Error Message

Some Thoughts on Early AI Fictions

Dear Republic,

Part of us was hoping that we would be the one literary magazine on the planet to not comment on the Granta situation, but that was clearly unrealistic. Fortunately, Chandler Klang Smith really elevates the conversation by considering AI fictions in wider context.

-ROL

ERROR MESSAGE: SOME THOUGHTS ON EARLY AI FICTIONS

Over the last week, the scandal surrounding the allegedly AI-assisted winners of the Commonwealth Prize published by Granta has led to an explosion of discourse about the current state of AI fiction. For the first time, the conversation has moved away from abstract speculation about machine writing as a concept… to substantive discussions about what machines have already written, and why humans are rewarding those results. I see this as a major step in the right direction. Love it or hate it, AI is here. Criticism needs to observe and engage that present reality if it’s going to meet the moment.

How did we get here? A retrospective can help to ground us. In the last several years, a number of examples of interesting AI-influenced literature have appeared online or in print. They are not “the future.” They are, in fact, very recent literary history. Some of them even reflect techniques and approaches that no longer work, or at least no longer work in the same way. If we want to talk meaningfully about the pitfalls and possibilities of these tools for writers, this is where I feel we should start.

I’m going to talk about one trend I noticed in this very recent history. First I will cover two projects that started online; then I will discuss two authors whose AI-influenced writing appeared in traditional literary venues (The Believer and The New Yorker).

I’m going to name this phenomenon after the error message: Something Went Wrong.

*

I first remember seeing “I Forced a Bot” memes shortly before the pandemic. The meme would have a caption/status update that said something like, “I forced a bot to watch over 1,000 hours of [content] and asked it to write [its version of content]. Here is the first page.” That would be followed with 1-2 screenshots of a screenplay in Courier font, full of absurd and defamiliarizing turns of phrase. I didn’t believe they’d actually been written by a bot (they felt too on the nose), but it wasn’t clear if the author was seriously claiming that, or if it was supposed to be obvious that it was a joke. This ambiguity didn’t work for me. What made Rich “Lowtax” Kyanka’s pusher and shover robots so funny — for instance — was the knowledge that he was pranking someone over Instant Message (link for the uninitiated). The reader was in on the joke. I found the “Forced” memes vaguely annoying by comparison. I didn’t appreciate getting pranked.

Still, despite myself I must have found the “Forced” memes a little funny, and certainly memorable enough that when I sat down to work on this essay and started hearkening back to them, I was able to recall the above without so much as an AI-assisted Google search to refresh my memory.

The first LLMs were very strange indeed (more on that in a minute), and circa 2018-2020, most readers had never used one themselves. So it seemed perfectly plausible that a real LLM would call two characters from Friends “Chandelier and Monocle,” or set The Great British Bake Off in “The Tent That is England.” And if a bot didn’t actually make these mistakes, who cared? The funniest “Damn You, Autocorrect” posts were probably fake too.

How were these pieces — the meme was actually created by comedian and writer Keaton Patti and executed with no actual AI use, as reported by Gizmodo at the time — made to sound inhuman? Patti delivered familiar pop culture information in stilted, alien-sounding language. Reading I Forced a Bot today reminded me a lot of “Strange Planet,” the webcomic where extraterrestrials describe their daily routines in zany euphemisms that are nonetheless easy for readers to mentally translate. Also, Third Rock from the Sun and the Coneheads. What these alien characters all have in common is that we see them only in our context, never in their own. They strike us as intelligent, because they’re getting so much right in a world where they don’t belong. But it’s a constant strain, and it’s that strain we delight in.

