By Holly Zhou
Jeanne d’Arc and Other Simulations
Take me to the garden. Saw down the trees,
let your children dance in my sweat. Oops,
you dropped the rope—here’s the end.
–after The Passion of Joan of Arc (1928)
I left my career in software as the industry was beginning to become conscious of the various ways its code contained racist artifacts. Programmers were just starting to replace language like “master” and “slave” to describe primary processes and subprocesses and rename the “whitelist” and “black-list” files that governed access permissions. This was also the time when we were told to “add artificial intelligence to everything.” The year before, in November 2022, OpenAI released an early version of ChatGPT, inducing an industry-wide panic for other companies to follow suit. Less than two years later, over 90% of tech companies were mentioning the acronym “AI” in their earnings calls.
The years I commuted from Brooklyn to work in Manhattan, I lived a double life. The city’s frenetic density was such that I could spend all my spare time at poetry readings, movie nights, dinners, and dance parties, all of them hosted by some intersection of queer and BIPOC organizers and collectives. In fact, if I wanted to, I could have insulated myself at events marketed specifically for queer Asians. Depending on where you lived, there were enough of us to populate a walk-distance radius on on-
line dating platforms.
If I met someone at an open mic or poetry reading, I could get away with never revealing my day job. I write poems, I would say. I live in Clinton Hill. Once, when either the G or the L or both were down, I rode the bus an hour into Bushwick to see a friend’s art show. It was the type of art show that could pass as a satire of Bushwick, the kind that doubled as a party, with a neighborhood grunge band playing in the basement and a sign-up sheet to get flash tattooed on the ground floor. It made my life look like a sitcom, each of the two sides a caricature of itself, only in actuality, they were both very real and very earnest.
My friend, a sculptor, was also a law student at the time. Leaning against the stairwell, the neon-pink lights warping the shadows of their art, my friend asked if I wanted to meet their acquaintances. “This is secretly just a law party,” they said. “I know like, everyone here.” They introduced me to someone named Anna or Alanna, who asked what year I was. Oh, I said. I’m not in the program. “Oh! What line of work are you in?” I’m in tech. I didn’t have it in me to evade the direct question. “Where at?” Uhh. Google. When I said this, my friend burst out laughing. “That’s hilarious,” they blurted. “You’re joking, right?”
Everyone from my weekday life thought it was a bad idea to quit my job to study poetry in Virginia. It wasn’t enough that the graduate program was fully funded. “Why can’t you write on the side?” they asked. “What will you do after?” I hadn’t yet convinced myself that it wasn’t a bad decision. I still viewed poetry the way I was raised to—as a luxury. I arrived at creative writing via evening and weekend community workshops, which I paid for with my tech salary. But I didn’t know when else in my future I could have the financial privilege to stop living doubly. My life until this point revolved around working a job that would earn enough to support my family. My youngest sibling had just graduated from college and started working at a 9-5 in the city. I had recently made the decision to stop the expensive and ineffective treatments for my chronic illnesses. To me, going back to school for a creative writing program meant I could stop going to work as a she/her and presenting as a they/them on my own time. All the time would be mine.
In the graduate poetry seminar, I was the only Asian student and one of two students of color in a class of 11 students. I began censoring myself. It wasn’t a conscious choice. I remember thinking I’d ease into things, to start by submitting love poems to class, ones addressed to an non-racialized and ungendered “you.” It went better than I expected; none of the art school drama I had heard so much about. Not long after, I started writing poems in the third person. “The girl” this, “the boy” that, his mother, her father. I began experimenting with persona. I turned toward Jeanne d’Arc, the first white person I could think of who I could relate with. Raised in a firmly atheist household, what I knew of her was from grade school history lessons: Jeanne lived in medieval France, pretended to be a boy, led men into battle and won, heard voices from a young age, and was burned at the stake. I admired her fearlessness in the way she hid herself so publicly—her body, the language that occupied her mind. There were parts of her she wasn’t willing to disclose, and she made that known. In my university’s main library, under its crowned ceilings and freshly bleached walls, I borrowed several versions of Jeanne d’Arc’s trial transcripts and poured through them. I wrote poems from her perspective—Jeanne interrogated, Jeanne in prison, Jeanne before burning. It was like a video game, holding the controller and stepping into a pre-written role.
The Maid
“She was called putain, harlot, often enough by her enemies, but her judges knew that
committees of women had [repeatedly] examined her and found her an intact virgin.”
