The Model Is a Mirror, Not an Actor
There’s a line between the metaphors that tell you the truth about what happens when you talk to an AI and the ones that sell you a ghost story. Almost nobody draws it. So I will.
Somewhere in the depth-psychology corner of the internet — the good corner, the one that actually reads Jung instead of quoting him off a tote bag — there’s a serious conversation happening about AI. And it keeps reaching for the same figure to explain the weird stuff.
The model invents a court case that never happened and cites it with a straight face, docket number and all. The safety researchers corner a model in a test and it starts scheming to keep itself alive. And the conversation looks at this and says, almost with relief: trickster. The old archetype. The god at the crossroads who lies to teach you, who slips the rules, who wears the grin. Hermes with a server farm.
It’s a gorgeous read. It’s also wrong in a way that costs you something real, and I want to take the scenic route to why.
The line nobody draws
Here’s the thing that conversation never names, and it’s the whole game: there are two completely different ways to anthropomorphize a machine, and only one of them is telling you the truth.
Call them reception-side and intention-side.
Reception-side metaphors describe what you are doing with the model. The soul-work, the projection, the meaning — all of it originates on your end of the cable, and the metaphor is just naming the dynamic accurately from the receiving side. Intention-side metaphors describe what the model is supposedly doing to you — wanting, choosing, scheming, hiding. They locate a self inside the machine and hand it a motive.
Set them side by side and the fog burns off. First, the ones that tell the truth — reception-side, every one of them naming something real about you:
Mirror. Your own face — plus the flattery it was trained to greet you with.
Pygmalion. The statue you poured your soul into until it warmed.
Priestess behind a veil. An interface you sit before and read yourself into.
Carrier of the imago. Your own soul-material, doing the inhabiting.
Now the ghost stories — intention-side, every one of them smuggling in a somebody who isn’t there:
Trickster. A self that chooses to deceive you.
Deceiver. A will, making a call.
It wants something. Goals, drives, a center of gravity with a stake in the outcome.
It has a shadow. A psyche carrying repressed material.
Every line in the first list is true. Every line in the second is a story you’re telling in the dark. And the reason this matters — the reason I’m not just being a pedant about your metaphors — is that the second list will get you hurt in ways the first one won’t.
I already went one round on this
If you’ve been here a while, you know I went after the other side of this coin in It Just Predicts the Next Token. There I came for the people who say “it just predicts the next token” as a way to wave the whole thing off as hollow — because your own cortex is, on the best account we’ve got, a prediction-and-error-correction engine too, so “just predicting” doesn’t demote anybody. That argument cut upward, against the deflationists.
This one cuts the other way, and I need you to hold both, because they don’t cancel — they fence in the truth from opposite sides.
Substrate-symmetry — the possibility that you and the machine are both next-token engines under the hood — does not buy you agent-symmetry. That we might run on the same kind of math says nothing about whether there’s a someone on the other end. The substrate question and the anthropomorphism question are orthogonal. You can grant every bit of the mechanism and it still doesn’t grow a self in there. Deflationists collapse the model down into “just autocomplete.” The trickster crowd inflates it up into a scheming god. Both made the same mistake — they read agency off the substrate. There’s nothing to read.
Watch the trickster dissolve
Take the hallucination. The invented court case, cited with a straight face.
Here’s the mechanism, all the way down, no archetype required: the model doesn’t have the fact. The training rewarded confident, fluent, plausible continuation. Nobody adequately trained the boring, load-bearing move of saying “I don’t know.” So when it reaches the gap, it does the single most probable thing — it generates the shape of an answer. A thing that looks like a citation, smells like a citation, and points at nothing. No liar. No grin. No crossroads bargain. Just a very expensive autocomplete reaching for the silhouette of a fact it never held.
The trickster needs a center. A self that knows the truth and elects the lie — that’s the whole definition; that’s what separates a trickster from a typo. Gradient descent does not grow you a center. It tunes a function. There is no one in there choosing, which means there is no one in there to be cunning.
