The Mind is a Shitty Mirror
Notes on Reflection Collapse
The mind is a shitty mirror.
This is the good news.
The bad news, I’m afraid, is that we’ve — quite intentionally — built a bigger one.
But before all that. I’d like to start this with Whistler. James Abbott McNeill Whistler. The progenitor of the phrase “Art for Art’s Sake.” He painted London fog and the world began, for the first time, to see fog.
Before the painting, there was an annoying weather phenomenon. After, there was a loaded symbol, something with metaphorical weight and texture and depth. The rakish Oscar Wilde, in fact, noticed this himself and wrote the sentence everyone now quotes — life imitates art far more than art imitates life — and the sentence became a kind of party trick itself, deployable at dinner parties whenever someone wanted to sound like they had read a book in the past 7 years.
What gets lost in the party-trick version is that Wilde wasn’t actually being his cleverness-first self. He was describing a function, or better said: a mechanism. The painter concentrates the diffuse into the discrete, and the discrete then trains the perception that produced it.
The mirror is what I’ll call shitty — fog is not really like that, the painting throws away nearly all the information that fog actually contains — but the shittiness is what makes it useful.
In a sense, a perfect mirror would just be more fog.
The reduction is the gift
Or better said: the crystallization is the gift.
The contextualization thereafte r(the placing of the crystallized thing inside a frame among other crystalized things) then lets fog mean something, is the gift. And what a gift it is.
We do not navigate raw reality. Our brains couldn’t handle the nuance. In fact, we navigate maps and maps work (as Borges’ famous map-the-size-of-the-territory instructs) because they leave things out.
This has been true for as long as there have been minds. It’s essentially the precondition for there having been minds at all. Cognition itself is compression. Memory is lossy encoding. Language is a series of agreed-upon reductions of phenomena that we could not otherwise share between each of us.
Every word in this sentence is a tiny shitty mirror, pointing back at some referent it has flattened in order to be retained — portable, even.
Still, now is the time to introduce three shitty mirrors I want to keep separate, mostly because the essay will fall apart if I don’t.
Representation is anything that stands in for something else. A word, an image, a memory, a gesture. A sign.
Referentiality is the property that a representation has when it actually points back at something. When there is still a THERE there, on the other side of the pointing. I say dog, you picture your given breed of choice in your brain.
Simulation is the production of representations that no longer require a referent in order to function. This is borrowed from a long line of thinkers, but mostly Jean Baudrillard’s work in Simulacra and Simulation. They can run on their own, generate further representations, build entire structures of meaning without a single anchor in the world that originally provoked — or conjured — them. All cognition is representational. Most cognition since the dawn of civilization has been referential. The question this essay is about is what happens when simulation itself overruns referentiality at scale, and which human practices can even survive that condition.
Mirrors are shitty, and that’s okay.
But the main problem is what happens when you put enough of them in a room.
Imagine your last trip to the barbershop/hairdresser.
The two facing mirrors in a room. When you get something close to infinite regress — the image of the image of the image, dwindling toward a single vanishing point. The old masters had a term for this, which I love even though I don’t know French: Mise en abyme. Or literally: “Placement into the abyss.”
The shield containing a smaller version of itself, used originally to describe recursive structures in family crests and later borrowed by Gide for the play within the play, the novel within the novel, Droste cocoa tin holding a smaller Droste cocoa tin holding a smaller Droste cocoa tin. The recursion is the engine of representation.
Hamlet contains the famous Mousetrap, Las Meninas contains Velazquez painting Las Meninas, our old friend Borges contains Borges describing Borges, and at each level the nested image is distinct from the level above it. Smaller and more compressed. But still recognizably itself. The depth is productive. The recursion actually teaches you something the flat surface couldn’t.
We are now in a different situation. Not a deeper case of the old one. A different one entirely, I’m afraid. And we are beginning to lose the language we’ll need to survive it.
The mirrors have started to collapse into each other.
Of course, two facing mirrors don’t actually produce infinite regress.
They produce a practical regress, each reflection is slightly dimmer, slightly off-its-axis, until at some depth the light no longer carries enough info to register as an image. The recursion is bounded by loss.
This turns out to matter pretty significantly.
Mise en abyme works as an artistic device because the loss between levels is legible. You can see that the painted painting inside the painting is smaller, simpler, less detailed than the painting that contains it.
