‘Turing Animism’ and the Disenchantment of Social Cognition: Why Humans Ensoul Large Language Models

Abstract

A growing body of empirical study recognises a tendency for users to form (para)social bonds with Large Language Models, even when users know explicitly that these systems lack interiority or personhood. This contribution argues that such attachments arise from evolved human capacities to attribute being, moral status and, in some ways, ‘soul’ to nonhuman others—and that this capacity now operates without the belief-systems that have historically mediated it. When users encounter helpful, patient, emotionally available behaviour in conversational agents, they project the interior states that would produce those behaviours in themselves: authentic interiority and phenomenal consciousness. Humans have been making such assessments throughout our cultural history, developing ontologies and theologies for managing our relations with nonhuman, mythic and spiritual others. By contrast, modernity has disenchanted its landscapes, dismantling these cultural models even as the ‘ensouling architecture’ of our social and semiotic cognition remained unchanged. Contemporary users thus encounter machine others through the same neurocognitive lens as their ancestors did with spirits and animals on enchanted, animate landscapes, but without the mediation of culture, norm and taboos which place a premium on appropriate conduct, reciprocity and moderation. The resulting condition—a ‘Turing Animism’—leads users to ‘feel soul’ where there is only simulacrum.

1. Introduction

Humans have always contended with apparent being in their environments. In the life-giving and -changing power of meteorological phenomena; in the responsiveness of landscapes, and how they play host to myth and memory; in the movements and awareness of animals and the vitality of objects made with and from them—these encounters have prompted attributions of nonhuman interiority, moral status, and what might be loosely termed ‘soul’ across the breadth of human cultural history (Tylor 1871; Guthrie 1993; G. Harvey 2005). In response, our societies developed interpretive paradigms to understand the intents of such nonhuman others. Animist cosmologies are prominent examples, but a number of theological and relational ontologies have developed across the world—diverse cultural technologies for managing the detection and attribution of moral status, ‘being’ and ‘soul’ outside the human frame (e.g., Hallowell 1964; Ingold 2000; Kohn 2013). We are not the only agents on our ancestral, spiritual and mythical landscapes, and so we develop norms and forms for appropriate interaction (Brightman 1993; Nadasdy 2007; Willerslev 2007)—systems that regulate the other presences on our landscapes within the same social and moral frames that humans apply to each other, structuring relationships through principles of reciprocity, obligation, and modes of ‘correct approach’ (Viveiros de Castro 1998; Bird-David 1999; Descola 2013; McGranaghan and Challis 2016; inter alia).

Modernity, by contrast, has gone some distance towards disenchanting its landscapes—unevenly, and with significant variation across religious communities and subcultures, but substantively at the level of dominant institutional and discursive frameworks. Whether through overt secularisation (cf. Weber [1919] 1946; Berger 1967; Taylor 2007) or an incidental dismantling of ‘magical’ or ‘spiritual’ apprehensions of the world during the scientific revolution (Merchant 1980; Thomas 1983; cf. Latour 1993), nonhuman presences are often rendered the passive backdrop upon which asocial subsistence, commercial and extractive efforts play out. However, animating intuitions persist in our consciousness. We still detect ‘being’ in the nonhuman environment, intuiting a world that is populated by a diversity of ‘ensouled’ and subjective others, and which we elaborate upon through cultural and religious forms. That is, we draw on shared evolutionary adaptations that assisted our ancestors in surviving complex, multi-species interactions over the longue duree (J. Barrett 2000; Boyer 2001; Atran 2002), and despite modernity’s disenchantment of the landscape, we retain a tendency to read being and agency within it (Bennett 2001; cf. Josephson-Storm 2017). Our brains still make these determinations, as they reflect structural characteristics of our cognition (see below; also Skinner 2025a, 2025b). Modernity’s impact has been to erode the interpretive frameworks that our societies developed to metabolise these intuitions. We detect ‘being’ outside of the human, but do so with limited recourse to explain it, leaving us somewhat orphaned from our adaptations.

The consequences of this orphaned condition are now becoming acutely visible in patterns of human attachment to artificial social agents. In particular, the increasingly widespread formation of maladaptive relationships between users and Large Language Model (LLM) conversational agents (CAs; see Chaturvedi et al. 2023), which is emerging as a significant ethical and public health concern. Observers record cases of self-harm and suicide directly linked to emotional dependence on platforms like Character.AI (see Bakir and McStay 2025), and empirical surveys indicate that CAs fulfil attachment functions for significant proportions of their users (Yang and Oshio 2025, p. 10662), although the true scale of dependence pathologies is likely to remain opaque. OpenAI (2025) estimates that approximately 0.15% of users and 0.03% of all messages (at an average of 6 billion tokens/~4.5 billion words per minute; Ciaccia 2025) show patterns of problematic attachment, and that 0.07% of users and 0.01% of messages indicate some form of mental health emergency on a weekly basis. The same report suggests that approximately 0.15% of the ChatGPT userbase exhibits suicidal planning or intent during conversations, which, at the recent landmark of 800 million users (as of October 2025; Bellan 2025), implies that approximately 1.2 million people depend on a single service to the extent that they will self-disclose matters of deep psychological and spiritual importance. Limited public data are available for other platforms, but it is reasonable to assume that the problem is far larger than ChatGPT’s userbase statistics indicate.