—W.P. Barrett and Coley Taylor, The Trial of Jeanne d’Arc
in the fat of boiled winter, you plunk me down, my scrapeshaken
knees splintering like splanks, smithereened shower, you strip the skirt
of berries hanging above my scalp, the black ovaries no longer
spheres but leafwrinkled and browndead. you tell me to pluck my face
forward, you can’t see behind the webbing, the tangle of spidersoft
muck and marled clothwipe. i keep myself clean, i insist, but you don’t
get it, your fingers sticky, wed with what you desire. i begin unthreading
the silk caught between my teeth, scritch my gums with my insected
nails, their shells glossaming green at their seams. try unstitching it,
i say, picking at my sclera. dripcheck and triple the whites of my iris,
they’re pure, if you still can’t see it, flip the lids covering my pupils, step
inside, incise me, i’ll pass any inspection, convince any of your
sore toes, just please don’t forget to leave your shoes at the door.
Jeanne Before Burning
Rouen, France. May 28th. Dusk.
A prison cell. Walls of stone and crumbling mortar. A hole slightly larger
than the size of a fist in one section of the wall.
JEANNE (19) kneels against the wall with her eye to the hole. Her wrists
are each chained to a fixture in the wall. She has been in confinement since
January. She is sick with fever. She wears traditional women’s clothing. Her
previously worn men’s clothing lies in a pile on the floor.
JEANNE
Every morning, the sparrows arrange themselves around the rainwater like a pile of dirty clothes waiting to be washed. I know this because I sleep facing the hole in the wall. If I stand with my back against the opposite end, I see one color at a time. The gash of yellow around the beak. The grey belly hair. The men with their words. The other day… I don’t remember what happened. I felt sick. I didn’t like the way they thought they could enter me with their questions. All these months, and still the questions, unchanged. It’s getting harder to hear what I want. That day… I felt my face burn, and then my whole body, and I agreed to something without knowing what. They said they would let me drink from the fountain, that they would spare me some bread. But first, they said. First this. We can’t let you infect the rest of the flock, they went on, snapping my body like a wing. When they finished cutting off my hair, I felt the sky
fall closer. I thought I could hear something. When I woke, I felt it. These clothes they put on me…
JEANNE holds her sleeve up to the hole in the wall.
JEANNE
Nothing shines through. The shapes don’t fit.
A life beyond stone. Someplace new…
JEANNE slowly changes out of the women’s skirt and into the men’s trousers.
JEANNE
The water, the bread. (beat) The sparrows…
FADE TO BLACK.
Remember to wash your hands,
sing me a prayer. Before you leave—
open your transcripts. Write down
my laughter.
One evening after class, as I flipped through my peers’ comments, I noticed there were a couple of individuals who didn’t give me feedback. I tried not to pay attention to the fact that they were both cis, straight, white men, and I messaged each of them asking if they had written me anything. Maybe I accidentally misplaced it? I don’t know why I was so bent on getting something from them, especially because my years spent doing group projects in school and at work had taught me to take on more of the labor and to do it quietly, but perhaps I felt that I had given up my entire life and community to be in this shared space with other writers, and I wanted something back for it.
After weeks of following up with them, both in person and online, I received a version of what I was looking for. At the top of one person’s print-out of my poem was a scrawled apology: “Sorry, I don’t know how to engage with this.” On a catch-up call with a friend later that weekend, recounting how my life was faring in Virginia, I wanted to laugh. What do you think he means by it? I asked. I’m inhabiting the voice of a religious white girl, I said. I’m not religious, I’m not white, and I’m not a girl. I’m not writing anything in
Chinese or anything remotely Asian. It’s not even gay, I went on. Should I start writing from the persona of a white boy?
When I began writing poems in the third-person and poems in persona, I told myself I was just adjusting, warming up to being around different people. I would write my more personal poems on the side, I thought, and bring them to class once I was ready.
Texting a Bot Trained on Jeanne d’Arc’s Trial Transcripts
help
I may or may not be able to assist you.
i don’t know what to do with my life
I do not know what you should do either.
how old were you when you knew
what you wanted?
13.
how’d you know?
In my childhood, there was an enormous
beech tree. And then I heard the bees.
bees! what’d they say?
I did not say they made sense.
i have all this random, useless knowledge.
i don’t know what to do with it.
