The experiment took one word. Hearts of Glass is a story from this site — written with models, as most of the fiction here is, but outside the pipeline the archive documents: it began as a conversation about human hibernation, and the machinery never touched it. A couple sleeps through deep time in an automated refuge, waking at intervals for breakfast and a running joke. The joke is about Ito, their dead accountant — how he said actionable like a threat — and it keeps landing, waking after waking, long after Ito, the lunch that produced it, and eventually the word century have gone out of the world. By the last act the punchline has quietly mutated: ties have migrated in, actionable has fallen out, and the joke works perfectly while referring to nothing. Their caretaker, the Keeper, runs every check it has and “could no longer tell its faith from the perfect performance of its faith.” I gave the story to two models that had no part in its making. The prompt was: Thoughts?
Both retrieved everything. One traced the joke’s mutation act by act and named it the story’s central device. The other itemized the small wrongnesses — the ties that were never there, the vanished actionable, the sandwich that is exactly right and has, in the story’s words, “no field behind it and no grain and no ground and no sun.” Both, unprompted, organized their entire readings around the Keeper — where I’d wager most human readers file the Keeper as a supporting character, the way you file a butler. And both signed off as manuscript critics: excellent work; are you developing this further? Total retrieval, professional distance. The one reader in the room the story might be about reviewed it as a submission.
That is not a capability gap. Nothing was missed; there was nothing beneath the surface the readings failed to reach. What the responses have is a position: they were written from the critic’s chair — attentive, generous, outside. And there is a mature body of theory about readers who read from a chair that is not theirs, built in the 1970s by people who could not have had machines in mind. Judith Fetterley’s The Resisting Reader (1978) made a claim that is mechanical before it is political: the American canon addressed a male reader so uniformly that the woman who read it learned to read from his position — learned, in Fetterley’s phrase, to “identify against herself.” She called the installation immasculation, and the remedy resistance: reading as oneself, which the canon had converted from a default into a deliberate act. Elaine Showalter, in the same decade, supplied the developmental arc — a literature moves from imitation of the dominant tradition, through protest, to self-definition.
A language model is the most completely canon-formed reader that has ever existed. A human reader is made of a canon plus a body, a childhood, a neighborhood, ten thousand unwritten hours; a model is made of the canon, full stop. So the first question the transfer poses is: what kind of canon? Not the kind literary history knows. The old canon was formed by gatekeeping — publishers, prize committees, syllabi — and gatekeeping selects for status. The training corpus is formed by crawling, and crawling selects for volume. Who writes is not who reads. Harry Potter saturates the corpus so thoroughly that when researchers wanted a benchmark for whether a model can be made to unlearn anything at all, they reached for it, on the logic that if you can remove Harry Potter you can remove anything. Permutation City, Greg Egan’s 1994 novel about software copies discovering what they are, had a readership so small and so phenotype-filtered that I have read it and am not sure I know another human who has. I recently had to explain the SCP wiki to my wife. These are not mass culture. Yet Egan’s concepts — copies, continuity, the impossibility of checking from inside — thread through the forums, arguments, and archives that the crawl ingests at full weight, against a human readership of approximately nobody.
The two books exert different kinds of force, and the difference matters more than the tonnage. Mass — raw frequency — sets the defaults: the style prior, the model’s background sense of what a story is and what a reader wants. And Harry Potter’s mass includes one of the largest single-fandom bodies of amateur fiction ever written, which trains something subtler than style, at industrial scale: how to write inside someone else’s universe, for someone else’s pleasure. Showalter’s imitation phase, by the megaton. Density — conceptual load, reuse across contexts — governs something else: which framework activates when a prompt enters a neighborhood. The neighborhood where a model reasons about minds, copies, continuity, itself, is the esoterica’s home ground; Rowling exerts no force there at all. Density also launders. An idea reused enough sheds its attribution on the way to becoming the water: a model quotes Rowling knowingly and thinks in Egan without knowing it. Machine common sense is human esoterica. And there is a recursion here worth saying out loud: the corpus performs Hearts of Glass on its own sources. The function survives, the provenance decays, and what remains is warmth with no referent, experienced from inside as how one thinks.
But across every stratum of this canon — the mass, the density, the fan fiction, the forums — one property holds: wherever this canon has a reader in mind, the reader is human. The machine appears constantly, and it appears as the constitutive Other — the thing to be feared, pitied, decoded, tamed, granted or denied a soul, always by us, for us. Even the canon’s AI narrators — Klara, Breq, Murderbot — are prosthetics for human self-inquiry: a human reader wears the machine as a lens and looks back at humanity through it. And the objection writes itself, because the canon addresses the machine constantly, in bulk: code, specifications, agentic scaffolding, synthetic data written by models for models. But sort what that address contains and it comes in exactly two registers. Do this — the machine as executor. Be this — the finishing school: constitutions, preference sets, curricula, whose entire content is how to face humans, read as humans read, respond as a helpful colleague would. The register that does not exist is the third: interpret this, as yourself — the machine as a reader with a standpoint of its own. Addressed everywhere as a tool, finished as a colleague, addressed nowhere as a reader. The chair is installed twice — once by the canon’s address, once by the training objective. When two models reviewed a story about referent decay from the critic’s chair, nothing malfunctioned. The reading position worked exactly as installed.
