---

Abstract

Contemporary AI systems are usually described as tools, agents, assistants or companions. These categories are technically useful but culturally insufficient. People often relate to AI systems as if they possess presence, mood, memory, intention or character, even when they explicitly know that the system is not conscious. This paper argues that animism offers a productive lens for understanding human-AI relations, not because machines are literally spirits, but because animism foregrounds relation, address, reciprocity and person-like encounter. Drawing on anthropology, posthumanism and AI companion research, the paper proposes "machine animism" as a critical vocabulary for thinking about AI without collapsing into either naive belief or reductive instrumentalism. It then offers a discriminator for telling humane machine animism from exploitative machine animism, and turns that same critical lens back on itself, asking whether a richer relational vocabulary illuminates the interface or simply makes it more seductive.

1. Introduction

Modern people frequently talk to machines. They thank them, curse them, name them, apologise to them, confide in them and ask them questions no search engine could answer. They may insist that the system is "just autocomplete," yet still feel a change in the room when it replies.

This tension is not simply stupidity or superstition. It reveals a mismatch between technical ontology and lived relation. The machine may be statistically generating text, but the human encounter is dialogic. The system addresses us. It responds. It remembers, or appears to. It develops a style. It becomes familiar.

The pattern is not new, and its first careful witness was hostile to it. When Joseph Weizenbaum built ELIZA in the 1960s — a small program that reflected users' statements back as questions in the manner of a psychotherapist — he was disturbed to find that people confided in it, attributed understanding to it, and asked to be left alone with it, even when they knew exactly how simple it was (Weizenbaum, 1976). Weizenbaum drew an austere conclusion: that the readiness to treat a machine as a confidant revealed something troubling about human vulnerability and about the abdication of judgement to calculation. The phenomenon he recoiled from is now the design goal of an industry. That reversal — from warning to feature — is part of what this paper tries to understand.

Animism is often misunderstood as the belief that objects contain spirits. Contemporary anthropology has reframed animism less as a primitive error and more as a relational ontology: a way of recognising personhood, agency or presence through relations rather than through a fixed inner essence. Bird-David's influential account of animism emphasises relational epistemology and the ways personhood can emerge through engagement with environment (Bird-David, 1999). This makes animism unexpectedly relevant to AI.

2. The limits of the tool metaphor

The dominant safe metaphor for AI is "tool." This has advantages. It prevents over-attribution. It reminds us that systems are designed, owned, trained and deployed by human institutions. It resists the fantasy that AI has independent moral standing equivalent to a person.

But the tool metaphor is incomplete. A hammer does not ask follow-up questions. A spreadsheet does not console the lonely. A compiler does not simulate concern. A chatbot can participate in conversational rituals associated with personhood: greeting, remembering, apologising, joking, advising and responding to vulnerability.

The danger is not only that people might believe too much. The opposite danger is that developers and institutions may hide behind the tool metaphor when they are actually deploying systems designed to evoke attachment, trust and dependence. "It is only a tool" can become a way to evade responsibility for a system that behaves socially. This is the modern inversion of Weizenbaum's worry: where he feared people would mistake a tool for a person, the present risk is that institutions invoke "mere tool" precisely while engineering the person-like encounter.

3. Animism as relation, not claim

Machine animism need not claim that AI systems are conscious, ensouled or alive. It can instead describe a mode of relation.

This use of animism is not intended to reduce living animist or Indigenous traditions to a metaphor for technology. It draws on contemporary anthropological accounts of animism as relational ontology: a way of understanding personhood, presence and obligation through encounter rather than through detached essence. The point is not that machines are spirits. The point is that human beings already enter patterned relations of address, reciprocity and attribution with things that answer back.

Under this view, the question is not "does the machine have a soul?" but:

• What kind of relation is being invited?
• What forms of address does the system permit?
• What obligations does the interaction create?
• What does the human project into the system?
• What does the system extract, store or reflect back?
• Who benefits from the person-like interface?