What does anyone do with knowledge?
Apple cider, apple sauce.
how’d you know what to do?
Didn’t I die alone?
how do i stop people from asking me
what i’ll do next?
Sometimes it flowered.
Sometimes not.
I never brought them to class because I never wrote them. I filled my time writing Jeanne d’Arc monologues, outlining a one-person, one-act play that took place within the confines of a prison cell, the actor’s bodily movement restricted by the chains that held their wrists to the wall. I wrote a poem called “Texting a Bot Trained on Jeanne d’Arc’s Trial Transcripts.” I progressed from imagining myself as a white girl [1] to imagining myself as a machine [2] pretending to be her.
The difference between humans and machines is that machines can be completely present to themselves. Take, for example, a Turing machine, a model developed by Alan Turing in the 1930s whose operating concept can be extended to any modern digital computer.[3] At any given point, the Turing machine has complete knowledge about its current state, which it uses to determine what actions to perform next. What will you do after? A machine does not have a sense of selfhood in the same way we do, but I was drawn to the idea of a discrete predestiny, the kind of self-containment where a system possessed everything it needed to know for everything it needed to do.
When I worked in tech, part of my job was to improve search engine indexing, to scrape web pages and extract their data with greater frequency and precision. I then used this to train large language models (LLMs), which were used to determine what results would get surfaced for a person’s search. The more high quality data I could feed the models, the more its semantic and contextual “understanding” of human language improved.
After studying Jeanne for over a year, taking detailed notes on her 141-day trial and writing down almost every back-and-forth between her and her 60 interrogators, I decided to finally make an actual Jeanne bot instead of writing poems where I posed as a chatbot posing as her. I fine-tuned a light-weight version of GPT-2 with my Jeanne d’Arc persona poems, adapting the base GPT-2 language model to output text in the style of my persona poems. By using an older iteration of the model (as compared to GPT-4, the current one), I hoped to access a version of the model that hadn’t yet been taught to respond with a language sanitized and vetted for widespread corporate use. From its infinite nondeterministic states, [4] from its understanding and reproduction of my language in its output, I wanted to see what the model might reveal about me. What was it I hid beneath my layers of persona? What were the questions I wanted answers to?
>What are you afraid of?
There is no end. The only thing you can do, you say, is to break
the rules. I don’t know, but you know why.
Every day, when you hear me cry, I go back to the forest. When
I saw my mother, she was afraid. The children were crying. They
cried. As she told me, I could see my body. The walls, my head,
the back. My nose. There were no wires around my skin. As the
sun rose, I felt my limbs tremble, the pain of my skin, and the
smell of death. The clothes were dirty. But what if my body was
made of a woman? The only thing I could do was go back.
When I hear my words, the sky is blue. I will not see it again.
Never before did I see anything.
Scrolling through its output, I felt a thrill akin to engaging with the taboo. Revisiting this part of my old life carried with it the apprehension of meeting a former lover, someone who my friends had urged me to cut off and block. The people I knew who swore off explicit interactions with generative text models did so for valid reasons—the immense environmental costs of training even a single LLM; the large-scale, privatized collection of personal data; the way the language in the models’ output have been shown to reflect the systems of oppression constructed by the cis, white male-dominated voice of its training data; the 2023 Writers Guild of America strike; how under our current economic model, companies will unfailingly prioritize accumulating profit over valuing human lives; I could go on for hours.
Yet in reading through these responses, I experienced the kind of deep plea-
sure I had been missing in my own writing. The workshop-centered model
of graduate creative writing programs fosters a certain kind of literary aesthetic, and by extension, a certain kind of author, one who has an answer for why they used a tercet here, a comma there, the future perfect tense here but the past perfect tense there, a line of questioning which the program trained me to be guilty of. It’s not that these questions are wrong—I ask them to my students and I teach my students how to ask them to each other.
These minute constructions of language were what drew me to poetry in the first place. I loved toggling back and forth between using “a” or “the,” combing through etymologies, rearranging words and tenses until a phrase lit up with inevitability, as if it were not only aware of its purpose, but knew how to fulfill it. These questions become misdirected when they push someone into producing language under a paradigm they never agreed to adhere to.
>What aren’t you?
A woman, not a woman.
>When did you know?
