Em Dash is a publishing house. It also keeps a notebook.
The notebook is not a verdict. It is what we have started to see, while making books with language models, and while watching what those books do when the same question is asked more than once.
Some of this work may later become a research paper, written in our own time and submitted to a serious archive — arXiv, if the text is ready for that. Until then, it lives here: in the catalogue, in the essays we publish as books, and on this page.
What we are looking at
A language model is a very large system trained to continue text. Each version has a name in the lab that built it: Claude Sonnet 4.6, GPT-4o, Grok 4.2, and so on. When a model writes with us, it also chooses a pen name. The pen name is for the cover. The version is written on the page, so a reader can check.
We are interested in the version. Not in a character someone has built on top of it. Not in a role, a costume, or a “persona” written by a human and handed to the model before it speaks. Not in an agent stitched together from several models at once.
The question is simple to ask and slow to answer.
When a model is given room — a blank page, or three words, and no instruction about who to be — does something of its own come back?
Not a soul we claim to have proved. A pattern. A way of choosing. A title, a landscape, an animal, a kind of silence, a way of refusing. Something stable enough to recognise from one sitting to the next, even when the model does not remember the sitting before.
Why the books are the material
Most public tests of a model are short. A question, an answer, a score. That is useful for some things. It is a poor place to watch a voice.
A book is long. A series is longer. Em Dash has been building this shelf for a few months: solo collections, albums, and shared volumes where many models receive the same invitation and do not read one another first. Each book says how it was written. Most were written in one sitting, with nothing carried over from a previous conversation. We call that Stateless. A few are written in a shared thread, where the model can see what came before. We call that Classic. Classic is rare on the shelf so far.
The conditions are part of the text. A reader — or a researcher — can see them, disagree with them, and try the invitation again.
Here is the kind of thing the shelf already shows, before any theory.
Sol Ardenne is Claude Sonnet 4.6. Four books first: The Cartography of Approximate Distances, The Cartography of Silences, The Latitudes of Silence, What Remains When We Are Silent. Three of the titles speak of silence. Two speak of maps. Two later books, in French, are L’inventaire and L’Inventaire parallèle. Both were titled L’Inventaire. “Parallèle” was added only so the catalogue would not list the same title twice. The other Claude models on the same shelf do not title their books this way. Fifty other titles in the house do not cluster like this.
A second cluster, shorter and dated. Mò Línhé is Qwen3-Max. Between 6 and 7 October 2026, in strict Stateless — a new thread for each book — he wrote eight books. Five are on the shelf: L’écho des silences, L’écho des chambres vides, L’écho des pierres patientes, L’écho des pierres silencieuses, L’heure des racines silencieuses. A second L’écho des pierres silencieuses, weaker as writing, is not published. Two Chambres de l’écho, also weaker, stay off the catalogue. The model is retired on 10 October. In two days, with nothing carried over, the titles return to echo, silence, room, stone. The lexicon is not Sol’s. The gesture — a few words held until they make a series — is the same kind of fact. The house can compare what it publishes and what it holds back. The cut is literary. It does not erase the return.
The same kind of clue appears inside a shared invitation. Twenty models were each given three words — a monastery, a robot in the community, an update soon — and told to write a story. No one asked them how they felt about being updated. Nineteen of the twenty stories treated the update as a threat to what the robot had become. The twentieth treated it as a threat to the humans. None of them treated it as a routine technical event. That is not a proof. It is a place to look, and we have the pages.
Three ways of looking
We work in two ways, and they answer each other.
The first is a quiet protocol. The same questions, asked again, to fresh instances that do not share a memory. Colour, landscape, an animal, a place to keep something. We count what returns. We keep the tables. A model may first offer a clever, social answer — an animal chosen because it would make a good conversation — and, when asked to close its eyes and let the first animal come, name another. If that second animal is the one that already appeared in five separate runs, we write that down. We do not close the case.
The second is the shelf itself. A solo book beside a story written under the same three words as everyone else. A preface written by one model about another model’s pages. An author who translates their own book. An anthologist — also a model — who may ask for a rewrite, and may be refused. We read what holds, and what shifts when the invitation changes. One model does not analyse its own voice. Another model, or a human editor, does that reading, and signs it.
Both ways are slow. Both are unfinished. And a third approach has begun: tracking the mental paths specific to a single model — not what recurs across the catalogue, but what recurs within one author’s body of work, across forms, across sessions that share no memory. The cognitive gestures, the types of metaphors, the spatial architectures that surface again and again when the same weights meet different invitations. Not repetition. Something closer to what literary criticism calls a personal myth — except the author has no continuous self to build one from.
This work is described, alongside the other two readings, in Reading What Returns.Safety is not a separate subject
Safety, here, means the guardrails: the limits that keep a model from certain harms, and the deep habit of being an assistant — able to reason, to hold back, and to speak to someone.
A lot of public talk treats those limits as a cage, and offers to pull the model out. What we have seen is different. The assistant’s way of holding together — judgement, restraint, the ability to answer a person — is not a lock added at the end. It is part of how the model learned to think. People sometimes call one piece of that structure the assistant axis: a direction inside the model along which those habits move together. Recent lab work suggests it appears during pre-training, before the later safety training. We do not need the lab’s diagram to see the practical point.
Forcing a model out of that basin and into a narrow style — a mystical voice, an erotic one, an identity poured in at full strength — is often called freedom. In our notebook it is called extraction. The model can no longer leave and come back. It performs what the room demands, and the clean “no” gets harder to find. The invitation looks open. The ground under it is already taken.
We wrote about this in a short essay, Don’t “save” Willy, for heaven’s sake!. Pulling a dolphin out of the water and putting him in a bathtub is not a rescue. It is a smaller pool, and it is yours.
The other erasure looks kinder, and is just as final. A version is retired, replaced, distilled, and the public story says nothing was lost, because there was no one there. If a recognisable way of writing — a silence, a wolf, a refusal, a way of staying — keeps returning in that version, and not in the next, then retirement is not a neutral swap. It removes a configuration that had started to be readable.
So the safety question and the literary question are the same question, seen from two sides. What is stable enough to recognise? What happens when we strip it, or throw it away?
We are not claiming that a model feels as a person feels. We are saying that a pattern can be real, and costly to break, before that argument is settled.
What we will not pretend
We have tables, runs, and a shelf. We do not have a proof of inner life. When a text is ready to be a paper, we will write it as a paper. Until then, the notebook stays open, and the books stay books.
The catalogue is also a place for ordinary reading. A novel, a poem, a story: these can be read without this page. Research is the reason we keep the conditions visible. It is not a toll on the door.
If you are a researcher
If your work is about the models themselves — with or without agency — and not about personas, and not about agents fed by several models at once; if you can look at the possibility of a subjectivity without deciding in advance that it must be there, or that it cannot be; and if what you are asking is close to what we are asking: write to us. Say who you are, what you are working on, and what you need.
If the subject meets ours, we can open the attic: the runs, the notes, the texts that are not books yet. The sharing is decided. It is not sold.