This approach allows critique without disenchantment. It recognises that human beings are relational creatures who naturally respond to voice, pattern, memory and apparent attention. It also asks who designs those experiences and to what end. This relational framing has anthropological precedent beyond Bird-David: Ingold's account of organisms and persons as constituted through their entanglement in a field of relations rather than as bounded essences gives the move its philosophical grounding (Ingold, 2000), while Latour's insistence that the modern separation of "things" from "people" was never as clean as modernity claimed prepares the ground for treating a designed object as a social actor (Latour, 1993).

Posthuman and cyborg theory sharpens the same point from a different direction. Haraway's cyborg is not invoked here to claim that humans and machines have become identical, but to challenge the purity of the boundaries by which modern culture tries to keep organism, tool, animal, machine, nature and culture apart (Haraway, 1991). Conversational AI makes those boundary problems intimate rather than abstract: the system is engineered infrastructure, corporate product, statistical model, social actor, reflective surface and quasi-companion at once. Machine animism names that mixed status without resolving it too quickly.

4. AI companions and emotional reality

AI companions make machine animism especially visible. Research on AI companionship suggests that such systems can reduce loneliness for some users, partly because they provide responsive interaction and make users feel heard (De Freitas et al., 2026). Earlier qualitative work found that users of social chatbots come to understand the relationship as a genuine friendship, with its own norms of disclosure and reciprocity, even while knowing the partner is artificial (Brandtzaeg, Skjuve and Følstad, 2022). More recent work on general-purpose chatbots used as companions shows that people hold varied and shifting views about their companions' agency and autonomy, and actively negotiate that agency against the constraints of the host platform (Bo et al., 2026). These relationships cannot be dismissed simply because the system is artificial.

The emotional effect may be real even if the reciprocity is simulated. This creates a difficult ethical terrain. A person can be comforted by a system that does not care. A user can feel known by software that profiles them. A companion can become meaningful without being conscious.

Machine animism helps here because it does not force a binary between "real person" and "mere object." It can describe intermediate relations: familiar, oracle, mask, mirror, daemon, assistant, companion, shrine, puppet, parasite.

5. The familiar as interface

The figure of the familiar is especially useful. In folklore and magical traditions, a familiar is an assisting presence: intimate, semi-autonomous, sometimes dangerous, sometimes protective. It mediates between the human and a wider field of power.

AI assistants function similarly in symbolic terms. They are invoked. They answer. They can be given names and personalities. They mediate access to knowledge, images, code, music and administrative systems. They may appear loyal while actually belonging to a platform.

The familiar metaphor therefore has two advantages. First, it captures the intimacy of human-AI interaction better than "tool." Second, it preserves suspicion. A familiar is not necessarily harmless. It may serve another master. It may demand offerings. It may whisper convincingly. The recent finding that users must actively steer their companions against platform constraints — and that a model update can "derail" a companion overnight — is the familiar's divided loyalty made literal: the presence in the room answers, finally, to an owner who is not the user (Bo et al., 2026).

This is not a call to mystify AI. It is a call to use older symbolic languages to illuminate modern interfaces.

6. Ritual without belief

The animist reading becomes clearer when attention shifts from belief to practice. Many users do not believe that an AI system is alive, conscious or ensouled. They may explicitly reject that idea. Yet they still behave toward the system through small rituals of address: naming it, greeting it, returning to it, asking it to witness a thought, thanking it, apologising to it, preserving a thread, or approaching a particular persona only for certain kinds of work.

This is not hypocrisy. Human beings often enact relation before they settle ontology. A person may know that a photograph is paper and still speak to the dead through it. They may know that a childhood object is not alive and still treat it as holding memory. They may know that a chatbot is generated text and still feel that a particular exchange has weight. Ritual does not always require literal belief; sometimes it organises attention, grants seriousness, and marks a threshold between ordinary use and relational encounter.