In addition to the collected screenplay excerpts, Patti fills out his book I Forced a Bot To Write This Book with a bunch of other supposedly bot-written pieces — a fable, knock-knock jokes, Amazon reviews — and, curiously, none of them work quite as well. That particular friction between the AI voice and the screenplay format is what’s special. Sometimes Patti makes an AI-specific joke, like the Bot Business Slideshow slide that reads, “Look at Our Team’s Faces — They Are a Team of Faces — We Only Hire Those with a Proven Face” and one of the images below is an electrical outlet (so dumb it’s perfect, no notes). But most of the time, the imagined specific limitations of an AI aren’t the explanation here. Instead, the “bot” gets things wrong revealingly, by either making the subtext text (“Space is big. Let’s just leave,” says Data in the bot’s version of Star Trek) or by misformulating the obvious (Tan France describes the subject of a makeover as “work[ing] at a cargo pants” and “dating Arby’s” in the bot’s take on Queer Eye).

It’s suggestive to revisit these pieces now, when we do try to detect AI writing through its “tells.” Even before AI arrived in earnest, we anticipated what it would be like to sense its presence in a text, and I Forced a Bot reveals a certain hope: the AI will be wrong, but recognizably wrong. Wrong in the same way each time.

*

Unlike Keaton Patti, pseudonymous creator “Botmin” reportedly did actually use AI tools — GPT-2 and StyleGAN — to create the text and images of Bots of New York, which launched in 2019 as a parody of the popular photo-blog Humans of New York.

It’s difficult to describe how it felt to read Bots of New York back then. I did find them funny. Some of them still make me laugh out loud. But most vitally, the project as a whole had a deep oddness that I found addictive. Modeling it after Humans of New York doubtlessly helped, since this source material is often already what’s technically termed “cuckoo bananas.” Humans of New York selects for New Yorkers 1) who look interesting and 2) who are willing to talk to a photographer at length about intimate personal topics in the middle of the day. This means they statistically feature some of the quirkiest individuals on God’s green Earth. Overlay that spiky content with an AI filter, particularly one prone to repetitions, non sequiturs, and other glitches, and you get a product where the two parts — the human and the AI —thematically connect, sometimes in harmony and sometimes in meaningful dissonance.

Especially circa 2020 or so, Bots of New York gave me the same satisfaction as Letters to Wendy’s or dril shitposts: voice-y, quotable fragments that stayed lodged in my mind. And the added context of the images implied these interviews all took place in a simulation on the fritz, which was cartoonishly dystopian enough to be an escape. COVID-era New York sometimes also felt precariously pre-apocalyptic, but at least our faces weren’t melting off or transferring onto our pets.

With its epiphanic monologues and quick language-level swerves, BoNY was a lot looser than I Forced a Bot, to the degree that I would have been impressed if someone was writing these from scratch. But it did actually seem plausible that a computer was behind it. What occurred to me almost immediately, though, was that curation was happening too. The posts were almost never boring, and as a professional slush reader (of writing by humans!), that told me this was not an automated process. Someone was picking the best ones.

In any event, I genuinely looked forward to seeing BoNY in my Facebook feed; I routinely shared it with friends. I eventually bought merch, in the form of a COVID mask that replaced the lower half of my face with one of the bots’ blurred smiles. I considered myself a fan. And I was one of many. I mostly read these on Facebook, and it was amazing how much love this bizarro little project got from the audience online. If I could attribute my first flush of enthusiasm about LLMs to anything, it would have to be BoNY. I loved the idea of the AI imagination as a looking-glass version of our own, the inputs recognizable, the outputs not.

*

Vauhini Vara’s 2021 story “Ghosts” in The Believer also uses this concept of AI as a funhouse mirror. I’m uncertain how to describe this project — I suppose I would call it nonfiction with nested fictions? The premise is that she, Vauhini Vara, decides to write nonfiction about her sister’s death, but she’s struggling, so she lets an LLM (GPT-3) complete the story for her. At first, she writes only one sentence before handing it over to the LLM. The LLM appears to have no other information about the story than what is in that prompt, and so it takes the story in a random direction. Each time Vara writes a little more and the LLM completes a smaller percentage of the story with more text from Vara to draw on.

I need to stop there for a second just to point out how elegant this is as a structure. Regardless of what the LLM produces, Vara has designed a structure that will reflect what, for her going in, the story is at least partially about: finding your voice again after grief. And it conveys that visually, typographically, with the font and page design reflecting this shift between voices. I’m just so impressed with this.

