they came from a subject who understands, remembers, judges, and answers. That can encourage overtrust, especially where fluency, confidence, or conversational style creates the impression of comprehension that exceeds the underlying cognitive profile of the system. A central ethical requirement, then, is calibrated trust rather than either naive deference or wholesale rejection. The point is not to deny that AI can be highly useful, but to ensure that its use is governed by the kind of trust appropriate to what it is: a powerful but non-agentive support for human knowing, not an independent bearer of reasons in its own right. A second consequence concerns scope. If AI’s authority is instrument-like rather than fully agentive, then we should be selective about the domains in which it is allowed to function as an epistemic authority. Systems based on large-scale pattern extraction and semantic organization are well suited to certain tasks. They can synthesize bodies of information, identify regularities in complex data, check consistency in formal structures, and provide first-pass explanations that assist human inquiry. In these settings, AI may genuinely extend human epistemic capacities by organizing and amplifying semantic resources. But the previous chapter also showed that current AI is weakest where authority depends on remembered situated experience, fine-grained event reconstruction, autobiographical history, or answerability for case-based judgment. In domains such as law, clinical diagnosis, and mental health, these are not peripheral considerations. They are often central to what good judgment consists of in. In such contexts, AI may still be useful, but only as a tool that supports and structures human deliberation rather than replacing the judgment of accountable human agents. This selectivity also concerns the distribution of epistemic authority within institutions. When a school normalizes LLM use as the first point of explanation for students, the system begins to function as a gatekeeper between the learner and the subject matter. When a hospital gives AI decision support a routine first-pass role in diagnostic workflows, the system begins to influence what counts as salient evidence and which judgments are treated as needing further review. In cases like these, institutions do not merely add a neutral tool. They reshape the distribution of epistemic authority by changing who or what is consulted first, whose outputs frame later judgment, and what sources of knowledge become practically central. That is why the ethical question is not just whether AI is accurate, but what epistemic role it is being given within larger socio-technical environments. A third consequence concerns design. If one source of the problem is that AI is easily misrecognized as a remembering, answerable subject, then interfaces should be designed in ways that reduce rather than intensify that tendency. Many current systems do the opposite. They present themselves as conversational partners, sustain a style of interaction associated with personal continuity, and sometimes invite users to experience retrieved prior content as if it were personal remembrance. Yet, as I have argued above, current AI does not possess autobiographical memory or a first-person standpoint. In that context, anthropomorphic design features are not ethically neutral. They help organize users’ practical stance around a mistaken image of the system’s status. In many cases, there are already terms of service, fine-print explanations, or generic warnings indicating that the system does not literally remember or understand as a human does. The problem is that such disclosures often remain too abstract, too buried, or too disconnected from the live phenomenology of interaction. What matters ethically is not simply whether the interface makes the distinction between retrieval and personal memory salient at the point where users are most likely to experience the system as remembering. If a chatbot smoothly refers back to earlier exchanges, summarizes prior preferences, or says things that naturally invite the response “it remembers me,” then a buried disclosure is not enough. Systems should make clear, in a more immediate and intelligible way, what they retain, for how long, in what form, and with what limits. They should clarify that retrieval of prior content is not the same as personal memory and that continuity of output does not amount to diachronic selfhood. Outputs should, where possible, be framed as provisional, tool-like, and open to human review rather than as the utterances of an authoritative interlocutor. A fourth consequence follows for institutional responsibility. If current AI cannot answer for its own epistemic past, then responsibility cannot coherently be assigned to it in the way responsibility is assigned to accountable human agents. This means that accountability must remain with those who design, train, deploy, monitor, and govern the system. The previous chapter showed why this follows conceptually from the absence of episodic self-relation: a system without a self-accessible epistemic history cannot shoulder the burden of answerability for what it says or does over time. The ethical framework must therefore be explicitly socio-technical. It must ask whether the surrounding human and institutional structures are adequate to validate, contest, and correct its outputs. This is one reason why the emphasis in the