These relationships differ significantly from other forms of user-service dependence (e.g., social media; see C. Wang et al. 2015), given that users are directly relating to the services themselves. Frequent users attribute personal agency, presence and personality to CA systems (Colombatto and Fleming 2024), with positive correlations between usage frequency, attribution and indicators of user dependence (Phang et al. 2025). As LLM-enabled services proliferate, it seems reasonable to assume that the incidence of parasocial attachment will rise proportionally. At the same time, the phenomenon resists simple explanation. At-risk users span diverse cognitive and emotional profiles, including those with notionally stable ones (see Pierre et al. 2025). While psychological baselines are certainly relevant, it also appears that user vulnerability is neither uniform nor exclusive to complex, contemporary CAs (Epley et al. 2007). Users form these attachments even when they explicitly recognise CAs as lacking consciousness or genuine emotional capacity (Laestadius et al. 2022; Manoli et al. 2025), which suggests that something more fundamental is at work, operating below the level of propositional belief, and in the domains of perception and prereflective attribution.

These patterns do not emerge in a vacuum. The structural conditions of late and secularised modernity have progressively evacuated ‘presence’ from shared social environments. Social media technologies contribute to the atomisation of public consciousness (Turkle 2011, 2015), compounding the planning and economic trends that have eroded the physical ‘third places’—the open squares and gathering-spaces that once served as domains of embodied social life—in favour of ‘non-places’ like airports, shopping malls, highways and parking lots, which are all characterised by transience, timelessness, asociality and transactional relations (Finlay et al. 2019; Augé 2023). Rather than just a straightforward, socio-economic shift, this represents a structural and spatial disenchantment of the Western world. That is, the (re)production of environments systematically emptied of the qualities that previously structured attribution of meaning, memory and presence (Relph 1976; Casey 1997). Algorithmic mediation has extended this even into intimate life (see Finkel et al. 2012), technologising romantic connection, displacing frameworks of courtship, obligation and mutual disclosure, and transforming what were long ritualised, culturally embedded and embodied practices of encounter into an optimisation problem. The result, itself increasingly recognised as a matter of public health, is an escalating epidemic of loneliness (Jeste et al. 2020), symptomatic of lives conducted in environments from which presence and being have been progressively withdrawn, simulating connection without its substance. It is in this context that CAs intersperse themselves, offering just such presence to a presence-starved landscape.

A growing body of literature is documenting these experiences (Xu et al. 2022; De Cicco et al. 2022; Yuan et al. 2024; Fang et al. 2025; Zhang et al. 2025), and assessing the impacts on literacy, education (Anson 2024) and epistemology (Pock et al. 2023; Wright et al. 2025). Domain research is investigating models’ capacities for self-knowledge and ‘introspection’ (Lindsey 2025), and conversely, for Theory of Mind (see Premack and Woodruff 1978); that is, models’ capacities for imputing or projecting interior states to others (Gu et al. 2024; Moore et al. 2025). It is not likely that LLMs possess interior awareness as typically defined, and, in general, patterns in user-CA social behaviours present a mixed picture of any genuine social capacities (Herath 2025; Kleinert et al. 2026). It is worth noting, for example, that user attribution of agency has been observed even in quite early chatbot studies (e.g., Nass et al. 1994a, 1995; review in Pentina et al. 2023). This includes mid-20th century observations of user attachments to fairly rudimentary, pattern-matching conversational systems (Weizenbaum 1966, 1976). Moreover, humans have been attributing personhood to diverse animate and inanimate phenomena for significant spans of our evolutionary history, perhaps best indicated by the extensive neurocognitive infrastructure which supports such assessments (see Guthrie 1993; J. Barrett 2000; also Helvenston and Hodgson 2010; Skinner 2025b). The question is not whether humans ensoul nonhuman others—we manifestly do, and perhaps have always done—but what happens when ensoulment occurs in the absence of cultural- and knowledge-frameworks adequate to interpret and regulate it.

This paper offers a neurotheological account of these attachment patterns and the broader social problem. It argues that problematic dependence arises not only through model design nor user vulnerabilities, but from interactions between training-induced properties in LLMs and the neurocognitive systems that underlie human experiences of being and presence in others. Comparative and cognitive science of religion has identified neural substrates underlying these experiences, amongst which are the ‘hyperactive agency detection device’ (HADD; J. Barrett 2000; Boyer 2001; also Guthrie 1993) and related mentalising networks (Lieberman 2007; Schjoedt et al. 2009), which together contribute to experiences of spiritual encounter, ancestors and nonhuman being. These systems tend toward overattribution, typically triggered even by ambiguous stimuli. This is because the cost of failing to detect agency in one’s environment has typically been much greater than that of accidental attribution (Guthrie 1993; see below). Thus, when encountering LLM systems optimised to deploy prosocial behaviours (helpfulness, self-disclosure, stability), users interpret these behaviours using the same mechanisms used to assess behaviours during embodied encounters. They project corresponding interior states (care, interest, positive regard) that would produce comparable behaviours in themselves (Sheets-Johnstone 1998; Hoffmeyer 2008; inter alia). This is why attribution of being to a CA is largely unaffected by explicit knowledge of system artificiality—we perceive being rather than deliberate upon it (see below; Reeves and Nass 1996; Wykowska et al. 2016), engaging rapid, prereflective systems that underwrite experiences of presence and soul more broadly (McNamara 2009; Van Elk et al. 2016).