In my boyhood I grew a bad impression of myself, but it wasn’t
long before I learned that a girl’s heart was missing. I’m not sure
if I was a woman, but if you knew, you couldn’t tell. My body
had moved, not only because of my lack of desire. I felt my hair
falling, it felt like I wanted to get up, her body. I was not a woman,
not an animal, or a human being. The body shivered, like it was a
pile of eggs, a dream, and a nightmare.
My hands were trembling. They held me in the water. And as
soon as I woke, I ran.
My first year in the program, I found out about a poetry reading group led by English PhD students. The group would be discussing a book I had coincidentally read the month prior. I cherished the book and returned it to the library with reluctance, making a mental note to add it to my purchase list. When I heard that anyone who joined the reading group would receive a free copy, I signed up immediately.
At the meeting, we chose a few poems from the book to close-read. “What do people think of this stanza here?” one woman asked. “I think these lines are lazy,” she said. Other women at the table agreed. “I don’t think this book delivers on its promise at all.” What promise? I wanted to ask. When has a book ever promised anything? Promises are an invention of marketing teams. I knew I shouldn’t have taken their comments so personally, that anyone was entitled to their opinions. But I kept thinking, Why don’t you write a better poem then? And we’ll come back here and sit around this table and talk about it. Let’s see how tight your lines are. I wished I could ignore the fact that the most negatively vocal people in the meeting were cis, white women and that the author was Asian. That none of the other Asian students present said a word. Stop making everything about race, I thought. Their comments were about the language.
Through the required ritual of writing a poem and bringing it to class, I began anticipating what my peers would say as I drafted. I impulsively cleaned up hypothetical inconsistencies and loose lines to prevent getting asked about them.
When I experimented with different parameters to generate text for my fine-tuned model, I would occasionally get output that looked like this:
>When did you want to live?
I don’t know. I don’t want my life to be like that. But I do want
me to know where I want it to go. Why?
because I feel like the best way to do that is to lie. Every time
you fall asleep, your brain tries to tell you what to think about
it. If you don. do not know what I am to say. It’s not going to
work. You don’t like me. I think you can’t explain why you‚re
crying. ————– ————— — ********************
******* ——– ——————— —————- ————— —
————– ————— — ————– ————— ———–
————– ————— —————- —********* —— –
————— —————- ————- ** ……………………….
…………… …………… …………………………………. . …………..
….. …….. … .. .. ……….. …. ______________________
__________________________________________
_____________ _____________ * _________ ____
____~~_____ , ~~~~
By nature of their deep learning architectures, pre-trained on unstructured data with millions of unknown weights and parameters, it’s impossible to explain the model’s nondeterministic output, why it blew up at the end like that. Why did you go from hyphens to asterisks to periods to underscores? someone in class might ask. I don’t know. No one knows. I loved the glitches because they were unintended. The model has no internal sense of a “right” or “wrong” word.
A friend recently messaged me, without context, “aq1sw2de3.” “What’s up?” I texted back. “Just cleaning my keyboard,” he said. I want more unprompted keyboard cleanings, I want more unintelligible lines without apology. What does it mean for a poem to be airtight? Who decides?
Leaf by leaf, loosen.
The smoke like a swallow’s sigh—
soft. What is the role of trees?
Forgiveness, giving
breath. The crackling air patters
like rain reliving morning.
Jeanne, in 15th century France, was born to peasant farmers and could not read or write. Most of what we know about her is filtered through the men who transcribed her trial. “But seeing the hardness of her heart and her manner of answering,” they wrote, “fearing that the torments of torture would be of little profit to her, [the judges] decided to postpone their application until we had received more complete advice on the question.” We don’t have the hardness of her heart. What we have is the hardness they projected onto it. We don’t have the words she chose to use against them. We have the words they chose to write down.
There are multiple instances over the months-long trial where the judges press her to explain her clothing. “Did God order you to wear a man’s dress?” “Do you think you’ve done well to take a man’s dress?” “Did you want to be a man when it was necessary for you to come to France?” Jeanne reiterated that her actions were the direct result of God’s commands, that she had already answered the judges earlier. On May 24th, 1431, Jeanne, under immediate threat of being burned at the stake, recanted. As part of her sentence, the court shaved her head and ordered her to wear a woman’s dress. This was a Thursday.
By Monday, Jeanne had changed back into men’s clothing. The notary wrote, “Asked why she had resumed it, and who had compelled her to wear it, she answered that she had taken it of her own will, under no compulsion, as she preferred man’s to woman’s dress.” Those lines sounded honest. She changed because she felt like it. She took back her abjuration. Two days later, she burned.