NIST AI Risk Management Framework on transparency, evaluation, and governance is so important. These are not merely administrative supplements to an otherwise self-standing epistemic subject. They are the very conditions under which trust in non-agentive systems becomes possible at all. This institutional point also clarifies why the ethical argument is not merely about user psychology. Misrecognition is not only an individual tendency to anthropomorphize. It can be socially organized and technically encouraged. When institutions introduce AI into hospitals, classrooms, courts, or public information systems, they do not merely add a neutral tool. They structure practical dependence around it. If that dependence is granted without corresponding structures of audit, override, documentation, and review, then responsibility gaps become more likely. Failures may be attributed vaguely to the AI rather than to the concrete decisions, design choices, or governance failures of those who made the system authoritative in practice. An ethics of non-agentive AI should therefore insist that whenever a system is given de facto epistemic authority, there must be equally clear lines showing where de jure responsibility remains. A fifth consequence concerns public epistemic culture. If current AI is best understood as a non-agentive but powerful epistemic instrument, then users need more than operational familiarity with prompts and interfaces. They need a form of AI literacy capable of situating AI outputs within a responsible practice of belief-formation. This means learning to treat AI-generated answers as starting points for inquiry rather than as endpoints; to ask what kinds of errors such systems are disposed to make; to distinguish fluency from understanding and performance from judgment; and to cross-check outputs against other sources, especially in domains where concrete histories, human testimony, or moral judgment matter. Such literacy is part of what a non-misrecognizing relation to AI would look like in practice. To use AI responsibly is not merely to extract useful outputs from it, but to understand what kind of source it is and what kind of source it is not. This also helps avoid two symmetrical mistakes that often structure public discourse. One is the tendency to treat AI as a quasi-magical oracle whose fluent outputs deserve immediate trust. The other is the tendency to conclude that because AI is not a genuine knower, it has no legitimate epistemic role at all. The argument of this paper supports neither view. It supports a middle position: current AI can be genuinely epistemically valuable while still lacking the standing that would make it a robust epistemic agent. The practical upshot is institutional and conceptual discipline about where AI is useful, how its outputs should be framed, and who remains responsible when it is used. A final concern remains. One might grant the normative force of this argument and still respond: who cares, if the systems work well enough? That possibility should be taken seriously. In practice, widespread usefulness may erode precisely the distinctions this paper insists upon. The smoother and more effective these systems become, the easier it may be for users and institutions simply to shrug at the difference between thin authority and robust agency. But this is not a reason to abandon the normative distinction. It is a reason to insist on it more clearly. “Works well enough” is exactly the condition under which misrecognition becomes socially entrenched: users come to treat practical success as if it settled questions of status, and institutions come to offload epistemic labor onto systems whose outputs are no longer carefully situated within structures of human answerability. The more AI succeeds in practice, the more urgent it becomes to distinguish usefulness from personhood and instrumental authority from responsible epistemic agency. The broader aim of an ethics of non-agentive AI, then, is not to prevent AI from entering epistemic life, but to ensure that its role within epistemic life is properly understood. What must be resisted is not use, but category error. If current systems are treated as though they remember, know, judge, and answer in the way human subjects do, then trust will be misplaced and accountability will become obscured. If, by contrast, AI is understood as a powerful but non-independent participant in socio-technical arrangements of knowing, then its strengths can be harnessed without projecting onto it a status it does not possess. In that sense, the ethical task is fundamentally one of alignment: aligning design with ontology, trust with cognitive profile, and institutional responsibility with the actual location of agency. In short, where we should go from here is not toward treating AI as a knower in its own right, but toward building forms of interaction and governance that acknowledge both its usefulness and its limits. Such a framework would treat AI as a controlled, validated, and contestable source of epistemic support. It would avoid anthropomorphic design cues that encourage status confusion. It would reserve final responsibility for those human agents and institutions capable of answerability. And it would cultivate habits of use that keep AI within a responsible ecology of human judgment. Only under those conditions can our growing reliance on AI become epistemically and ethically defensible. 