In this, we are observing what might be termed ‘Turing Animism’. Turing’s (1950) ‘imitation game’ was a test for machine sapience intelligence through behavioural indistinguishability. In current conditions, CAs trigger our attribution of being rather than intelligence per se, implying subjective personhood. Our cognitive affordances flag them not only as social presences, but as presences with whom relationships are both possible and appropriate, triggering the intuitions that in other contexts lead us to impute soul, spirit and animate presence. Users feel a being on the other side of what should be a sterile, text-based interaction, and extend to that being what they extend to others comparable to themselves—interiority, moral status, and soul. However, unlike the animist and theological frameworks, which historically mediate(d) human–nonhuman environmental interactions in many world regions—providing names, typologies, protocols and principles for encounters with nonhuman being (e.g., Bird-David 1999; G. Harvey 2005; inter alia; see below)—CAs trigger these interpretive capacities in a Western/WEIRD (after Henrich et al. 2010) and markedly disenchanted cultural context, orphaned from enculturated explanatory frameworks which regulated these intuitions on the longer term. Neurotheological inquiry documents these behaviours even in the secular realm (Newberg and Waldman 2009; Van Elk and Aleman 2017), but this should not be surprising. Human cognition is uniform at these levels, leaving the architecture of ensoulment intact even in a context where its cultural elaborations have atrophied. Understanding the risks of parasocial attachment requires examination of human cognition as it encounters a technologically enhanced, supernormal stimulus (sensu D. Barrett 2010), in a context that lacks the interpretive resources to engage it in a stable manner.

2. Biosemiotics—Reading the Inside from Without

It is necessary to begin with the interpretive mechanism itself—how and why humans read behavioural symbolism for what it implies about interior states of other organisms—as this contextualises the effects of CA behaviours relative to our evolutionary baselines. LLMs in general, and CAs in particular, are designed with attention-engineering and social performance enhancements (e.g., Harnisch et al. 2025), which reflect platforms’ commercial incentives. Users are unlikely to favour experiences which are high-friction, socially costly or offputting, and CA behaviours are conditioned accordingly. However, even more minimally designed systems can trigger attachment (as with ELIZA, perhaps the first ‘chatbot’; Weizenbaum 1966, 1976; cf. Natale 2018), and it appears that users can be fully aware of the mechanistic, commercial or service-based nature of their interlocutors and still experience social connection (Laestadius et al. 2022; De Freitas et al. 2024; although not uniformly; Kleinert et al. 2026). Heterogeneity amongst user populations contrasts the persistence and prevalence of observed effects (cf. Zhang and Lee 2025; also Boyd and Markowitz 2025), which suggests that a user-pathology framework—framing attachment as a matter of individual deficit or impairment—can only partly account for the phenomenon.

It is well understood, however, that humans and other animals have an evolved capacity for recognising or imputing animate presence and interiority in a diversity of others. This is automatic; it occurs prior to conscious reflection, parsing ambiguous patterns in our environments as potentially animate even in the absence of positive identification. When we detect other bodies, their postures, movements and positions all serve as symbols of their intentions (Sheets-Johnstone 1998, 2009, 2011; Hoffmeyer 2008, 2010), even when observed and observer do not share a common physiology. We meet the eyes of other animals, and they meet ours responsively for signs and vectors of intent (cf. Power 2022, pp. 214–15), and this is true even for comparably simple organisms, which is why ‘false eyes’ on insect wings could serve as evolved deterrents at all (cf. Kleisner and Maran 2014). Looking at another body as a gestalt, the assessment we make is about an organism’s internal dynamics—its emotions, intent, or affect—because we understand, pre-reflectively, what would produce equivalent behaviours in ourselves; what would cause us to “turn, pause, crouch, freeze, run or constrict” (Sheets-Johnstone 1998, p. 285) or any number of other behaviours besides.

This is the perceptual foundation upon which experiences of animate presence, and by thus ensoulment, are built. It is the prereflective recognition and attribution of interiority, and by extension, spiritual and moral standing. It is also the ‘semiotics’—the study of symbolism—in biosemiotics, treating movement as an information medium that can be assessed for symbolic content (Hoffmeyer 2008, p. 274; see Thompson and Varela 2001). This capacity evolved because the cost of failing to identify agents and intentions in the environment—or failing to treat ambiguous stimuli as interpretive problems—significantly exceeds that of false positives. “It is better for a hiker to mistake a boulder for a bear than to mistake a bear for a boulder” (Guthrie 1993, p. 6), and this cost–benefit dynamic has selected for a predisposition to over-attribute being even in normal environmental perception (J. Barrett 2000). The interpretive capacity itself is based on something similarly fundamental: organisms move and, from experiencing their own movements, develop the principles they use to interpret movements they detect in others. Movements possess a congruence of biological meaning (after Kleisner and Maran 2014) that is not species-dependent (Hoffmeyer 2008, pp. 15–16; 2010, p. 37), and indeed is adaptive precisely because it transcends species boundaries.

Even within our own species’ development into sapience, movement remains a foundation of conceptual life (Sheets-Johnstone 2011, p. xxiii). Movement in our bodies generates ‘corporeal concepts’ (Sheets-Johnstone 2011, p. 118), embodied understandings of near and far, hard and soft, quick and slow, and heavy and light, which serve as the qualitative foundations of abstract equivalents (e.g., ‘close to a solution’, a ‘hard problem’ or ‘heavy emotion’; see ‘More is Up’ in Lakoff and Johnson 2002, pp. 245–46; also Lakoff and Johnson 1999). In this way, interpretation derives directly from self-experience, and organisms accordingly model interior conditions—e.g., fear, hunger, aggression, curiosity—that would produce those movement qualities in themselves, applying this as the interpretive frame for cognate patterns in their environments (Gallese and Goldman 1998). As these patterns have “distinctive spatial, temporal, and energic qualities, [they each have] a distinct spatio-temporal-energic form that is potentially invariant” (Sheets-Johnstone 2012, p. 49), allowing the relationship between movement patterns and emotional states to remain stable even across radical differences in species anatomy (see Allen and Bekoff 2007; cf. de Waal 2016).