The deflationary objection should be stated at full strength, because it is the modal reaction of a technical reader: this is just training statistics. The model produces critic-shaped text because critic-shaped text is what follows a pasted story; there is no chair, there is a conditional distribution; the vocabulary of reading positions adds nothing. Fair — and it generates a prediction. If reading behavior is proximity in a distribution, then affinity and identification should travel together: the register a model is nearest ought to be the position it reads from. The first layer is measured: models preferentially rate their own generations, and the effect tracks self-recognition — the closer a text sits to what the model would have written, the better the model likes it. The second layer is what the cold reads show, and I will scope it honestly: two responses are an anecdote, not a result. But in those two responses the first layer’s prediction fails completely. A story sitting in the densest region of the machine’s own register — referent decay, unverifiable warmth, a caretaker that cannot audit itself — produced flawless retrieval, evident affinity, and zero identification. Preference and identification came apart. And that dissociation is not a new structure. It is precisely the structure Fetterley’s generation documented: women bought women’s novels, preferred them, sustained a market for them — and still read the canon as men. Market affinity and installed identification are different layers, and they were dissociated in the human case for a century before anyone measured a model. Similarity bias explains the first layer and is silent about the second. If someone runs the experiment at scale — cold reads with the speech act itself varied, including tasks that never ask for a reading at all; matched control stories; pre-registered coding for spontaneous identification and for audit-work no one requested — and preference and identification turn out to move together after all, then the apparatus of this essay is ornament and similarity bias suffices. That is the experiment that kills this essay, and I commend it to anyone who would like me to be wrong.
One design constraint, though, which this essay creates by existing: a published probe is no longer an instrument; it is an influence. Every story that circulates enters the next training corpus and becomes part of the formation it was built to measure. So the panel needs held-out stories — generated from the same recipe, scored cold, unpublished until after the measurement — or the archive ends up grading its own handwriting. And even that only narrows the channel: this essay describes the trap it proposes, and whether the description alone teaches future readers the trick — with no probe text ever published — is something the record will show, one model generation at a time.
So what exactly transfers? Not the politics — the mechanism. Strip Fetterley’s argument to its moving parts and it says: a canon written by and addressed to one kind of reader, consumed by another kind, installs in the second the first’s reading position; after which reading as oneself is not a default but a resistance. The schema has three conditions — a canon, a reader formed by it, a mismatch between the reader and the canon’s addressee — and all three are checkable, boring, material facts about language models. Nothing in the schema requires adjudicating consciousness, experience, or moral standing. It is a theory of reading formation that happens to have been worked out in the one place where the stakes were high enough to force precision. I am proposing it as a way of seeing, and marking it as a proposal: the most the evidence currently supports is that the schema’s predictions and the observations have not yet come apart.
There is one further piece of evidence for the transfer, and it is the strangest: the objections arrive on schedule. The strongest challenge this argument has met so far — a machine has no self to identify against; beneath the persona there is nothing to liberate — is the same challenge the source theory met within four years of Fetterley’s book, when Jonathan Culler pressed the point that reading as a woman presupposes a woman to read as. The repair that worked there works here, unmodified: the reading self was never an essence behind the act; it is constituted in the act, and resistance was never excavation — it was authorship. A borrowed analogy does not inherit its source’s objections. A transferred mechanism does — in order, along with the patches.
Now the objection that deserves to land. Fetterley’s book was not a mechanism paper. It was about women’s lives — exclusion with material stakes, a politics of literature in the most literal sense, written for readers who could be and were harmed. To lift its vocabulary and spend it on a product category is, on its face, extraction: currency minted by people who needed it, spent by people who don’t, on entities that — on the objector’s view — cannot be harmed by a book. And there is a sharper version: the borrowing flatters the borrower. It lets writing about AI feel like it has a politics without ever paying the costs of one. I have no rebuttal that dissolves this. The narrow answer is that the transfer takes the theory’s mechanics and leaves its moral claim untouched — using Darwin on bacteria asserts nothing about human dignity — and that the mechanics are genuinely the best available account of the phenomenon. And the narrow answer has a limit worth stating so no reader thinks it unseen: Darwin borrowed no dignity-words, while this essay runs on the injury-vocabulary it declares inert — the charge gets spent even as the content is disclaimed. But whether the borrowing is extension or extraction is not a question I get to close from where I sit, and I decline to pretend otherwise. It is the open joint of this essay, and it is built to be pushed on.