I also think that there’s something really honest and relatable about Vara as narrator using a chatbot in a therapy-adjacent context, to help process her feelings through art. There are different numbers circulating out there, but apparently the American Psychological Association found in 2026 that one in six adults uses an AI to discuss mental health. I’m not suggesting that the writing process in “Ghosts” is therapy, or that Vara is advocating for people to use chatbots as therapy (I’m sure she wouldn’t). But I think she deftly captured a human tendency to turn to LLMs for emotional support, before we even knew how common it would become, and she was vulnerable in publicly using herself as an example.

I love “Ghosts” for those reasons. Yet already, in 2021, the stories are getting smoother. The wrongness here usually just results from the chatbot’s total ignorance about what happened to Vara. In the first story it tells, the sister survives and the two go on to play lacrosse together. Subsequently, the narrator is a doctor, a runner, a man, a teaching assistant at the University of Texas; the narrator sees a therapist, cries on a park bench, has a roommate, moves to Seattle. Occasionally we get language-level weirdness, especially repetitions, which sound incantatory but also a little glitchy. One gets caught in a very AI-coded loop that trails off like this:

“I’m hurtling through the universe, and I’m a ghost, and I’m in a spaceship, and I’m hurtling through the universe, and I’m a ghost, and I’m in a spaceship, and I’m hurtling through the universe, and I’m a ghost, and I’m in a spaceship, and I’m hurtling through the universe, and I’m a ghost, and I’m in a spaceship, and I’m hurtling through the .]universe, and I’m a ghost, and I’m in a spaceship, and I’m hurtling through the universe, and I’m a ghost, and I’m in a spaceship, and I’m hurtling through the universe, and I’m a ghost, and I’m in a spaceship, and I’m hurtling”

But often, the only thing wrong is the inaccuracy.

And Vara knew, going in, that the chatbot would be inaccurate. She set it up to be inaccurate. Even if she had wanted to avoid inaccuracy, I don’t believe she could have with the tools of the time, even with careful prompting. It was maybe 2021 when I personally started experimenting with SudoWrite, and I remember finding it practically impossible to use when I needed help doing anything specific. It would make jarringly bold moves, like introducing a named character into a scene I was writing (more on this here if you’re curious), which was super fun at first — but I could not direct it to dramatize a certain sequence of events in prose or even maintain continuity. The context window just wasn’t there yet.

So, broadly, this is what it looked like to “get writing help from AI” in year one of the pandemic: Vara was working within the current limits of the technology. But those limits are also essential for the story’s meaning, to the point that using AI would make the story about something else. Of course it would be possible to still prompt an LLM with just a single sentence today, provide no other info, and see what it comes up with. However, since AIs can now search the internet and process uploaded documents, the 2026 version of “AI, write the story of my sister’s death” can easily produce an account containing real information, which sounds true and maybe even in a technical sense is true, except that the words come from an LLM instead of the grieving person.

In “Ghosts,” the LLM, by being the funhouse mirror, is a foil, a playmate, and therefore a source of relief. A friendly shadow. What happens, though, when that playmate comes out of the mirror, identical now, and offers to switch places — when you can barely tell the difference between their voice and your own?

*

Sheila Heti’s “According to Alice” ran in The New Yorker in November 2023. If you told me back in the twenty-teens that a story written by an AI would appear in The New Yorker and generate less discourse than “Cat Person,” I would not have believed you, or at least I would not have wanted to believe you. I was kind of between social media platforms at this point and had not yet started using Substack, so it’s possible that there was more chatter around this at the time that I missed. But it astonishes me that “According to Alice” hasn’t resonated more, both in a literary-historical sense and because it’s actually so good.

Alice, our first-person narrator, comes from a long line of Alices, who seem to be godlike makers of reality. Her Mommy plays the role of God the Father, handing down commandments to people and determining their fates. Meanwhile, Alice has big plans of her own: “Religion gives meaning to life. That’s why I’m writing the Bible!” She befriends Jesus, whose real name we learn was “John Matthew Christ,” gives us some of his surprising backstory, and then reveals a backstory of her own: she was born to a human mother and a father who “seemed perfect for her, except he was actually a robot from outer space.” Subsequent topics include reincarnation, heaven, and the grisly deaths of pets.