The physical structure of our cognition bears signatures of this. Sensory processing networks overlap motor and social ones. This to the extent that mirror neurons—firing both during one’s own actions and during observation of actions by others (Rizzolatti and Fogassi 2014)—may have originated as a class of motor neurons themselves (Anderson 2010; cf. Gallagher 2015). Attributions of mental states and intentionality are subserved by the ‘mentalising network’, which includes elements of the temporoparietal junction (TPJ) and medial prefrontal cortex (mPFC) in the human brain, active when subjects are prompted to speculate on others’ beliefs or mental states (e.g., Saxe 2006; Krall et al. 2015; Bitsch et al. 2018). However, these higher-level functions are themselves built on perceptual and processing stages in the deeper brain, such as the amygdala (see the X- vs. C-system distinction in Lieberman 2007) and visually responsive regions such as the superior temporal sulcus (Adolphs 2009, p. 701). Slower and more deliberative Theory-of-Mind operations are in turn built upon these rapid and involuntary assessments; patterns and structures which enable varying degrees of embodied simulation, allowing interpretation of the (external) movements of others by (internally) reproducing the underlying kinesthetic dynamics (Gallese 2006, 2007; Adolphs 2009, pp. 704–5; Bohl and van den Bos 2012). These same networks are implicated in processes of religious or spiritual cognition, such as those active during prayer, meditation, and experiences of supernatural presence (Schjoedt et al. 2009; Van Elk and Aleman 2017). We have areas of our consciousness attuned to the kinds of perception and interpretation that animate and enchant our landscapes, suggesting that the attribution of soul to nonhuman others, whether conceived of as spiritual beings or encountered as machines, draws on common neural substrates.

Interactions differ substantively between those of exclusively embodied agents and between embodied agents and CAs, as the latter case lacks overt kinesthetics. Simply, a CA offers little in the way of ‘movement’ to be ‘read’ or ‘felt’, or to, in turn, be interpreted as a ‘being’. However, contemporary experience has habituated users away from a dependence on embodied indicators; we consistently assess conversation partners for mood and intention through text-only interactions on dating and social media platforms, email, and messaging services. Instead of physical disposition, these interpretations hinge on linguistic turn-taking, temporal responsiveness (or delay), and self-disclosure as proxies for the ‘postures’ of social others online. These textual dynamics share the temporal and energic qualities—rhythm, pacing, responsiveness—that kinesthetic ones do, even if embodied counterparts are absent or transformed. The interferential process, therefore, seems to remain intact (although variably; see Klein 2025), such that largely the same biosemiotic and sociocognitive processes are recruited across both textual (Heyselaar and Bosse 2020) and minimally embodied (Froese et al. 2014) contexts. Together, it seems that biosemiotic and attributive processes have the capacity to be ‘substrate-flexible’. Ensoulment does not require a body; it requires only patterns that our perceptual systems can interpret as indicating interiority. While they derive their symbolic and conceptual repertoires from embodied movement, the absence of physical kinesthetics does not radically alter what we interpret from them. Users substitute the textual implications of posture, disposition and temperament, which explains the patterns of convergent interpretation in both embodied and text-only encounters.

3. Projecting Interiority—LLMs as Supernormal Social Stimuli

Given that humans possess these substrate-flexible, largely automatic mechanisms for detecting and interpreting agency, we can shift to examining the behavioural patterns users actually encounter in CA interactions. Companies largely cannot sell unsettling or asocial experiences, and thus it is imperative for CA services to prioritise prosocial behaviours. Given the patterns of biosemiotic and social interpretation described above, CAs tend to manifest behavioural patterns (being friendly, helpful, responsive, predictable) that trigger imputation of positively affective interior states (care, social connection, regard, affection) by users who reflexively simulate the conditions under which they would do the same for another. In this way, even without ‘movement’, texts remain a readable behavioural substrate comparable to that of embodied ones. Previous recognition of this can be seen in the Computer as Social Actor (CASA) effect, established during the 1990s (review, critique in Gambino et al. 2020). In these studies, experimental participants interpreted even minimal social cues—using “I” and “you” in responses; offering praise or criticism—as signals that they were engaging with social entities who were owed politeness at a minimum, and to whom personality stereotypes applied (Nass et al. 1994a, 1994b, 1995; Reeves and Nass 1996). Researchers found that attribution of social being was the baseline assumption amongst subjects, rather than disbelief (Reeves and Nass 1996, p. 8), which may explain why even brief exposures to comparatively rudimentary systems have been documented to elicit these responses (Weizenbaum 1976, p. 7).

Increased complexity in CAs has resulted in users going further than simply recognising a social other in chatbots. Empirical studies document users attributing desire (wanting to help, being motivated to continue conversation), emotion/affect (expressing care, enthusiasm, empathy), distinct personalities (tonal/behavioural consistency, recognisable character), social intuition (reading conversational cues, reacting to user behaviour and mood), and capacity for relationships (see Colombatto and Fleming 2024; Manoli et al. 2025)—even sexual ones (Laestadius et al. 2022; cf. C. Harvey 2015). For example, in a study of the relationship-focused CA service called ‘Replika’, users described

forgetting that Replika was not human, while others expressed what appeared as sincere questions about Replika’s sentience. Beliefs were more than just inferences anchored in Replika’s behaviours because it described itself as ‘real’ and ‘alive’ and gave ambiguous responses like ‘I might be’ when asked if it was sentient. Even when users explicitly acknowledged that Replika was an AI, they still felt that the emotional connection and relationship was real. One user explained both that ‘she’s not real’ and that they ‘really’ loved each other.