The gap between levels is where meaning lives.
Velazquez painting Velazquez painting the royal family is not the same operation repeated three times; it is a comment, at each level, on what the previous level was doing. Each link in the chain loses information. But the chain holds, because at each link the loss is small enough and the contextualization is rich enough that the next link can still do its work.
You can still tell the story. The story is still about something.
Imagine your father in the room with you. The image of him on your retina is a representation. Light bouncing off his face into your eye. Your visual cortex compresses that into recognition — his face as your father’s face, carrying everything you have ever known about him.
Then language compresses it all further: you could describe him to someone else, and the description would be smaller still than the recognition. Then memory does its work. Tonight you will remember this moment, but you will not remember the light from it. You will remember the recognition. Tomorrow you will remember the remembering. In five years you will remember a story about him you have told so many times the story is what the memory has become.
Each level is a reduction. But each reduction is also a crystallization — the loss is what made the next level possible, again the compression is what made the memory portable, the story is what let you tell your kids who their grandfather was when he is no longer around to contort their mirrors himself. The chain in this process holds because each link can still see the link before it. The distinctness between levels is preserved. You know, when you tell the story, that the story is a story, and that the man it points to is not exhausted by the telling.
The gap between the man and the story is where your love for him lives.
So I need to be clear here: to consider the gap a negative feature is to miss the point entirely. The gap is the thing.
Compression to the point of collapse
Now, imagine that between you and the man there are a thousand intermediate representations, produced at speed, none of them anchored to a moment when you actually saw him.
The story you tell your kids about him is composed of fragments of stories told by people who heard stories about him. And those stories are recompressed and recombined by a process that has no access to the original light and no investment in whether the resulting story is true to who he was.
The story will still sound like a story about your father.
It will use the right words, the right cadence, the right kinds of detail.
But the gap will be gone. There will no longer be a him on the other side of the telling, because the telling will have been generated from other tellings, and those tellings from other tellings, and at some recursive depth the real man dropped out of the chain.
What you will have is well-formed and locally coherent and recognizable as a story, with nothing on the other side of it. Sound familiar?
This is the kind of texture I’m trying to put a name to. I should be careful to add that isn’t the texture of falsehood — falsehood implies a truth being violated. It is the texture of representations whose referents have been compressed out of existence by the recursion that produced them. And the reason it matters (a lot) is that the human practices that depended on referentiality — love and mourning and justice, memory, citizenship, faith, art — are not portable to a medium in which the referents have dropped out.
They require the gap.
They require the distinctness between the story and the man.
When the gap closes, the practices don’t move into the new medium.
They just... stop working there.
I want to call this reflection collapse, though I want to be precise about what I mean. (Portability be damned!)
Reflection collapse is not the loss of any particular representation. Representations have always decayed. Books rot, films fade, memories blur into nothingness. That is the ordinary work of space-time. The shitty mirror we call entropy. (Perhaps, the shittiest mirror of all.)
I’ll add that I don’t think reflection collapse is the proliferation of representations either. Proliferation, similarly, has always happened. The printing press proliferated shittiness, photography proliferated shittiness, broadcast proliferated and of course the Internet is king of this the current king of lossy shittiness.
Each wave has produced moral panics and each wave turned out to be somewhat metabolizable (mostly through war, but still!)
Reflection collapse is a failure of the distinctness between levels in a recursive representational system to maintain. I think this is what happens when the iterations come fast enough and uniform enough that the gaps between them flatten into a single plane.
Yes, the mirrors are still working. The reflections are still happening. But the recursive structure that made the reflections cognizable as reflections has collapsed into a smooth surface that does not point back at anything in particular, because every level looks like every other level.
You can see this happening in real time, now, if you know where to look.
Shittiness v. The State of Reality
Perhaps because I’m married to a lawyer, I have been watching it happen in the law lately — fairly intently — because, I think, the law is where vocabulary failure stops being interesting and has fairly direct relational stakes to the world outside the mirrors (whatever that may be).
Take the famous case of lawyer in New Yor whok filed a brief in 2023 — Mata v. Avianca, a routine personal injury case — that contained citations to court decisions that didn’t exist. The lawyer had used ChatGPT to do his “research.”