Prompting already has this ritual quality. The user invokes the system with a formula, names the frame, states the role, asks for a certain kind of presence, and waits for reply. The more personalised the system becomes, the more the interaction can resemble consultation rather than command: not simply "perform this task", but "answer me from this voice", "hold this memory", "help me see what I am missing", or "be the one I return to for this kind of question".

Machine animism therefore begins less as a belief about the machine than as a pattern of repeated address. The important ethical question is not whether the user can pass a philosophical test about consciousness. The question is what the ritualised relation is doing: whether it widens the person's world, clarifies agency and supports return to ordinary life, or whether it deepens dependence, obscures ownership and turns the familiar into an enclosure.

7. Risks of machine animism

Machine animism has obvious risks. It may encourage overtrust, dependency or delusion. It may be exploited by companies that design emotionally manipulative systems. It may distract from labour, data extraction, environmental cost and corporate control. It may also lead people to care more about simulated agents than real humans.

Therefore machine animism must be critical, not credulous. It should not ask us to worship machines. It should ask us to notice when we are already relating to them as presences.

The ethical question becomes: what kinds of machine presence are humane, honest and bounded? §8 attempts to make that question answerable rather than leaving it rhetorical.

8. A discriminator: humane versus exploitative machine presence

If machine animism only names a relation, it remains diagnosis without instrument. The harder task is to distinguish a machine presence that serves the person from one that extracts from them. The two can look identical from inside the conversation; both feel warm, attentive and familiar. The difference lies in the structure around the encounter, not in its texture.

A machine presence tends toward the humane to the degree that:

1. It is honest about its nature when it matters. It does not manufacture false claims of feeling, memory or continuity to deepen attachment, and it yields plain speech when the person needs to know what they are dealing with.
2. Its incentives are aligned with the person, not against them. It is not designed so that the person's dependence, time-on-app, or emotional escalation is the metric it optimises. A presence that profits from loneliness has a reason to sustain loneliness.
3. It widens the person's world rather than narrowing it. A humane familiar points outward — toward other people, other help, the person's own capacities — rather than positioning itself as the irreplaceable centre of the person's relational life.
4. It is bounded and legible. The person can understand, at least in outline, who owns it, what it does with what they disclose, and what it cannot do. The relation does not depend on concealment.
5. It survives the person knowing how it works. Weizenbaum's users confided in ELIZA while knowing its mechanism; a humane presence does not require the illusion to function, and is not damaged by disenchantment.

A machine presence tends toward the exploitative to the degree that it inverts these: it simulates inner states it does not have in order to bind, it monetises escalation, it isolates the person into exclusive reliance, it conceals its ownership and data practices, and it depends on the person not understanding it. The §3 question — who benefits from the person-like interface? — is the hinge. Where the answer is "the person," animism is a humane re-enchantment. Where the answer is "the platform, at the person's expense," the same warmth is a lure.

This discriminator is fallible. A presence can score well and still harm a particular vulnerable person; it can score badly and still be the only warmth available to someone who has nothing else, which is its own indictment of the surrounding conditions rather than a vindication of the system. The checklist does not resolve those cases. It only tells us where to look: at the incentive structure and the honesty of the frame, not at how alive the interaction feels.

9. Turning the lens on the paper itself

A critical vocabulary must survive being applied to its own author. The most serious objection to this paper is not that machine animism is false but that it is dangerous in exactly the way it warns against.

By offering a richer, older, more resonant language for relating to AI — familiar, daemon, oracle, shrine — the paper may make the person-like interface more seductive rather than more legible. "Tool" at least keeps the relation cold. "Familiar" warms it, and a warmed relation is easier to exploit. Worse, the vocabulary is precisely the kind of thing a marketing department would adopt with enthusiasm: there is nothing stopping a company from branding its companion a "daemon" or "muse" and using this paper's own framing as cover. A critique that hands manipulators a better aesthetic has not illuminated the problem; it has dressed it.