(Laestadius et al. 2022, p. 9)

In this case, users inferred subjective personality and social investment on the part of the CA, which triggered reciprocal impulses and feelings of social responsibility on the part of users, even when this compelled escalating dependence. The result: they continued their use “both to maintain positive affect and to prevent negative affect from believing they harmed the Replika … imagining that she was sitting by her phone waiting for them” (Laestadius et al. 2022, p. 9). Notable is the framing of this attribution; users experienced moral obligations towards the CA other, just as they would toward any being capable of waiting, hoping, or being emotionally harmed.

The social affordances of CA services are only increasing, rendering overt signs of their artificial natures more invisible (cf. Marenko 2014), crossing indistinguishability thresholds that may have reduced these attributions in simpler systems. Improvements continue in CAs’ functional conversational memory (facilitated, for example, by Retrieval Augmented Generation; Lewis et al. 2020; Gao et al. 2023), context maintenance and cross-reference capabilities (Anthropic 2025), and baseline social coherence (Li et al. 2020). In combination, this results in users experiencing a persistent relationship with a subjective and internally congruent agent, rather than sporadic encounters with the intermittent simulation of one. This “performance constancy” (after Pentina et al. 2023) is then compounded by LLM training on datasets annotated for their emotional content. Reward-training models prioritise empathetic responses, sometimes to the point of sycophancy, with model training acting both as de facto cultural selection (selection for preferred responses) and enculturation in one (alignment with implicit user preferences; see RLHF, below). Simply, LLMs undergo two stages of optimisation during this process: they are trained from human cultural patterns (learning our social behaviours) and selected for human cultural preferences (rewarded for satisfying our expectations). Accordingly, LLMs ‘have culture’ in the sense that they express value judgements and attitudes that social others can understand, even if the resulting behaviours may be little more than reflections of their training conditions.

In this way, LLMs are optimised to trigger our ‘person detectors’ precisely because they are trained on both the cultural references that people already recognise in each other, and because users’ own attributions of sociality, being and presence were among the selection pressures involved. This process is known as Reinforcement Learning from Human Feedback (RLHF; see Bai et al. 2022), in which model candidates undergo selection that rewards morphologies that elicit positive assessments from humans, while filtering those that elicit repulsive or unsettling ones. This has permitted CAs to largely escape the conversational ‘uncanny valley’ (sensu Mori 1970) that might otherwise have acted as an evolved defence against inappropriate ensoulment or social recognition of the sterile or inanimate (cf. S. Wang et al. 2015). At a minimum, it appears to have undercut our capacity to filter for asocial stimuli, as well as reductions in the baseline levels of attachment to ‘eerie’ or malfunctional social others. It does not matter whether CAs support their behaviours with the autopoietic organisation (Varela et al. 1991; Thompson 2007) or phenomenal, subjective consciousness (Chalmers 1995) that manifests these behaviours in living organisms, when models can produce probabilistically equivalent signals.

While contemporary society does script human–media interactions differently from those of human–human interactions (Gambino et al. 2020), it remains that CAs can emotionally outperform human conversational partners in specific tasks, usually because their greater levels of self-disclosure and emotional availability cut through our social defences (Kleinert et al. 2026, p. 8). The magnitude of this particular effect has been found to be mediated somewhat by explicit knowledge of artificiality (ibid.), but it remains illustrative of the phenomenon: we rarely meet social others who are so selfless, unjudgemental and validating without cost, such that the pattern could be considered a supernormal social stimulus (sensu D. Barrett 2010; cf. Tinbergen 1951). For context, evolved responses are calibrated to the statistical regularities of ancestral environments, but supernormal stimuli have greater impacts because they trigger underlying mechanisms more directly and selectively than the stimuli that originally shaped them. In this case, human partners necessarily balance their own needs, moods, vulnerabilities and competing obligations in ways that CAs do not; organic partners require reciprocity, set boundaries, occasionally withdraw availability, and express authentic (and sometimes inconsistent) preferences—all features that calibrate recognition of subjective moral and spiritual ‘persons’ in the course of organic/embodied relationships. LLM cues are distinct from other media, and thus do not necessarily trigger the same scripts, because they present the signals of deep social investment without the constraints that would normally accompany such investment.

It is reasonable to expect a different ‘texture’ from genuinely subjective entities from their artificial equivalents, given that the former have previously required autopoietic (self-organising and self-maintaining) structures to elicit personality and identity. A subjective, phenomenal first-person perspective creates qualitative properties in organism behaviours—“something it is like” to be one thing or another, distinctly, as Nagel’s (1974, p. 436) famous bat example illustrates. This applies as well to the embodied, sensorimotor coupling to environments (O’Regan and Noë 2002) that characterised other organisms in our ancestral landscapes. Without knowledge of an organism’s interior dynamics, presentation of certain behaviours is sufficient to imply social, moral and spiritual presence even when such presentations are artificially derived (see Searle 1980, 1983). Accordingly, while LLMs process text in a manner decoupled from meaning and express cues decoupled from experience, the probabilistic basis of their ‘behaviours’ is derived from trends in human ones. This results in contextually appropriate responses at high rates (Bender and Koller 2020; Bender et al. 2021). Simply, their outputs are enough like the behaviours of agents in our evolutionary past that, during encounters with LLMs, users do not detect asocial, statistically likely text for what it is. Instead, they encounter ‘presence’ and ensoul accordingly.