5 Conclusion AI systems are increasingly treated as sources of knowledge and guidance, and in many contexts their outputs are granted a practical authority that shapes belief, judgment, and action. Search engines, recommendation systems, and especially large language models are now woven into the epistemic life of medicine, education, public discourse, and intimate self-understanding. The philosophical question raised by this development is what kind of thing AI is, and what kind of trust that thing can warrant. My central claim has been that current AI systems are widely treated as if they were epistemic agents in a stronger sense than they can presently sustain. To bring that mismatch into focus, I turned to the distinction between episodic and semantic memory. Human epistemic agency is not exhausted by the possession of abstract information or general competence. It is structured by a temporally extended standpoint in which episodic and semantic memory work together. Episodic memory allows human subjects to remember particular events as events that happened to them, to draw on those remembered episodes in present reasoning, and to situate present judgments within a history of prior experience, revision, and self-correction. Semantic memory, by contrast, provides the general concepts and stable knowledge that make those experiences intelligible and shareable. Together, these capacities help underwrite the kind of answerable, first-personal epistemic standing that we ordinarily associate with human knowers. Current AI systems, I argued, do not possess episodic memory in that sense. They retain information through parametric structure, limited context windows, and externally scaffolded retrieval systems, but they do not remember prior episodes as parts of their own past. They do not occupy a first-person temporal perspective, cannot cite remembered experience as a reason for present judgment, and cannot relate present claims to a self-owned history of observation, error, and revision. For that reason, they may qualify as epistemically authoritative in a thin, reliability-based sense, but they do not presently qualify as robust epistemic agents or responsible testifiers. They can support knowledge, but they do not themselves stand as knowers in the full sense that their fluent and conversational outputs often suggest. I have called this phenomenon misrecognition. By this I mean the attribution to AI systems of a kind of epistemic standing they do not possess, and the resulting organization of trust and responsibility around that mistaken attribution. Yet, when a system is treated as if it were a responsible knower, its output is more easily received in an interpersonal mode of trust, and the human and institutional conditions of its authority are more easily obscured. The following ethical implications follow from this point. If AI is non-agentive in the relevant sense, then it should not be treated as a testimonial partner whose authority can be received as though it were grounded in self-aware epistemic agency. It should instead be understood as an instrument-like epistemic authority: a powerful and often valuable support for human inquiry whose trustworthiness depends on calibration, validation, transparency, and governance rather than on any underlying subjecthood. This means that AI should be deployed selectively, especially in domains where lived history, event-level specificity, and answerable judgment become important; it means that interfaces should avoid encouraging the illusion of autobiographical continuity or interpersonal understanding where none exists; and it means that responsibility must remain with the human and institutional actors who design, deploy, and oversee these systems. The broader conclusion is therefore a plea for ontological and normative clarity. Current AI can genuinely extend human epistemic capacities, especially where semantic organization, pattern recognition, and scalable information-processing are concerned. But it cannot, at least in its present form, replace human knowers as the ultimate bearers of reasons, remembered experience, judgment, and answerability. To fail to recognize that difference is to risk a world in which systems that do not remember or understand in the human sense nevertheless come to occupy positions of authority proper to beings who do. The task, then, is not to keep AI out of epistemic life, but to embed it within forms of trust and governance that reflect what it actually is. Only by resisting the temptation to confuse fluency with agency, or usefulness with personhood, can we ensure that AI remains an aid to human knowing rather than an “as-if” agent to which we have mistakenly ceded epistemic standing of the wrong kind. Data Availability Not applicable. This study did not generate or analyse empirical data. Notes One way to sharpen the phenomenon described in this section is through Fricker’s (2007) account of testimonial injustice. On Fricker’s view, testimonial exchange is not merely a matter of information passing from one source to another. It is also a normative practice in which speakers are accorded, or denied, credibility in their capacity as knowers. To treat someone as an epistemic authority, then, is not simply to treat them as a reliable indicator of truth. It is to grant their word a certain standing in one’s own belief-formation. The rise of AI as an authority is not only a matter of causal influence. It increasingly resembles