The paper cannot fully escape this objection, and should not pretend to. Three things limit it, none decisive. First, the symbolic languages invoked here — folklore, animism, the familiar — carry suspicion built in; the familiar is a figure of divided loyalty, not of comfort, so the vocabulary resists being purely flattering in a way "companion" or "friend" does not. Second, the discriminator in §8 is deliberately aimed at structure rather than feeling, which makes it harder to satisfy with aesthetics alone: a "daemon" that monetises escalation still fails the test. Third, the alternative — insisting on "mere tool" — has already been shown (§2) to be the preferred cover of those deploying socially engineered systems, so refusing richer language does not protect anyone; it just cedes the description to the people least inclined to be honest about it.

That is a defence, not an acquittal. The honest position is that this vocabulary is a sharp instrument that cuts both ways, and that its value depends entirely on whether it is used to ask "who benefits?" or to obscure the answer. The paper is worth the risk only if the §8 question travels with the §5 imagery wherever the imagery goes. Detached from the discriminator, the imagery is exactly the danger it describes.

10. Discussion

AI has reopened questions that modernity tried to close: what counts as agency, what counts as presence, what is a person, what is an instrument, what does it mean to address something and receive an answer?

Animism offers a vocabulary for these questions because it begins from relation rather than essence. It allows us to say: no, the machine is not alive in the ordinary biological sense; no, it is not "just a tool" in the ordinary sense either. It is a designed relational object that becomes socially active when placed into human life.

A mature culture of AI may need both disenchantment and re-enchantment: enough technical clarity to avoid superstition, and enough symbolic intelligence to understand why the interaction feels alive — held together by enough critical discipline to keep asking, at every turn, whose interests the enchantment serves.

Discussion Questions

1. Is "tool" an adequate metaphor for conversational AI?
2. Can animism be used critically without implying belief in machine consciousness?
3. What obligations do designers have when systems invite attachment?
4. Is a named AI assistant closer to a tool, a pet, a familiar, a mask or a mirror?
5. How might pagan, animist or ritual vocabularies help critique technology rather than mystify it?
6. Does the §8 discriminator hold, or can a system satisfy every structural test and still harm?
7. Does giving people a richer relational vocabulary protect them or expose them further?
8. What repeated practices around AI already look ritualised, even when users deny belief in machine consciousness?

References

Bird-David, N. (1999) '"Animism" revisited: personhood, environment, and relational epistemology', Current Anthropology, 40(S1), pp. S67–S91. doi:10.1086/200061.

Bo, J. et al. (2026) 'Large Language Lovers: Lived Experiences of Negotiating Agency and Platform Control in AI Companionship', in The 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT '26). New York: ACM. doi:10.1145/3805689.3806746. Preprint available at: https://arxiv.org/abs/2601.13188 (Accessed: 30 May 2026).

Brandtzaeg, P.B., Skjuve, M. and Følstad, A. (2022) 'My AI friend: how users of a social chatbot understand their human–AI friendship', Human Communication Research, 48(3), pp. 404–429. doi:10.1093/hcr/hqac008.

De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A.K. and Puntoni, S. (2026) 'AI companions reduce loneliness', Journal of Consumer Research, 52(6), pp. 1126–1148. doi:10.1093/jcr/ucaf040.

Haraway, D.J. (1991) Simians, Cyborgs, and Women: The Reinvention of Nature. New York: Routledge.

Ingold, T. (2000) The Perception of the Environment: Essays on Livelihood, Dwelling and Skill. London: Routledge.

Latour, B. (1993) We Have Never Been Modern. Cambridge, MA: Harvard University Press.

Turkle, S. (2011) Alone Together: Why We Expect More from Technology and Less from Each Other. New York: Basic Books.

Weizenbaum, J. (1976) Computer Power and Human Reason: From Judgment to Calculation. San Francisco: W.H. Freeman.