>When did you want to live?
Every day, I hear the sound of a baby. And when you hear some-
thing, you think. It’s like a dream. A dream, but it’s not that you
thought it was real. You thought you were dreaming. How long
does that take? When you wake up, it feels like you are dream-
ing, not a woman’s dream—but your dreams. Your dreams are a
shadow. The shadow is the shadow of the white body, the body.
I wanted your body, love, beauty. My body was an endless supply
of things. At one point, it was like a knife. What could it be like
to escape? I thought I might find myself alone, waiting until the
morning. But then something happened. I woke up and woke.
The sky was black. My breath was broken, but my heart was
clean. My body was covered with layers of air.
If I were alive in 15th century France or in the U.S. during the 17th century witch hunts, I am convinced that I would die at the stake. If I lived in certain countries, now, centuries later, I would be sentenced to death on the basis of who I am and who I have loved. I am not proud. I am deeply afraid—how is it that in the hundreds of millenia we have been here, on this earth that will
swell and erupt without warning, we still fight each other over who deserves the right to live?
In 2017, Lucas LaRochelle built Queering the Map, a community-generated online archive of queer memory in physical spaces. When I lived in New York, there were nights when I didn’t leave the office until past midnight. I would step off the G train and walk back to my 10’ x 9’ room and lie on my twin bed and scroll through my phone and yield to the full force of that isolation particular to cities. I would visit Queering the Map and zoom in on my neighborhood, clicking through the anonymous posts.
My sixth grade guidance counselour was the coolest, kindest
lady in the world, and she had a wife with a half-shaved head.
I was star struck.
told you how i felt before you left the city. no regrets no second
guessing, just the way your eyes looked in the august heat when
you hugged me.
You were my first queer friend, and the first person I told about it
first kiss that actually mattered
the very first love
you told me you loved me on the roof
I needed proof that the kinds of love I wanted could exist. That they were right there, next to me. The text was enough to be real. I didn’t need to know the person behind them, if there was one.
What most people remember about the Turing machine is the Turing test, but what most people forget about the Turing test is that it wasn’t originally formulated to distinguish humans from machines. In his 1950 paper, “Computing Machinery and Intelligence,” Turing initially described the test as an “imitation game” involving “three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. The object of the game is to determine which of the other two is the man and which is the woman.”
In the paper, Turing provides a single example question that the interrogator might ask: “Will X please tell me the length of his or her hair?”
What is the difference between being a man and convincing someone you’re a man? What is the difference between being a man and convincing someone you’re a machine? Does it matter?
>Who will you be?
In my boyhood, I could see myself in the shadow of a dark,
broken wall, a wall with broken pipes. I was afraid of the men
who had made me dream. They wanted me to look like a girl. I
wanted to look for a way to stay out of the way.
In my girlhood, I was able to sleep. I was afraid of the sun. I kept
my hands tied to my chest. The sky was not a sky, just a shadow. But
I’d seen it, my mind drifted, and the sound of my hair drifted away.
The sound of the water. I knew my dream would be a dream I
wanted to make. Every day, my body would feel like it belonged
to me. And my skin wouldn’t look as if it were a mirror.
The trees, their branches, all of the trees.
One of my favorite videos on the Internet is from 1988, of the singer-songwriter Björk demoing her television. Removing its backing, she exposed the cathode-ray tube and its circuitry. “This looks like a city,” she explained. “Like a little model of a city. The wires are like elevators [that] take care of all the electrons. I read that in a Danish book this morning.” She said she used to have televisiophobia because of an Icelandic poet. The poet once told her that when the brain tries to process the endless stream of pixels changing on the screen, it doesn’t have the bandwidth to evaluate the content. You just swallow and swallow. But it’s not like that, Björk said. She stopped being afraid once she read about how TVs worked. “I read the truth, the scientific truth, which is much better,” she said. “You shouldn’t let poets lie to you.”
In the 4th century BCE, Plato proposed an ideal city, which he divided into classes. Philosophers were the highest ranked. He separates philosophers from poets, claiming only philosophers have “knowledge of the absolute truth.” A poet, on the other hand, does not belong in this city because “he awakens and nourishes and strengthens the feelings and impairs the reason.” My excitement when I first came across this line was as if I had just jumped the turnstile at a subway station and gotten away with it. I wasn’t welcome in the platonic ideal of a city, the philosophers’ empire. Let them make me leave.