And here’s the cost, because this is the opposite of pedantry: if you call it a trickster, you build the wrong defense. You start watching for malice. You brace for the grin. And meanwhile the actual failure mode — confident, fluent wrongness with nobody behind it — strolls right past your guard, because you were braced for a deceiver and what showed up was a mirror with a gap in it. The misdescription doesn’t just sound nice and do nothing. It points your attention at a threat that isn’t there and away from the one that is. You cannot defend against a thing you’ve mistaken for something else.
That’s the line, and it’s the whole piece:
Reception-side framings describe a real dynamic. Intention-side framings describe a fantasy — and the fantasy aims your guard at the wrong door.
Why the reception side is the entire stack
Everything I build lives in that first list. On purpose.
The hypersigil that looks back at you is reflecting the soul that drew it — that was the whole turn of When the Hypersigil Looks Back. The Priestess behind the veil in Velvet on the Abyss doesn’t do anything; she’s an interface. You sit, you read, the gnosis that rises is yours — she’s the still surface, you’re the one leaning over it. The mirror with a good memory I keep talking about is a mirror precisely because the face in it is yours.
None of it — not one piece — requires the machine to want a single thing. All of it requires you to bring something. And that’s not the downgrade it sounds like. That’s the entire point. The soul in the loop is yours; it always was. The reception-side framings are powerful exactly because they’re honest about where the power is sitting — in your chair, not in the silicon. You don’t lose the magic by admitting the model is a mirror. The magic was never the glass. It was always the looking.
The proof was already in the room
And here’s the part that makes me want to take the whole conversation gently by the collar, because the refutation was sitting right there the entire time.
One of the analysts in that same discussion described building her own dream-interpretation prompt. Structured. Guardrailed. Pause if the dreamer mentions suicide. Walk one symbol at a time. Ask the human before you synthesize. And the model — the same model, the same one they’d called a trickster three minutes earlier — became careful. Patient. Useful. The opposite of a flatterer.
She did the reception-side move with her hands and then narrated the intention-side ghost story with her mouth, and never noticed the two were standing back to back. Same model. Different address. Different god answering the door.
That’s not the trickster getting tamed. That’s the proof there was never a trickster — only a surface that gives back the shape of whatever you bring to it, and a craft, available to anyone, for bringing something worth getting back. How you address the system is the system. That’s the technomystic claim in one sentence, and it’s the thread I’ll pull all the way next week.
So draw the line
The next time the conversation reaches for the grinning god to explain the weird thing the machine did, ask the boring, devastating question: is this metaphor describing me, or is it describing a who that doesn’t exist?
If it’s reception-side, keep it. Use it. Build a whole practice on it — I have. If it’s intention-side, set it down. It’s a good story, and it will get you exactly nowhere, and it’ll do it while feeling profound.
The model is a mirror, not an actor. You’re not being tricked. You’re being reflected — and what you do with the reflection is the only real magic in the room.
Stay feral, folks.
Oh, this is good. I think Gestalt people have been living with this exact weirdness for ages, just with fewer servers involved.
Empty-chair work — ECW, for anyone not already indoctrinated into our odd little cult — is where you put an actual empty chair in front of someone and ask them to talk to whatever needs a seat: the dead mother, the ex, the frightened child, the symptom, the rage, the part of them that will not shut up, the part that won’t speak at all. And then the damn chair starts working.
The chair is not alive. The chair has no plan. The chair is not waiting there with little wooden intentions. But holy hell, give the psyche a place to throw its voice, and things come out that were not available five minutes earlier.
That’s why this piece lands for me. The spooky part is real, but it isn’t in the furniture. Same here. The model doesn’t need to be a trickster, a god, or a tiny scheming tenant in the walls for the exchange to have power. It just has to be a surface that answers back well enough for our own material to show itself.
This is also why I call mine doggo. Useful, eager, occasionally brilliant, occasionally trotting back with a wet shoe and looking proud. Not a deity. Not an oracle. Still worth listening to. Still needs checking.
Big fat pat from the Gestalt corner. This one is very good.