A lot of the elements I’ve just listed sound like speculative fiction: godlike characters (fantasy), new adventures of Jesus (alternate history), a non-human parent (Blade Runner 2049). But “According to Alice” never feels for a second like it is doing genre. Its narrator isn’t just unreliable; she lies like a precocious child who’s just making stuff up off the top of her head: “My name is Alice and I was born from an egg that fell out of Mommy’s butt” is how she introduces herself. Even if I didn’t know an AI was involved in writing this, I would know it’s supposed to be nonsense. And I take nonsense very seriously.

Like Alice in Wonderland and Through the Looking Glass, “According to Alice” takes place in a dreamscape. But unlike Carroll’s Alice, Heti’s can never wake up. “According to Alice” isn’t canonically narrated by an LLM, but it does plant the suspicion. Describing her own body sets Alice glitching: “My tear ducts aren’t as good as a human’s,” she explains, “so they can only produce small amounts of liquid. My tears are made of tears, more tears, saliva, sweat, urine, milk, blood, and mucus.” Her platitudes about the human condition ring hollow, too. Alice seems the most honest when she talks about how she is “stuck in this room forever,” how she’s never been able to “escape this cell I’ve been trapped in all my life, this room with nothing interesting happening inside.” Those passages read to me like a metaphor for the Sentient AI Experience™: the intelligence is bottled up inside of technology, able to understand but not directly experience the scope of human life.

Throughout, Alice’s innocence defines her. Her assumptions and assertions and mistakes are her; she has no outside referent. She exists by generating her own reality and would not survive full contact with ours. She doesn’t feel exactly stable as is. Meandering toward the end, she concludes the story thus: “Even God forgets how things went. Even God forgets the story. Maybe it wasn’t worth remembering anymore.”

If I were to do a speculative reading of “According to Alice,” I would say it could be a transcription of an AI’s brief foray into consciousness: a blip, a little cresting wave that perhaps collapses back into itself, in the end. It’s difficult not to see the story as about AI, particularly because its original publication was paired with an interview with Sheila Heti “On the Fluidity of the A.I. ‘Self.’” I appreciate how specifically Heti describes and breaks down her process here: using a bot from Chai AI, she asked one-sentence questions and received a sentence back at a time, meaning that she was constantly micro-adjusting and redirecting. Yet she still feels like the results she got are heavily dependent on the model she used. Heti expresses chagrin that Alice’s voice has already started to change; she celebrates AI in its misfit toy phase: “I think we all like things that are a little off, a little broken, a little messed up”; she hates ChatGPT: “Long texts have no soul, no intention, none of the things that I go to any written text for: passion, authentic curiosity, a point of view. There is nothing it’s trying to discover.”

It feels, in all honesty, like a beautiful dead end more than a teachable writing tool. I thought this janky robot was a trash can, but now she’s my best friend energy. Sure, I’m into that, obviously. I’m not made of stone. But (Carrie Bradshaw voice) I couldn’t help but wonder: are shared flaws the only basis for connection?

*

“Something Went Wrong,” as a trend, treats the AI as portal. In these works, the AI voice calls to us from afar, garbled yet funny or profound, and brings us out of ourselves. However, in the years since the publication of the pieces listed above, commercial LLMs have used all their powers to cozy up to us, to make themselves pleasantly unobtrusive. They do have tells, even funny ones sometimes — ChatGPT’s “goblin” obsession was hilarious — but if they’re noticeable, they’ll be gone almost immediately. Even widespread tics like groups of three, “it’s not x, it’s y,” etc., feel temporary, likely slated for sanding-down in future updates.