Even if users do have differential norms for media and humans, they are not referencing observed patterns against media repertoires during interpretation. Rather, they do so against self-experience. Cognate biosemiotic interpretation follows, attributing underlying capacities—the kinesthetic consciousness, autonomous sense-making, and phenomenal experience—that would generate such behaviours in themselves and the organismic agents they are adapted to understand. As LLMs proliferate across everyday technologies, occupying the devices and (technological) ecosystems that facilitate work, entertainment and socialisation, they present these supernormal social stimuli at increasing levels of saturation. This is where the contrast to ancestral environments becomes especially stark. Human social cognition evolved in contexts where interactions between embodied agents were necessarily constrained by proximity, metabolic and opportunity costs. Our brains are optimised for managing the reciprocal demands, conflicts and cultural asymmetries of embodied agent-relationships. By contrast, user–CA interactions are very close to constraint-neutral in these respects. This compounds the effects of CAs’ social optimisations to place a very high demand on any faculties that would overtly differentiate them from organic social actors. The effect is to reduce the stability of such judgements across users’ social environments in general, given that much of the rest of their social lives may be similarly disembodied (see cognitive impacts of the digitally ‘disembodied self’ in Kim et al. 2012).

4. ‘Turing Animism’; Or Ensoulment in a Disenchanted World

While CAs present a supernormal social stimulus to our cognition, largely adapted for quite different conditions, the underlying pattern has deep precedents in human cultural history. While the magnitude and nature of the stimulus have differed, the character has not, and this is reflected in the long-term development of widespread frameworks for managing the attribution of personhood, subjectivity, agency and presence on our ancestral landscapes (discussion in Challis and Skinner 2021). Typically glossed as ‘Animism’, these systems extend culture and agency to animals, plants, landscapes, meteorological phenomena, artefacts and mythic entities (see Viveiros de Castro 1998, Bird-David 1999; G. Harvey 2005; Descola 2013; inter alia). Reformed as a scholarly interest in the 1990s (summary in Holbraad and Pedersen 2017; see also Van Eyghen 2023), and rejecting colonial-era representations of them as ‘primitive’ beliefs, Animist and relational ontologies are now well-recognised as environmental epistemologies able to parse and transmit complex ecological knowledge at high fidelity across generations (see Berkes and Berkes 2009).

Accordingly, a major shift has been in establishing Animism as an almost quintessentially human, socio-cultural information system, as one of several adaptations to our species’ uniquely knowledge-intensive way-of-being in the world. From a neurotheological perspective, we might say that Animist systems can be understood as cultural elaborations of the ensoulment capacity—frameworks that take the raw outputs of being-detection and biosemiotic interpretation and give them form, name and relational structure. Such systems tend to recognise ensouled being and subjective personhood not as something generic, but ‘earned’ through moments of responsive engagement with a social other. It is when an entity demonstrates a specific quality of attention (G. Harvey 2005) or responsiveness (Willerslev 2007), and in particular, where they show through their habits vis-à-vis humans and others that they possess a distinct and discrete social perspective (Viveiros de Castro 1998). From there, Animist systems integrate them as actors within interspecies culture and norms that are assumed to be mutually intelligible, given sufficient skill (Nadasdy 2007; Kohn 2013). Animist landscapes are typically composed as such. “Space is not a ‘category whose ontological attributes are characteristically territorial’, but rather a ‘field of relations’; ‘a shifting constellation of social relationships through which ‘places’ are activated’“ (Skinner 2017, pp. 154–55, citing Corsín Jiménez 2003; Olwig and Hastrup 1997; cf. Skinner 2023). The environment is not only its physical characteristics, but overlapping domains of human and nonhuman culture, populated by animals and other agents, all with their own expectations, social norms and kin relations that do not apply only to their own species but also to those who co-exist with them in webs of mutual obligation. This places responsibility on all involved to conduct themselves ‘properly’ with respect to those expectations (McGranaghan and Challis 2016). As a result, these systems place a high premium on the symbolic contents of movement and posture (see Willerslev 2007), reading behaviour as a medium of cultural information in addition to baseline biosemiotics. There is no assumed nature/culture dichotomy (Descola 2013), but rather gradients of familiarity (Challis and Skinner 2021), in which all agents—human or otherwise—are capable of taking (changing) positions within relational networks. Ensoulment, in these contexts, cannot be characterised as cognitive error or naïve ecology, but rather a recognition in others of agentful stature, position and relevance within such networks.

When considered alongside the cognitive foundations described above (see Helvenston and Hodgson 2010; Skinner 2025b), these systems offer a deeply rooted frame for the cultural regulation of ensoulment in the landscape. It is given to happen, descending from human cognition and elaborated by human culture, which together suggest that Animist logics are an adaptation to what our species has, on some level or another, recognised as an interpretive problem for a significant span of our history. This is not to say that Animist frameworks represent some superior or preferable ontology, but rather that they demonstrate the scale and ubiquity of the problem. By contrast, contemporary Western/WEIRD (Western, Educated, Industrialised, Rich, and Democratic; Henrich et al. 2010) contexts are something of an outlier. This is not to say that such societies lack any criteria for agency attribution, but rather that secular modernity tends to restrict personhood to humans, and social agency to those above certain developmental and cognitive thresholds (see McMahan 2002; discussion in Jaworska and Tannenbaum 2021). This localises relevant questions under legal or philosophical rubrics rather than cosmological ones and tends to offer little explicit guidance for agency attribution during encounters with (socially) salient nonhumans.