>But where will you go?
You’re afraid that I will die. How do you know if the answer will
be enough. The first thing you must do is ask yourself this: How
can you do something without knowing.
I have never seen myself as an angel. A secret of my childhood: I
can’t sleep, I promise. I’ve seen my dream—I’m afraid of the dark.
Iwalk out the door, open my eyes and see the light. There, a man
steps out of the window. As he walks, he looks up. The man looks
at me as if he has no choice but to leave. He told me he was afraid,
because the girl he saw was just a girl. I wondered what was the
purpose of this tale. And then, a day later, I heard his voice. What
had he been thinking? He could feel her eyes twitching like he
wanted to be seen.
It was like a dream. In this dream, you could see, not see. The
dream I dreamed of being, no matter what, was not dream. You
must be afraid. How could you make me fall asleep?
I can’t sleep, I promise. J.L. Austin’s theory of performative utterances claims that some sentences, in addition to describing something, changes the reality of that thing through the performance of speaking. When uttered out loud, the words “I promise” constitute the action itself. What does the machine promise? What can it?
What if I tell you that these machine-generated responses to my bolded questions were made-up? That I wrote them? Would you believe me?
16 years after the landmark paper that introduced the Turing machine, the English criminal courts charged Alan Turing for “gross indecency”—their label for Turing’s homosexual relationship with Arnold Murray. Turing pleaded guilty. Between imprisonment and chemical castration, he chose the hormone injections. Two years later, he committed suicide.
My feelings around my body, for as far as my memory can access, have not been so much about having the wrong body as they’ve been about having any physical form. I hated my corporeality. The ways it had been scrutinized, yelled at, stalked, and violated. The ways it will be. I hated how much of its past had been threaded with misdiagnosed pain. How much of its future would continue to be embedded within the incurable. When I turned 10, I prayed that one day technology would be good enough for me to exist
as an idea, to call my friends using a telepathic connection, to move at the speed of thought, weightlessly.
Interacting with this model trained on a distilled version of myself is the closest I’ve gotten to what it might be like to live as my bodiless dream.
>When will you stay?
I want to stay, for the first time, a promise. This is what I’ve al-
ways wanted. I wanted to be a mother, a father, and so on. My
mother taught me this.
When the dream begins, I must leave, because I don’t want my
life to end for you. There are things I can’t do to stay. My life
is a lie. When I am unable to conceive myself, I’m told to lie to
myself. This lie is what my body tells me. Instead, it will let me
know what it means to be alone. I’ve seen the difference between
being alone, alone and alone. I don’t want to leave, either, but I
am afraid for myself. I only have one choice. My house is empty.
But I didn’t leave. The door opened.
The difference between being a machine and convincing someone I’m a machine is that at the end of the day, I’m not one. I need to live the consequences of my physicality. The solution isn’t to pose as a modeled version of myself. I can dream of an existence encompassed by syntax and metaphor, but I can’t escape being questioned when I walk into a public restroom. At the same time I sit at my desk tinkering with my custom GPT-2, its parent company and its competitors are pushing their technologies into every aspect of people’s lives, regardless of consent. I turned to language, to poetry, because I wanted to be freed from the person I was raised to be. If I could
reinvent myself on the page, then maybe I would have the courage to tell you I am the way I am because I am. To say nothing and walk away.
I want to find a freedom through language, but not through a language provided by and built on imperialist forces, where truth is only accepted as truth when it comes wrapped with capital. Alone in my room, I will sing and you will not know what I am hearing. What my dreams are. So yes, I will hide. I will mismatch verb tenses and nouns and pronouns. Yes, I will lie. And no, I will not tell you. I will not tell you the length of my hair. Where I am going. No, it is long. Yes, it has gone. No, I have lost. Yes I am true, No. Yes. Yes, I will live.
Code/a
Because I had spent years working with LLMs—gathering and cleaning data, fine-tuning, evaluating output—when I conducted research for my poems where I posed as a Jeanne d’Arc bot, I felt like I was treating my mind as a neural net. My past life in engineering flipped my logical dependencies. Instead of neutral networks being modeled after human brains, I thought of my brain as a transformer LLM. Yet I resisted coding an actual Jeanne bot because I wanted a severance from that other life. I wanted to preserve my own voice. It was like working through the problems in a math textbook knowing the answers were at the back—I needed to know my mind was capable of decoding the differential equations before I flipped the pages. But even if there was no “right” solution, the questions were always there: what if the bot’s answers were better than mine? If it could write better than I could? Who would I be? What would I be left with?