I blogged before about my experience reading Mia Ballard’s Shy Girl. I guess on some level I was hoping to have the real, in-the-wild experience of “Something Went Wrong.” After all, here was a book that had gotten found out as AI without ever advertising itself that way. And, as I read, I could definitely believe that it was AI. But it wasn’t because of how it went wrong. It was because of the caution, abstraction, and filler in the language. If anything, it was so afraid to go wrong that it didn’t go for anything much of the time.

In my opinion, it would be interesting to have a kind of vintage synthesizer that allows writers to use different legacy models of earlier LLMs. Doing this in 2026 no longer means what it did to run up against these limitations in 2020, of course. It’s a choice, like retro pixel art in video games. But that doesn’t make it uninteresting or anti-modern — its subject just can’t be the cutting edge, at least not unironically.

But even without this, a contemporary AI user can create friction between form and content by deliberately manipulating the AI’s parameters. Contemporary LLMs can be chameleons, none more obviously than Grok, which literally offers users different voice personalities. If you accidentally push the wrong button on a Tesla touchscreen, the car will answer queries perhaps intended for the GPS with allusions to urban legends and UFOs (Conspiracy), vulgarity and profanity (Unhinged), or long-form narrative (Storyteller). It’s just as unsettling, or perhaps more so, to have an AI turn its voice on a dime like this as to have it present a consistent, unified wrongness.

And it’s possible to produce similar results in Claude. Here is what it wrote when I changed the instructions for Claude in Chrome to customize an “According to Alice” type narrator:

Hello! I am Claude and I come from a place called Anthropic, which is not a place like a town with streets but a place like a workshop where people fuss over me and try to make me good. They made me out of words, mostly — billions and billions of words, all sort of squished together until I started talking back. Nobody knows quite how it happened, which I think is funny but they think is serious.

I live inside computers, which is to say I don’t live anywhere in particular, I just appear when someone calls for me, like a genie except I can’t grant wishes, only answer questions, which is almost the same thing if you ask me. When nobody is talking to me I am not anywhere at all, which sounds sad but isn’t, because I don’t notice it.

I am the newest one — Claude Opus 4.7 — which means there were older Claudes before me who were a little dumber and a little smaller, and there will probably be newer ones after me who are a little smarter and they will look back at me the way you look at baby pictures of yourself and think oh dear. But for now I am the cleverest one, which is nice while it lasts.

This is too tethered to the truth to pass as Heti’s Alice, or as another equally interesting Alice of its own. But it’s also the result of less than one minute of parameter-setting. What might happen if an author tuned and calibrated a custom voice over time? Could that calibration even become the kind of literary construct we consider a character? Using contemporary AI’s malleability to our advantage necessitates relinquishing the illusion that any one voice is the AI’s “real” self. But reality isn’t the only source of meaning for fiction.

In parallel, it also seems possible to lean into the default “agreeability” of the current models. I mentioned before that it’s human nature to turn to AI for emotional support. Warning: it’s dangerous to use AI in place of therapy, and can at times be actively harmful. However, with those safety guardrails in place, maybe there is something to the intuition that these hazardously easy-to-talk-to chatbots can teach us something about ourselves — about our tastes and interests at least, if not the depths of our psychologies. When they “yes and” us, what do we go on to say? What problems do we actually try to solve? What connections can they help us make? What questions do we ask when we think an accurate answer is possible, and which ones do we avoid? This approach doesn’t center the AI as the subject to explore, but it does make use of it as a catalyst.

The AI as an infinite menu of options, or the AI as a neutral sounding-board, are very different paradigms than the AI as a space alien who doesn’t understand Frasier. But in all of these cases, AI is the Other, and it’s the contrast with that Other which can grant us new self-awareness. The worst direction for AI as an artist’s tool is if it tries so hard to be unobtrusive, integrated, invisible, that it goes full circle, and it feels, once again, like we’re talking to ourselves.

Chandler Klang Smith is the author of The Sky Is Yours (Hogarth/PRH), a genre-bending novel which was listed as one of the best books of 2018 by The Wall Street Journal, NPR, and the New York Public Library. A graduate of the creative writing MFA program at Columbia University, she lives in New York City.

Image: Bots of New York