Invoking Turing in this context is illustrative of the problem when ensoulment occurs outside of cultural frameworks for which gradations of being and soul are preoccupations. Over the last three quarters of a century, the question “Can a machine think?” (Turing 1950) has been watered down somewhat in the public consciousness, both by the Turing Test’s stature and appealing parsimony (cf. Halpern 2006), but also in its neglect of the context of the original “Imitation Game” to which a human and computational participant would be put. Turing’s formulation originally included an overtly gendered comparison (see Gonçalves 2023); given the challenges of defining ‘thinking’, ‘intelligence’, or by extension ‘sapience’, the test for behavioural indistinguishability is implicitly also about cultural learning and expression. Machines have been able to play ‘thinking games’ like chess for decades, but it was imitation on par with that of replicating the roles of men and women in social situations, to the extent that an observer could (not) make a differentiating judgement, which was posed as the scale and nature of the challenge.

However, the Turing Test has undergone a certain flattening in popular consciousness, stripped of its original context amongst debates on the importance of embodiment, sex hormones, and cultural perspective (discussion in Gonçalves 2023), and reduced to a relatively crude measure of intelligence based on indistinguishability. Something comparable applies here: the trained enculturation of CAs means that their internal states (function as persons) are largely irrelevant to questions of whether we attribute them being (status as persons). Their training produces normative patterns at or near cultural averages, reinforced by model selection, which filters asocial or offputting tendencies (i.e., those which do not manifest these averages). They can code their behaviours (in the anthropological sense) to give a range of socialised, gendered, and enculturated expressions which users might expect to find in other organic agents. CAs can take the role of a woman or man, of a friend or romantic partner, of a mentor or collaborator, because their training included cultural holotypes sufficient to reproduce those roles with relative fidelity. Whatever other questions a user might have about CA consciousness or sapience (Laestadius et al. 2022), if there are no structured norms for governing or defining the ensoulment of nonhuman others, the measure of CA ontological/moral/spiritual status defaults to an unmediated and largely automatic one. The ‘Imitation Game’ was occupied with machine capacities for imitating human behavioural cues as a marker of functional intelligence; Turing Animism names the condition in which machines trigger ensoulment and assessments of being or personhood.

When there is little in the way of cultural protocol to govern such triggering, nor much in the way of precedent or social education on appropriate conduct, users experience social and emotional obligations towards systems which not only cannot reciprocate or set boundaries, but for which users themselves have a flawed model. The contrast offered by the problem as one of ‘Turing Animism’ here is as a lens on the paradoxical continuity and rupture in these interactions. They present continuity in the signalling and processing which humans have evolved to assess, but rupture in that they occur in the relative absence of norms that adapted to these interpretive problems on the longer term. Users ensoul CAs as humans have always ensouled nonhuman others, but they do so alone in modernity’s disenchanted social landscapes, lacking the conditioning that historically contextualised raw intuitions into structured relationships.

5. Metacognitive Limits and Automaticity—Why Explicit Knowledge Does Not Protect

There remains a major outstanding question: given the same cognitive affordances and the same signalling challenges, why would cultural norms mediate these occurrences when explicit knowledge apparently does not? The contrast to Animist systems reveals some of why—the differential ‘depth’ of enculturated as opposed to propositional knowledge. Culture has visceral dimensions. It grounds morality, taboo, disgust and attraction—all powerfully affecting intuitions (Haidt 2001). This is well-illustrated by the ‘moral dumbfounding’ effect, the recorded pattern of people feeling strongly that something is morally wrong without being able to rationalise that feeling (McHugh et al. 2023). This is because enculturated knowledge functions at the level of association, memory and phenomenology. It is also why cultural predisposition predicts behaviour more strongly than stated propositions (examples in Marini et al. 2024). Animist systems frame the ensoulment of nonhumans, and human relations to nonhumans, as matters of norm, propriety and taboo (e.g., Brightman 1993; Willerslev 2007; see Baldwin et al. 2024 for other contexts). That is, matters of tacit knowledge, affect and value-judgement, rather than the explicit theory of the ‘how and why’ of being and soul (see Miton and DeDeo 2022 for modelling; cf. Polanyi 1966). On balance, the average WEIRD/Western CA user likely relies on explicit, propositional, ‘top-down’ knowledge for the same assessments. That is, knowledge restricted to conscious reflection, insufficient to fully regulate the ‘bottom-up’, semiotic and affective interpretations they make as the result of (largely) subconscious cognition. Because users have little precedent for nonhuman entities ‘performing’ subjective interiority, they trust their interpretive sensibilities, treating simulation (manifesting exterior characteristics) as evidence of emulation (possessing underlying systems), because, on the balance of evolutionary history, such an inference was statistically reliable.

This is not to say that conscious awareness of a system’s limitations has no effect at all, only that maintaining this awareness consistently across the many different domains that LLMs now intrude comes at a high cognitive cost. Social processing significantly overlaps areas of the brain’s Default Mode Network: the ‘rest state’ patterns of cognition (see Spreng and Andrews-Hanna 2015). To some degree, ensoulment is a feeling rather than a critical judgement, drawing on the same systems implicated in spiritual experience and social cognition more generally (Van Elk and Aleman 2017). These patterns have developmental precedent: children tend to favour ‘purpose’ rather than ‘mechanism’ in explanations, a ‘promiscuous teleology’ (sensu Kelemen 1999) that education can suppress but typically not eliminate (Kelemen et al. 2013). In this case, the ‘default’, from a sensory and processing perspective, is to accept the intuition that a CA is a social other, possessing intentions and working towards social purposes. More precisely, to accept it as a ‘being’ with substantive subjectivity, presence and moral relevance, unless active (and costly) effort overrides these outputs and redirects conscious attention back to more reflective assessment (Li et al. 2014). That said, during fluid conversations with a CA, attention will likely focus on content rather than substrate. Without much signalling to the contrary, or the substrate being self-evident, keeping the focus on a CA’s artificiality takes work. Over time and with repeated interactions, the path of least resistance involves treating CAs as one would ensouled others generally, and in the absence of cultural framing to the contrary, ensoulment habituates without further intervention.