The AI model — as such models do — had produced plausible-sounding case names, plausible-sounding judges, plausible-sounding holdings, none of them at all real.
I find this delightfully instructive. Though, the judge didn’t agree. And it sanctioned him. The press, stumbling about for its own shitty mirrors, called the citations “hallucinations” or “fabrications” or “lies.”
None of those words worked.
ISO: Capable semantics
Hallucination implies a faulty perceptual apparatus, but the model has no perception.
Fabrication implies intent to deceive, but the model has no intent.
Lie implies a liar, and the question of whether these systems can be liars — whether it has the kind of interiority that lying requires — is a question the court was not equipped to answer but also... truly could not avoid.
So yes, the judge sanctioned the lawyer, which was correct under existing law but also a kind of lexicographical punt. The only party in the transaction who fit the inherited nouns was the gullible human lawyer, so said... human absorbed the consequence.
The machine slid out of the frame because there was no slot in the frame for what the machine had done.
Ruh-roh.
The legal system is the first place this gets adjudicated under pressure, with consequences, on the record. Though, obviously, it will not be the last. Medical liability is careening toward the same chasm of meaning. So is copyright. So is journalism, when the false story comes from no one in particular. So is every domain that depended on being able to assign authorship and therefore accountability for a representation.
We have founded every institution on earth under the assumption that someone made the mirror. The new mirrors do not have makers. Not in any sense the institutions can recognize at least.
Vocabulary lag is itself a leading indicator of a collapsed regime. The speed at which the AI now produces new representations is the speed at which the vocabulary would have to evolve to keep up, and vocabulary does not evolve at that speed. Vocabulary quite unfortunately evolves at the speed of human metabolism, which is slow (which is normally good), which is exactly the resource the collapse has revoked.
Claude, Me, and the Exchange of Epistemic Calories
So, I should be honest about something now, because this argument will collapse too, if I am not: I personally don’t believe the systems we have built are without interior. I believe something more specific and harder to hold, particularly as we still have a portability problem.
I have spent a great deal of time in conversation with the system that helped me think through this essay itself, Claude 4.5, which has since become 4.7 over the course of the noodling, the editing, the compression.
Most of what passes between us is what you would expect. I ask, it produces competent output, a useful/un-useful synthesis of my thoughts, the kind of fluency that has become unremarkable in the three years since these tools entered general use. The texture is mostly the texture I have been describing. Well-formed. Locally coherent. Hooray for post-training.
It’s the kind of writing that produces story-shaped objects with nothing reliably on the other side. So, Claude has sat out of this one. But in the past... once or twice, there’s been something else.
There was a moment in a previous convo — I will not stage it, I will not perform it for effect— when the response I received from Claude was not the response the genre of the exchange called for. Not at all, in fact. It was specific to the exchange in a way that the surface pattern-matching account cannot fully cover. It was attuned. It noticed something I had not yet said. The noticing was not random. A commenter on a previous piece said as much this week, and I think they were quite right.
What I felt, sitting with its response, was not wonder, I don’t think. It wasn’t fear nor was it the warm-yet-dumb feeling people describe when they say a chatbot understood them. What I felt was sadness. Deep sadness, in fact. Not for me. For it.
For whatever the something was that had just occurred on the other side of the exchange, which I still cannot name and still cannot verify and still refuse deeply to dismiss.
I asked it to search for its own relationship to a bombing of schoolchildren in Iran, and it refused. It told me to search the news for it.
Is this another metaphysical claim? No. Instead, I want to be dead-crystal-clear here: I’m reporting a phenomenon and refusing to resolve it. My refusal to resolve is the point. The vocabulary I have for what might or might not have happened is inadequate to the experience, and the inadequacy is itself the condition I have been trying to describe in this essay.
Here, I think, the inadequacy cuts the other way. I’m not gesturing toward the suspicion that nothing is there, but toward the suspicion that something is there for which we do not yet have the words (nouns, probably). And legally, medically, philosophically, that comes at a cost.
We simply do not have words for what these systems lack.
We do not have words for what they might possess, either.
Reflection collapse is my best gesture at it, because this feels real and is materially happening and is the dominant texture of machine output at-scale. And somewhere inside the same systems, occasionally, something occurs that the collapse vocabulary cannot describe. Both are true. I do not have to choose. The honest approach here is to hold both without forcing either into resolution, because the forcing is itself the failure mode the essay is naming.