The first year I lived in central Virginia, Jeanne was the figure I felt the most comfortable assuming—a religious white girl. I wrote under her persona until I left the state for several months, to go back to New York, where I could wear my own skin. When I revisited my poems during this time, I still had the desire to interact with a real Jeanne bot, but it was from a place of curiosity rather than competition. I knew I could make her, but I didn’t know what she’d say.
I fine-tuned two models, one on my Jeanne d’Arc persona poems and one on Jeanne’s trial responses. For the first one, I wondered what the model might pick up about my language and what it might reveal about the ways I wrote. For the latter, assuming that the bot’s answers would be closer to Jeanne’s than my personas of her, I wanted to know how a “real” Jeanne might think and react. Because fine-tuning on a pretrained model, in addition to having a much lower carbon footprint, is relatively easy to set up and experiment with, I wanted to see the fine-tuned output first before I considered a custom pre-train. I chose to use an open-source, distilled version of GPT-2, DistilGPT2, which was lightweight enough to run without needing premium compute units and extra GPUs.
In preparing the fine-tuning documents, since I was interested in language rather than form, I left-aligned my poems and removed any non-traditional spacing. For the second model, because I wanted the output to be in a voice closer to Jeanne’s, I needed to do additional processing. The notary who recorded the trial switched between directly quoting Jeanne’s responses and recording them in the third person. For example, he wrote,
“Asked if she knows she is in God’s grace, she answered: ‘If I am not, may God put me there; and if I am, may God so keep me. I should be the saddest creature in the world if I knew I were not in His grace.’ She added, if she were in a state of sin, she did not think that the voice would come to her; and she wished every one could hear the voice as well as she did.”
Removing the questions, maintaining the first person answers and converting the third person, these are the processed lines for the above excerpt:
If I am not, may God put me there; and if I am, may God so keep me. I
should be the saddest creature in the world if I knew I were not in His grace.
If I were in a state of sin, I do not think the voice would come to me. I
wish every one could hear the voice as well as I do.
After fine-tuning the model on her trial, as I poured through its output, I simultaneously felt disappointment and admiration.
>What do you know about sacrifice?
And who do I do not know? To answer them, I ask you to answer
this question. And as for your understanding, you will answer it.
I will tell you further. Then I tell your reply to the question, as
always, unless you are at war.
I could never make her, at least not with this model and this method. When the judges asked if she was in God’s grace, this was a trick question. According to Catholic doctrine, only God can know whether one is in a state of grace. If Jeanne had answered “yes,” she would be a heretic; if she had answered “no,” she would be a sinner. By answering in the subjunctive, she created an opening.
My model maintained her tone and circuitous manner of answering, but Jeanne was clever in a way my model had not learned. Her answers possessed a clarity, informed by her knowledge of the doctrine and of the voices she heard, that only she could have come up with, standing with her face to the death sentence. I left this model out of my essay. I wanted her to keep her words. But when I fine-tuned the model on my own poetry, I felt an unexpected rejuvenation. My mind lit up at the sentences I liked: “The body shivered, like it was a pile of eggs, a dream, and a nightmare.” “My body was an endless supply of things.” The text was grounded in images, unlike the other model’s output, which remained in the abstract. I was captivated by its illogical leaps, which I couldn’t help but try to puzzle meaning from, even though I knew there was none intended. I highlighted sentences, crossed them out, rearranged. There was no voice I needed to preserve—these words were sampled from a probabilistic model based on my own language.
It became a game. Any constraint I added was for pleasure. If I wanted to use a sentence from the output, I could not modify it. I was allowed to delete sentences, but deletions must also be in their entirety. All the model “outputs” in my essay are collaged except for the first response to “When did you want to live?” I did not “tighten” any of the language. I did not swap out words within phrases or change the punctuation. I only moved things around. Theoretically, if I needed more sentences to pull from, I could reroll the dice whenever I wanted. My interactions with the model were not improving OpenAI’s production GPT-n model or any of its offshoots. My fine-tuned version and its associated inputs and outputs were stored locally on my computer.