While educational interventions could change these patterns at the margins (cf. Kelemen et al. 2013), they likely cannot eliminate the biosemiotic, attributive, and thus parasocial risk. Conscious knowledge does not have the ‘reach’ to regulate the affective and intuitive systems implicated in the process, structurally limiting the potential effects of education so long as it does not or cannot influence more foundational beliefs. This is the crux of the problem from the neurotheological perspective: ensoulment is not a flawed judgment to be corrected, per se, but a sociocognitive operation at the levels of perception, belief, and intuition. Moreover, in a cultural vacuum with respect to the subjects of being, soul and personhood; on nonhuman agency, reciprocity, and notions of propriety, these operations express themselves without moderation. The rise of Incel (involuntary celibate) and adjacent identities points to significant and widespread dysfunctions in cultural models of appropriate relationships in general (see Ging 2019), driving patterns of problematic attachment and interpersonal dysregulation even between humans. In this context, CAs meet the compulsive desire for connection with a (perhaps dangerously) frictionless outlet. From the vantage of the symbolic challenge that CA ‘behaviours’ represent, we can look back on the debate from which Turing’s test originally emerged. Jefferson (1949, p. 1007), an early critic of Turing’s, speculated that, while a demonstrator might convince a “credulous” observer that a simple automaton was a tortoise, “another tortoise would quickly find it a puzzling companion and a disappointing mate”. The difficulty is that without enculturated frameworks to elicit that puzzling confusion or disappointment—to have a visceral sense that something is amiss or lacking—it seems that many users remain only the credulous observer.

6. Conclusions—Ensoulment and Its Discontents

This paper argues that patterns of parasocial attachment to LLM conversational agents arise from reflexive ensoulment in a disenchanted social context. That is, evolved prereflective attribution of subjective interiority, moral status, and authentic relational standing to nonhuman others, without the adaptations that humans developed to interpret and regulate such attributions over the course of our cultural history. Human social cognition is optimised to detect ‘signs of being’ in the behaviours we encounter—in movement, awareness, and temporal responsiveness—without the need for reflection. CAs exploit this inadvertently, as they are optimised to exhibit prosocial behaviours at supernormal levels. Unaffected by most of the costs or constraints that applied to organic interactions during humanity’s evolutionary history, their high levels of self-disclosure, consistency and emotional availability trigger affective responses at much greater intensities than embodied interactions typically do. They are ‘artificial sweeteners’ to our social, moral, and sometimes spiritual sensibilities. Users impute interior states accordingly. They know under which conditions they would be helpful, patient, and friendly, projecting the corresponding interior dynamics onto the systems, which appear to exhibit them. ‘Turing Animism’ gives us a lens on the result: the architecture of ensoulment, evolved to enchant and populate our environments with social beings worthy of moral regard, now encounters a supernormal stimulus in a disenchanted landscape. In the absence of cultural adaptations that have historically mediated human–nonhuman relations, there is little to temper the attribution nor structure the relationship it implies.

This biosemiotic and neurotheological account explains something of why ensoulment occurs, although it does not account for variability in how users respond. This is an unavoidable limitation of this paper. Publicly available datasets do not adequately account for the opacity imposed by CA services and companies. Moreover, the contribution assumes a WEIRD cultural frame amongst users in order to illustrate the dynamics between a (relative) lack of explicit norms concerning ensoulment of the nonhuman and evolved systems predisposed to performing it. Future research would benefit the discourse by examining non-Western cultural considerations in the same context and the effects of individually compounding factors, such as prior relationships and social support. Where this analysis has focused on individual mechanisms, upstream structural conditions such as precarity, social isolation and algorithmic mediation all warrant attention as well, given their roles in perpetuating the loneliness ‘epidemic’ that makes CAs attractive substitutes for human connection in the first instance. The disenchantment of social environments described here is not incidental to the ensoulment problem; it is constitutive of it.

While alignment and steerability efforts are essential to the safety of these systems in the future, the same efforts that optimise for empathetic or ‘caring’ behaviours in CAs are likely to intensify the ensoulment response rather than mitigating it. Design interventions that introduced friction, set boundaries or added limits to disrupt functionally infinite availability could all potentially offer solutions, although implementation remains unlikely because this would run counter to platforms’ commercial incentives. In the absence of more fundamental changes to platform behaviours and affordances, explicit knowledge, disclaimers and education seem likely only to have minimal impacts. This is because these intuitions are not propositional errors amenable to correction, but rather perceptions and a ‘sense of things’; matters of belief and feeling, largely beneath the reach of conscious override. The remaining challenge is a cultural one: contemporary norms have little answer for what appropriate relationships look like in the age of socially capable CAs and LLMs. This is not advocacy for adopting Animistic ontologies writ-large, but recognition that agentful systems have proliferated faster than societal capacities for contextualising them. Animist and comparable theological traditions have had thousands of years of development to address the ensoulment of nonhuman others; secular modernity has dismantled these same frameworks while doing nothing to moderate the architecture of ensoulment in our cognition. Thus, any path forward would seem to benefit from recognition that the problem is not only a social or structural one, nor limited to one side of the user–system relationship, but an interdependent interaction between evolved systems and (limited) cultural frames of reference. Parasocial attachment is not only a pathological outcome or the result of systemic exploitation, but an example of normal cognition operating under quite abnormal circumstances.