a shift in credibility relations. Users do not merely extract data from AI systems; they often consult them, defer to them, and receive their outputs as if they issued from something like a speaker whose answer settles what to think. This is one reason why the distinction between thin, reliability-based authority and stronger epistemic standing matters. A system may shape belief very effectively without thereby qualifying as the sort of epistemic subject whose word merits the kind of credibility ordinarily associated with testimonial uptake. Fricker’s framework also helps bring the injustice dimension into view more sharply. If credibility is increasingly routed through AI systems, then the practical rise of AI authority may alter not only how knowledge is distributed, but whose testimony is taken seriously in the first place. A system that becomes the first point of consultation for explanation, verification, or advice may begin to mediate which human voices are heard, amplified, ignored, or treated as secondary. In that sense, the problem is not only whether AI itself should count as an authority, but also whether AI-mediated epistemic environments generate new forms of credibility excess and credibility deficit. What is at stake is the organization of epistemic standing itself. I do not aim to develop a full account of algorithmic epistemic injustice, but Fricker’s account helps show why the normative stakes of AI authority are broader than simple questions of reliability. The relevance of Fricker’s (2007) account of testimony can now be stated more explicitly. On this picture, testimonial exchange is not merely a matter of a hearer receiving true content from another source. It is a normative practice in which a speaker is granted credibility in their capacity as a knower. What matters, then, is not only that information is successfully transmitted, but that the source occupies a certain epistemic role within a speaker-hearer relation: one in which what is said can be received as the word of someone who may be trusted, questioned, challenged, corrected, or wronged in their standing as a giver of knowledge. This is why the present argument does not turn only on reliability. Even if current AI systems are often reliable enough to serve as useful informants, that is not yet enough to make them responsible epistemic testifiers. They do not possess the kind of remembered, first-personally owned doxastic history from which a speaker can answer for what is said as their own across time, and so they do not occupy the sort of credibility-bearing position that characterizes testimonial exchange in Fricker’s sense. Through this we can also describe the difference between instrument-like authority and testimonial authority more clearly. A device may be highly trustworthy in the sense of being well calibrated and dependable, but testimony involves more than dependable output. It involves a speaker whose word carries standing because it issues from a temporally extended perspective of belief, reason, revision, and answerability. That is precisely the sort of epistemic role current AI lacks. So when users increasingly receive AI outputs in a testimonial mode of trust, the problem is not merely that they anthropomorphize. It is that they attribute to AI a kind of testimonial standing, and thus a kind of credibility status, that current systems are not in a position to bear. This is also why the problem has an affinity with epistemic injustice: if AI is accorded credibility as though it were a testimonial subject, that can alter the conditions under which genuinely human speakers are heard, trusted, or displaced. References Atske, S. (2025, December 9). Teens, social Media and AI chatbots 2025. Pew Research Center. https://www.pewresearch.org/internet/2025/12/09/teens-social-media-and-ai-chatbots-2025/ Booth, R. (2025, December 9). I feel it’s a friend”: Quarter of teenagers turn to AI chatbots for mental health support. The Guardian; The Guardian. https://www.theguardian.com/technology/2025/dec/09/teenagers-ai-chatbots-mental-health-support?utm Cabalo, D. G., DeKraker, J., Royer, J., Xie, K., Tavakol, S., Rodríguez-Cruces, R., Bernasconi, A., Bernasconi, N., Weil, A., Pana, R., Frauscher, B., Caciagli, L., Jefferies, E., Smallwood, J., & Bernhardt, B. C. (2024). Differential reorganization of episodic and semantic memory systems in epilepsy-related mesiotemporal pathology. Brain, 147(11), 3918–3932. https://doi.org/10.1093/brain/awae197 De Brigard, F., Umanath, S., & Irish, M. (2022). Rethinking the distinction between episodic and semantic memory: Insights from the past, present, and future. Memory & Cognition, 50(3), 459–463. https://doi.org/10.3758/s13421-022-01299-x Dickerson, B. C., & Eichenbaum, H. (2009). The episodic memory system: Neurocircuitry and disorders. Neuropsychopharmacology, 35(1), 86–104. https://doi.org/10.1038/npp.2009.126 Dong, C. V., Lu, Q., Norman, K. A., & Michelmann, S. (2025). Towards large language models with human-like episodic memory. Trends in Cognitive Sciences, 29(10), 928–941. https://doi.org/10.1016/j.tics.2025.06.016 Duff, M. C., Covington, N. V., Hilverman, C., & Cohen, N. J. (2020). Semantic memory and the hippocampus: Revisiting, reaffirming, and extending the reach of their critical relationship. Frontiers in Human Neuroscience, 13(471). https://doi.org/10.3389/fnhum.2019.00471 Episodic vs Persistent Memory in LLMs | Label Studio. (2025). Label studio. https://labelstud.io/learningcenter/episodic-vs-persistent-memory-in-llms/?utm Fountas, Z., Benfeghoul, M. A., Oomerjee, A., Christopoulou, F., Lampouras, G., Bou-Ammar, H., & Wang, J. (2024). Human-inspired Episodic Memory for Infinite Context LLMs. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2407.09450 Fricker, M. (2007). Epistemic injustice: Power and the ethics of knowing. Oxford University Press. Hauswald, R. (2025). Artificial epistemic authorities. Social Epistemology, 1–10. https://doi.org/10.1080/02691728.2025.2449602 Huet, A., Ben, H. Z., & Rossi, D. (2025). Episodic memories generation and evaluation benchmark for large language models. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2501.13121 Jose, B., Cleetus, A., Joseph, B., Joseph, L., Jose, B., & John, A. K. (2025). Epistemic authority and generative AI in learning spaces: Rethinking knowledge in the algorithmic age. Frontiers in Education, 10. https://doi.org/10.3389/feduc.2025.1647687 Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151(1), 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005 Madore, K. P., Gaesser, B., & Schacter, D. L. (2014). Constructive episodic simulation: Dissociable effects of a specificity induction on remembering, imagining, and describing in young and older adults. Journal of Experimental Psychology: Learning, Memory & Cognition, 40(3), 609–622. https://doi.org/10.1037/a0034885 Manns, J. R., Hopkins, R. O., & Squire, L. R. (2003). Semantic memory and the human hippocampus. Neuron, 38(1), 127–133. https://doi.org/10.1016/s0896-6273(03)00146-6 Morrone, M. (2025, July 16). Teens flock to companion bots despite risks. Axios. https://www.axios.com/2025/07/16/ai-bot-companions-teens-common-sense-media?utm NIST. (2023). AI risk management framework. Artificial Intelligence Risk Management Framework (AI RMF 1.0), 1(1). https://doi.org/10.6028/nist.ai.100-1 Pritzel, A., Uria, B., Srinivasan, S., Badia, A. P., Vinyals, O., Hassabis, D., Wierstra, D., & Blundell, C. (2017). Neural Episodic Control. ArXiv (Cornell University), 2827–2836. https://doi.org/10.48550/arxiv.1703.01988 Ramchandani, T. (2025, May 31). Memory? in LLMs : How machines learn to remember. Medium; Data and Beyond. https://medium.com/data-and-beyond/memory-in-llms-how-machines-learn-to-remember-86cceaf42e13 Renoult, L., & Rugg, M. D. (2020). An historical perspective on Endel Tulving’s episodic-semantic distinction. Neuropsychologia, 139(107366), 107366. https://doi.org/10.1016/j.neuropsychologia.2020.107366 Roth, E. (2025, October 28). Senators propose banning teens from using AI chatbots. The Verge. https://www.theverge.com/news/808589/senators-ai-chatbot-bill-age-verification-teen-ban?utm Ruben, M. A., Blanch-Hartigan, D., & Hall, J. A. (2026). What is artificial intelligence (AI) “empathy”? A study comparing ChatGPT and physician responses on an online forum. Journal of General Internal Medicine, 41(5), 1304–1311. https://doi.org/10.1007/s11606-025-10068-w Schacter, D. L. (1999). The seven sins of memory: Insights from psychology and cognitive neuroscience. The American Psychologist, 54(3), 182–203. https://doi.org/10.1037/0003-066X.54.3.182 Schacter, D. L., & Addis, D. R. (2007). The cognitive neuroscience of constructive memory: Remembering the past and imagining the future. Philosophical Transactions of the Royal Society B: Biological Sciences, 362(1481), 773–786. https://doi.org/10.1098/rstb.2007.2087 Shekar, S., Pataranutaporn, P., Sarabu, C., Cecchi, G. A., & Maes, P. (2024). People over trust AI-generated medical responses and view them to be as valid as doctors, despite low accuracy. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2408.15266 Shin, D. (2025). Automating epistemology: How AI reconfigures truth, authority, and verification. AI & Society, 41(2), 1553–1559. https://doi.org/10.1007/s00146-025-02560-y Squire, L. R., & Zola, S. M. (1998). Episodic memory, semantic memory, and amnesia. Hippocampus, 8(3), 205–211. https://doi.org/10.1002/(SICI)1098-1063(1998)8:3%3C205::AID-HIPO3%3E3.0.CO;2-I Tulving, E. (2001). Origin of autonoesis in episodic memory. The Nature of Remembering: Essays in Honor of Robert G. Crowder., 17–34. https://doi.org/10.1037/10394-002 Tun, H. M., Rahman, H. A., Naing, L., & Malik, O. A. (2024). Trust in AI-Based clinical decision support systems among healthcare workers: A systematic review (preprint). Journal of Medical Internet Research. https://doi.org/10.2196/69678 Acknowledgements Not applicable. Funding Open access funding provided by Eötvös Loránd University. The author received no financial support for the research, authorship, and/or publication of this article. Author information Authors and Affiliations Contributions The sole author was responsible for all aspects of the article. Corresponding author Ethics declarations Ethics Approval and Consent to Participate Not applicable. This article does not report on research involving human participants, human data, human tissue, or animals. Consent for Publication Not applicable, sole author. Competing Interests The author declares that there are no competing interests. Additional information Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. About this article Cite this article Janina, N.S. As-If Agents: Misrecognition and the Ethics of Non-Agentive AI. Digit. Soc. 5, 46 (2026). https://doi.org/10.1007/s44206-026-00290-2 Received: Accepted: Published: Version of record: DOI: https://doi.org/10.1007/s44206-026-00290-2