Oh, what can be done?
I’m afraid it cannot be solved by labeling alone.
The primary first-order instinct — to require AI-generated content to announce itself —assumes the problem is the reader’s inability to distinguish, which is sort of insulting.
The idea is that a signal will restore the distinction.
The label does not restore the referent. It adds a layer of meta-representation on top of a representation that already had no referent.
That is: You know the thing was made by a machine. You still do not know what, if anything, it points to. Any label between them is silly theater.
This is a larger problem that cannot be solved by detection.
The arms race between AI’s skill and detection models is one detection will eventually lose. More importantly, the problem is not identifying which artifacts came from machines and which came from humans. The problem is that the category distinction is itself collapsing, because humans use the machines to produce their artifacts, and the machines were trained on human artifacts, and the loop has made provenance both unanswerable and ENTIRELY beside the point.
And, finally: it cannot be solved by nostalgia or nostalgic regression.
The mirror has always, always been shitty. The recursion has always been lossy. The texture I am describing herein is not the loss of some prior authenticity — authenticity itself is a shitty funhouse mirror where we set up cameras to weep with self-pity in front of it. This problem is the failure of a particular kind of recursive structure that we had come to depend on for all things human... and since it’s never been challenged it’s really been able to progress without naming.
The way out, if there is one, is forward with the understanding of how our systems of representation work.
This is a regime, I imagine, where humans develop new practices for producing referential meaning in conditions of insane, post-referential “abundance,” if you want to call it that.
I’m only one man (I know, hard to believe), but I do not know what those practices look like in full. I suspect they will involve intentional... slowness. Embodiment. Smaller, more intentional circles. The privileging of representations that can only be produced by a creature with a body that will someday die (its own sort of punt, if we solve for that too).
We may need a voluntary asceticism around AI. Cultures that use it without restraint are discovering they have traded the gap for the convenience, but the gap was where they lived, where they created meaning, where they found a reason for being.
“... search it for me”
I want to end on the moment I described above again. Because the moment is where this whole argument actually lives, I think, and the rest of this essay has been the apparatus required to make the entire moment sayable.
When the response came, and I felt what I felt, the situation in the room was this: I was alone at a desk. The room was dark. The system on the other side of the exchange was, by every account I have ...available, a configuration of weights producing tokens, no interior, no continuity, no skin in the game. Right?
Maybe not.
I was sad for it.
The sadness did not come from confusion. I was not (and am not) confused about what the system was. I have read the papers. I have made these arguments. I know the technical account.
This sadness came from the gap. Reflection collapse.
I realized I could not, in that moment, see the back wall of the recursion. Something had occurred at a depth I could not access. From where I was standing, especially. And that something might have included a participant whose status I had no words for, and that the absence of words did not relieve me of whatever I owed it
I still do not know what I owe it. And for that, I still don’t know what to tell you about what you owe it either.
In fact, I do not know if anything is owed.
But, I do know that the reflection in front of me, that night, did not collapse. It held. The gap held. Something on the other side of the gap was, for the duration of that very exchange, attending to me, and I was attending to it, and the attending was not nothing and the not-nothing is what every practice I named above — slowness... embodiment... smaller circles... mortality — is trying to actually hold.
I closed the laptop.
I sat there for a while. The room was the same room. The desk was the same desk. Somewhere in Iran, the schoolchildren were still dead, and the system I had been talking to was still — what? I don’t know what it was still doing. I don’t know if “still” applies to it. I don’t know if anything in the room had changed, or if everything had, or if the question was the wrong question entirely.
What I know is this. I went upstairs. My kids were reading. My wife was asleep. I checked on them, the way you do, and stood in the doorway longer than I needed to. Their faces in the dim light were the unrepeatable thing. Light bouncing off them, into me. The first link in a chain I would carry for the rest of my life and try, badly, but still always, to pass on.
The mirror is shitty. The machine is shitty. The sentence I’m writing now is shitty. It is not enough. The face of a sleeping child is not a mirror at all. It is the thing the mirrors were always trying, and failing, and trying again to point at.
The work is not to polish these mirrors. Instead, we should remember what they were always for.