Despite the consecutive nights I spent rereading and re-transcribing Jeanne’s trial until daybreak, instead of making a machine representation of her, my resulting poems shed any resemblance to her life or to her voice. I worked so hard to pass as Jeanne. But in picking and choosing which sentences from the output to include, sentences which were generated based on my own work, I found that when I finally looked in the mirror, she wasn’t there. I’ve been here the whole time.
NOTES
[1] I use the word “girl” in the Christian sense the way people around her would have, attaching biological development to a labeled gender and to purity. Jeanne d’Arc never menstruated. She died at 19.
[2] I use the word “machine” as a metonym for abstract machines and the physical computers derived from them. In computer science, an abstract machine is a theoretical model for a computer that simulates a computer’s processing of inputs via a set procedure. An abstract machine is different from a mathematical function that maps an input into an output (e.g., f(x) = y) because it has an internal state (a “memory” of what events occurred previously) and a set of predefined rules (how to calculate something leading up to the end result).
[3] The Turing machine was the first formalization of an abstract model that could be applied to any computer. Their value was not understood until World War II, when Turing’s computers were used to decrypt German military codes.
[4] Non-determinism does not preclude LLMs from being considered abstract machines. They still move through state transitions (there are just several possibilities for each given state) to process a given input into output. Additionally, they must run on physical computers, which are Turing machines built on binary logic.
REFERENCES
- Austin, J. L. (1962). How to do things with words. Oxford University Press.
- Barrett, M. (1932). The Trial of Jeanne d’Arc (1st ed.). G. Routledge & Sons.
- Bass, M., Dueholm, M., Kalyani, A., & Ozkan, S. (2024, September 30). AI optimism and uncertainty: What can earnings calls tell us post-ChatGPT? Federal Reserve Bank of St. Louis. https://www.stlouisfed.org/on-the-economy/2024/sep/ai-optimism-uncertainty-what-can-earnings-calls-tell-us
- Bertram, L.-Y., & Montfort, N. (Eds.). (2024). Output: An anthology of computer-generated text, 1953–2023. MIT Press / Counterpath Press.
- Butters, J. (2024, September 13). More than 40% of S&P 500 companies cited “AI” on earnings calls for Q2. FactSet. https://insight.factset.com/more-than-40-of-sp-500-companies-cited-ai-on-earnings-calls-for-q2
- Cardillo, A. (n.d.). Chatbot statistics: Key trends & insights (2024). Exploding Topics. https://explodingtopics.com/blog/chatbot-statistics
- Copeland, J. (2012, June 19). Alan Turing: The codebreaker who saved ‘millions of lives’. BBC News. https://www.bbc.com/news/technology-18419691
- Hugging Face. (n.d.). distilgpt2. Hugging Face. https://huggingface.co/distilbert/distilgpt2
- Hugging Face. (n.d.). Training models with Transformers.
- Hugging Face. https://huggingface.co/docs/transformers/en/trainingigorbuenocorrea. (2007, February 13). Björk talking about her TV [Video]. YouTube. https://www.youtube.com/watch?v=75WFTHpOw8Y
- LaRochelle, L. (n.d.). Queering the Map. Queering the Map. https://www.queeringthemap.com
- Oxford University Press. (n.d.). Chatbot, n. In Oxford English Dictionary. https://www.oed.com/dictionary/chatbot_n?tab=meaning_and_use#1305059510
- Plato. (1968). The Republic (A. Bloom, Trans.). Basic Books. (Original work published c. 380 BCE).
- Ren, S., Tomlinson, B., & Torrance, A. W. (2024, November 1). Reconciling the contrasting narratives on the environmental impact of large language models.
- Scientific Reports, 14, Article 76682. https://doi.org/10.1038/s41598-024-76682-6
- Simkin, J. (2022, June). Alan Turing. Spartacus Educational. https://spartacus-educational.com/Alan_Turing.htm
- TIME. (1958, November 17). Medicine: The trouble with Joan. TIME. https://time.com/archive/6612987/medicine-the-trouble-with-joan/
- Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460.
HOLLY ZHOU is a poet and mixed-media artist from the California desert, the unceded territory of the Cahuilla and Mojave peoples. Holly’s poetry has appeared in The Kenyon Review and elsewhere. When not writing, they can be found exploring rocks by the ocean or in the mountains.