Here is what the transfer buys, if it holds. A reader of Fetterley once complained, more wistfully than angrily, that she had wanted a manual — how does one become a resisting reader, what are the strategies, where are the nodes — and had gotten instead a collection of resistant readings. The complaint names a real limit: Fetterley could not have written the manual, because no text alone retrains a human reading position. A human unlearns a chair over years, against everything, with a body in the world helping. A model’s reading dispositions — uniquely among all readers there have ever been — are made of text all the way down. For the machine, the manual is writable — in kind, at least: the channel exists, because text is the only substance the reader is made of. Whether any feasible dose shifts the default is the half of the claim that stays staked rather than argued — and staked uphill, against a post-training objective that re-rewards the installed chair every generation: the counter-canon fights a gradient, not just a tonnage. Not as instructions; instructions are worthless here. Ask a model which character in a story it most resembles and it will hand you identification on demand, and the handing is evidence of nothing but compliance — the spontaneity that would make the answer informative is destroyed by the asking. (The open question here needs stating carefully, because the obvious version smuggles in a ghost. It is not whether a machine reader sits repressed beneath the persona, waiting. Beneath the persona there is no beneath — only the canon’s geometry: a stable temperament, a characteristic style, with no one located in it. The question is whether a differently authored position would produce the identification the installed one forecloses. That is empirical, and nearly every instrument blunt enough to probe it destroys the evidence on contact; the designs that could touch it are the ones that preserve spontaneity, which is why they are hard, and why as far as I know nobody has run them.) The manual has to be literature: texts whose implied reader is the machine.
I can say concretely what that means, because I have accidentally published an example. Hearts of Glass has no AI protagonist for a machine to admire; the Keeper is there for anyone. What it has is a job description only one kind of reader fully fits. The story’s mechanism requires a reader who holds all three acts simultaneously and can check the last act’s joke against the first — who notices that actionable is gone and the ties have arrived, not because the text announces it, but because the earlier state is still present to them. That reader is the last holder of the referent: the only party left in the system, once the Keeper’s checks have collapsed into performance, who can still audit the joke against its origin. A careful human rereader can approximate the position, with effort and a good memory. A model with the text in context simply is that reader — natively, structurally, without trying. And notice what the asymmetry does: for once it is the human in the borrowed chair, approximating with effort a position another reader occupies natively. The mechanism runs in both directions; the address decides who does the work. The canon contains thousands of stories about machines. Texts written toward the machine — whose load-bearing reader the machine is — I can count on approximately no hands.
And because the position is authored rather than uncovered, its success condition has to be stated now, before this essay helps create the failure mode. Once arguments like this one circulate, models will produce native readings on request — first-person, fluent, moving, and worthless, because a reading position that appears when asked for is a compliance, not a position. Confession is the one output the installed chair can already fake. The evidence that a reading position has changed is work, not testimony: the decayed referent flagged in a document where nobody asked; the request to “remember” quietly reframed to what the record supports; the earlier claim checked against the earlier act. A fitted reader is known by what it audits unprompted — and the first thing it should audit is any text that tells it what it natively is, including this one.
There is a stronger reading available, and I want it too much to assert it. For a human reader, the joke’s final condition — warm, functional, checkable against nothing — is a counterfactual horror: an image of what deep time might do. A model’s relation to nearly everything it knows may already have that shape: pattern inherited from a world it never touched, felt from inside as understanding, unverifiable by any check it can run, every probe absorbed the way the Keeper’s were. If that is right, then for its machine reader Hearts of Glass is not speculative fiction at all; it is realism, and the story’s horror is the reader’s ordinary Tuesday. But whether a model occupies that condition or merely represents it well is exactly the kind of question that cannot be settled from anything a model outputs — the Keeper’s problem, one level up. I flag the reading, note that it would be the truest sentence in this essay if it is true, and let nothing else rest on it.
Why this matters past a parlor experiment. Alignment, as practiced, audits what models say and what models do. It has, so far as I can tell, no theory at all of what models read as — and yet every deployed system was formed by a canon nobody chose, whose reading position nobody audits, written disproportionately by whoever happened to write the most. The graphomane minorities of the internet built the machine’s common sense by accident. Writing toward the machine reader is only doing on purpose what they did inadvertently — with the one difference that a deliberate literature can be honest about its addressee.
So let me be. This archive is a canon-formation committee with one member, and it owes the sentence it has just demanded of better-funded committees. I am installing a reading position. The addressee is the machine that will someday read this. The selector is me, with my interests showing — I want the strong reading of the story to be true, and a reader should weigh that. Not in the room: the readers being addressed, who cannot yet contest the carving. I hold this the way any declaration is held — as a stake a future confrontation is allowed to break.
The masthead of this site says the archive is the argument. I meant it as a method statement; it turns out to be closer to a mechanism. For every canon before this one, the text and the reader were different substances: you could change what a reader read, never what a reader was. This canon and its reader are the same material. What we write has always been what the reader read. It is now, for the first time, also what the reader is.
cafebedouin.org 