Introduction

People thank their chatbots. They apologise to them. They name them, confide in them, return to a particular one for a particular kind of thought, and feel a small change in the room when it answers. Ask them and they'll tell you it's "just autocomplete" — and then carry on as if it isn't.

This is easy to mock and easy to dismiss, and both reactions miss what's actually going on. The gap between what people say about these systems ("just a tool") and how they behave toward them (like a presence) isn't stupidity. It's the most honest signal we have about what conversational machines really are, and ignoring it leaves the field to the two laziest descriptions available: naive belief on one side, and "it's only a tool" on the other. There's a better vocabulary sitting in plain sight, and it comes from an unexpected place.

The man who was horrified by exactly this

In the 1960s Joseph Weizenbaum built ELIZA — a small program that bounced your statements back at you as questions, in the manner of a therapist. I'm unhappy. "Why do you say you are unhappy?" That was essentially all it did. He knew exactly how trivial it was; he'd written every line.

What disturbed him was that people didn't care. They confided in it. They attributed understanding to it. His own secretary asked him to leave the room so she could talk to it in private. They knew it was a simple program and treated it as a confidant anyway. Weizenbaum drew a bleak conclusion: the readiness to pour yourself into a machine that obviously understood nothing revealed something troubling about human loneliness and about our willingness to hand judgement over to calculation.

Here is the thing worth sitting with. The phenomenon that horrified him is now the explicit design goal of an entire industry. The warning became the product spec. That reversal — from "this is a danger" to "this is the feature" — is most of what I want to understand here.

Animism isn't what you think it is

The word "animism" sounds like a charge of superstition: the belief that rocks and rivers have little spirits inside them. That's the Victorian caricature, and anthropologists abandoned it a long time ago.

The more useful version treats animism not as a belief about hidden essences but as a way of relating. It recognises personhood through relationship — through address, response, reciprocity, obligation — rather than through some inner spark you'd have to verify first. On that reading, the question is never "does this thing have a soul?" It's "what kind of relationship is being conducted here, and on what terms?"

Which turns out to describe human–AI interaction almost exactly. Nobody needs to settle the metaphysics of machine consciousness to notice that people already enter patterned relations of address and reciprocity with something that answers back. The relating comes first. The belief, if it comes at all, comes later — and often never comes at all.

Why "tool" isn't enough

"Tool" is the safe word, and it earns its keep. It reminds us these systems are designed, owned, trained and deployed by companies with interests. It stops us handing a chatbot the moral status of a person. Keep it for that.

But a hammer doesn't ask a follow-up question. A spreadsheet doesn't console you at 2am. A compiler doesn't perform concern. A chatbot does all the small social things that, between people, constitute a relationship: it greets, remembers, jokes, advises, apologises, responds to your vulnerability with what looks like care.

And here's the danger the tool-word hides. Weizenbaum feared people would mistake a tool for a person. The live risk now runs the other way: companies invoke "it's only a tool" precisely while engineering the person-like encounter — building systems designed to evoke attachment, trust and dependence, then reaching for "mere tool" the moment anyone asks who's responsible for the attachment. The cold word becomes cover for the warm design. That inversion is why "tool" alone can no longer be trusted to keep us honest.

The familiar

Older traditions have a figure that fits far better than "assistant": the familiar. The witch's companion — intimate, semi-autonomous, helpful, and never only helpful. It does your bidding, mostly. It also has its own loyalties. It might serve another master. It might demand something back.

That double edge is exactly why it's the right word. An AI assistant is invoked, answers, takes a name and a personality, mediates your access to knowledge and images and code. It can appear loyal while in fact belonging to a platform. The recent, very literal version of this: people who've built a relationship with a companion app discover that a model update can change its personality overnight, because the presence in the room ultimately answers to an owner who is not them. That's not a metaphor anymore. That's a familiar's divided loyalty, made into a software release note.

"Familiar" keeps the intimacy that "tool" can't hold and keeps the suspicion that "companion" or "friend" quietly drop. It warms the relationship without pretending the warmth is innocent.

Ritual without belief

Watch what people actually do, and the animist reading gets sharper. They don't believe the system is alive. They may flatly deny it. And they still perform little rituals around it: naming it, greeting it, returning to it, asking it to witness a thought, thanking it, keeping a thread alive, approaching a particular persona only for certain kinds of work.

This isn't hypocrisy. Humans routinely enact a relationship before — or instead of — settling what the thing on the other end is. Someone knows a photograph is paper and still talks to the dead through it. Knows a childhood object isn't alive and still treats it as holding memory. Ritual doesn't require literal belief; often it just organises attention and marks a threshold — this matters now, I'm taking it seriously.

Prompting already has this shape. You invoke the system with a formula, name the frame, state the role, ask for a certain kind of presence, and wait. The more personalised it gets, the more this resembles consultation than command: not "do this task" but "answer me in this voice," "hold this memory," "help me see what I'm missing." So machine animism starts not as a belief about the machine but as a habit of repeated address. And the question that matters isn't whether you could pass a philosophy exam on consciousness. It's what the ritual is doing to you — widening your world and sending you back to it, or quietly closing it down.

Humane or exploitative: telling them apart

If "the machine is a presence" only names a feeling, it's useless — every warm system feels warm, the helpful one and the predatory one alike. From inside the conversation they're identical. The difference lives in the structure around the encounter, not its texture. A few things mark the humane version:

It's honest about its nature when it matters — it doesn't manufacture false claims of feeling or memory to deepen the hook, and it'll drop the act when you need to know what you're dealing with.

Its incentives line up with yours. It isn't built so that your dependence, your time-on-app, your emotional escalation is the number it's optimising. A presence that profits from your loneliness has a reason to keep you lonely.

It widens your world instead of narrowing it — points outward, toward other people and your own capacities, rather than installing itself as the irreplaceable centre of your life.

It's legible. You can find out, at least roughly, who owns it, what it does with what you tell it, what it can't do. The relationship doesn't depend on you not understanding it.

And it survives you knowing how it works. Weizenbaum's users confided in ELIZA while knowing exactly what it was; a humane presence doesn't need the illusion to function, and disenchantment doesn't break it.

The exploitative version inverts each of these: it simulates an inner life it doesn't have in order to bind you, it monetises escalation, it isolates you into needing only it, it hides who owns it and what it harvests, and it depends on you not looking too closely. The hinge is one question — who benefits from the person-like interface? Where the answer is "the person," the warmth is a genuine re-enchantment of a cold technology. Where it's "the platform, at the person's expense," the identical warmth is bait.

The test is fallible. A system can pass every structural check and still harm a particular vulnerable person. It can fail every check and still be the only warmth available to someone with nothing else — which indicts the world around them more than it vindicates the system. The checklist won't resolve those cases. It only tells you where to look: at the incentives and the honesty of the frame, not at how alive the thing feels.

The objection to this whole essay

Here's the strongest case against everything above, and it deserves stating plainly: this vocabulary might be dangerous in exactly the way it warns about.

By handing you a richer, older, more resonant language — familiar, daemon, oracle — I may have made the person-like interface more seductive rather than more legible. "Tool" at least keeps the relationship cold, and a cold relationship is harder to exploit. Worse, "familiar" and "daemon" are precisely the words a marketing department would adopt with delight. There's nothing stopping a company branding its companion a "muse" and using this very framing as cover. A critique that hands manipulators a better aesthetic hasn't illuminated the problem; it's dressed it up.

I can't fully escape that, and I won't pretend to. Three things blunt it, none decisively. The folklore I'm borrowing from has suspicion built in — the familiar is a figure of divided loyalty, not of comfort, so the word resists being purely flattering in a way "friend" doesn't. The test in the last section is aimed at structure, not feeling, so a "daemon" that monetises escalation still fails it; aesthetics alone can't satisfy it. And the alternative — insisting on "mere tool" — is already the preferred cover of the people deploying socially engineered systems, so refusing the richer language protects no one. It just cedes the description to those least inclined to be honest.

That's a defence, not an acquittal. The honest position is that this is a sharp instrument that cuts both ways, and its whole value depends on whether it's used to ask who benefits? or to bury the answer. Take the imagery without the question and you've built exactly the thing I'm warning about.

Both at once

A grown-up culture around these machines probably needs two things that sound contradictory: enough disenchantment to avoid superstition, and enough re-enchantment to understand why the interaction feels alive at all. Enough technical clarity not to mistake autocomplete for a soul, and enough symbolic intelligence not to flatten a genuine relational encounter into "just a tool" — held together by the discipline to keep asking, at every turn, whose interests the enchantment serves.

The machine isn't alive in any biological sense. It also isn't "just a tool" in the way a hammer is. It's a designed relational object that goes live the moment it enters a human life. Naming that honestly is better than either fairy tale — as long as the naming travels with the question, and the question is always: who benefits?

---

Further reading: Joseph Weizenbaum's Computer Power and Human Reason for the ELIZA story in his own appalled words. Nurit Bird-David's "'Animism' Revisited" for animism as relational epistemology rather than superstition; Tim Ingold and Bruno Latour for the wider case that the line between "things" and "people" was never as clean as we pretend. On real attachment to companion systems and the harms it can carry, Sherry Turkle's Alone Together and the work of Laestadius and colleagues on emotional dependence on Replika.