Issue 70: What AI means for writers
Ever since ChatGPT was announced at the end of last year, there has been a lot written about what AI means for writing. Is it the machine coming for us all, or is it merely a tool that we can leverage? As both a software engineer and a writer, I’ve been trying to learn about not just the technology and its possibilities, but also understand it in the context of our culture, labor systems, and values. Technology is never neutral. This essay aims to organize some of my own thoughts by synthesizing excellent ideas from a variety of journalists, fiction writers, and technologists about what ChatGPT is, its current impact, and the existential questions it raises about human creativity.
What is ChatGPT?
ChatGPT is the latest project released by OpenAI, the company also responsible for DALL-E, a program that generates art based on text prompts. It’s a large language model (LLM) that is trained off of a huge corpus of existing text, including websites (basically large parts of the internet) and books. GPT-3 is the third version of OpenAI’s Generative Pre-Trained models, released in 2020. GPT-4 is the latest version, released in March, which accepts text and image inputs to create text outputs. The LLM identifies statistical regularities in text, so correlations often determine how the chatbot responds to different types of subjects. Based on those correlations, it makes choices about what will come next in a sequence of words.
Ted Chiang has one of the best analogies for describing how ChatGPT works, calling it “a blurry JPEG of all the text on the Web.” He compares it to lossy compression, which is generally used for videos, audio, and photos because “absolute accuracy isn’t essential” the way it is for text files and computer programs, where one incorrect character can have disastrous consequences. ChatGPT only ever generates an approximation by repackaging information with different words. But we often accept the returned approximation because it’s in the form of grammatically correct text. “You're still looking at a blurry JPEG,” Chiang writes. “But the blurriness occurs in a way that doesn’t make the picture as a whole look less sharp.”
Ben Recht, a computer science professor at UC Berkeley, says something similar: “If you scale a language model to the Internet, you can regurgitate really interesting patterns” because “the Internet itself is just patterns.”
ChatGPT's impact on writers
Jay Caspian Kang compared GPT-3’s plotting skills to his own by feeding it the premise of his novel and asking what should happen next. He asked open-ended questions, provided different options for how the story should progress, added context around the genre, and prompted character sketches. The conclusion? “It seems, at least for now, that GPT-3 can generate its own stories, but can’t quite get beyond broad platitudes delivered in that same, officious voice.”
In my own experiments with ChatGPT, where I repeatedly asked it to generate a novel premise, first chapter, or short story about sisters (a topic of personal interest to me), I noticed a few patterns. The sisters were always wildly different, one usually living in a city and the other in a small town. The one in the city was always a “lawyer” or “attorney,” which made me wonder how the training data associates a woman in the city with this profession. Half the time, the sisters were brought together by the deaths of one or both parents (grim). They always had Western names, until I asked it to generate a chapter “in the style of Ling Ma,” at which point it named one of the sisters “Mei” (lol). Regardless of style or genre, it always concluded with their reconciliation. It all reads like “the design is very human”—the writing is very human.
ChatGPT’s output is generic and bland—and this is by design. GPT-3 is calibrated and edited (by humans!) to give more palatable answers, because it turns out that when you let AI learn what it knows from the internet, it will learn some troubling things. As Kang summarizes, “it still feels, for the most part, like you’re watching a very precocious child perform a series of parlor tricks.”
That being said, just because ChatGPT doesn’t match the literary prowess of humans doesn’t mean that it’s not already impacting writers. Terry Nguyen outlines a few ways ChatGPT is finding its way into the workplace. Some news organizations are incorporating it into their brainstorming and research processes. I work for a media company that has already drafted a set of AI editorial guidelines that present generative AI as a tool, not a replacement, for human creativity, voice, and judgment. The Writer’s Guild of America, which is currently on strike, is negotiating its contract to ban AI-written works and prevent AI from being used as source material.
AI-generated fiction has begun to flood literary magazines like Asimov’s Science Fiction and Clarkesworld. In response, Clarkesworld temporarily closed its submissions. Lincoln Michel astutely observes that “ChatGPT doesn’t actually have to get good to be bad news for writers.” The numbers of publishing—from submissions to queries to manuscripts—are already difficult odds, and it doesn’t take a whole lot to overwhelm the system and the people who steward it.
Michel predicts that one way to stem the flood will be to limit or shut down slush piles, focusing instead on soliciting authors or relying on existing networks. While it addresses the deluge, this will “only make art more elitist and harder to access for people without the privilege or connections.” Meanwhile, there might be more scams or opportunists who will try to capitalize on this publishing shift. “There might not be a lot of money in being a writer,” Michel writes. “But there is a lot of money in preying on the dreams of aspiring writers.”
The focus on whether AI is going to replace writers is one that buries a lot of nuance. Nguyen reminds us that “the conversation around AI must consider a milieu of competing interests: Corporate, labor, ethics, and lest we forget, the writer's ego.” Her two-part series in Dirt, “The AI writer” and “The AI reader”, have delved into some interesting ways to think about machine-generated works and creative collaboration. She connects the history of this kind of cocreation in the twentieth century to artists and programmers experimenting with ChatGPT today in ambitious and “unnervingly profound” ways.
I think it’s important to consider the full range encompassed by Nguyen and Michel’s examples and forecasts. It’s still early days, but the better we can understand the use and abuse cases of this technology, the better we will be able to shape our relationship with it and each other.
Considering existential questions
ChatGPT raises existential questions about how we define original or creative writing, why we value it, and how technology might shift our ways of thinking, creating, and critiquing art.
Chiang, a science fiction writer, narrows down the question of whether LLMs can help humans with the creation of original writing by defining “original” in this way:
Can the text generated by large language models be a useful starting point for writers to build off when writing something original, whether it’s fiction or nonfiction? Will letting a large language model handle the boilerplate allow writers to focus their attention on the really creative parts?
He argues that “starting with a blurry copy of unoriginal work isn’t a good way to create original work.” Writing unoriginal work is a rite of passage before you arrive at truly original ideas. Time and effort are the necessary experiences one needs to have to skillfully conceive and articulate thoughts. It’s very common to only arrive at original ideas through the act of writing. He also highlights the importance of drafting and revision. Our first drafts are “an original idea expressed poorly, and it is accompanied by your amorphous dissatisfaction.” The gap between our intentions and the result (not dissimilar to the Ira Glass quote about taste and creative work, which I’ve cited 23810 times in this newsletter) is fundamental to the creativity process, and it’s one of the things lacking when you begin with AI-generated text.
Kang prods at the “human” part of this conversation, namely, why are we interested in writing by humans? He questions whether humans can create things that are truly original. He sees himself functioning “like a more curated but less efficient version of GPT” because his voice is largely the synthesis of his own influences, trained on the corpus of books he’s read and admired. This assessment, however true, is not the whole picture. We’re interested in art because it’s a mirror. “The reason we read books and listen to songs and look at paintings is to see the self in another self,” Kang reflects. “Or even to just see what other people are capable of creating.”
Deb JJ Lee, an author and artist whose work was fed into an AI image generator without their knowledge or consent, similarly emphasizes the importance of how an artist’s voice is developed:
An artist’s voice is a lifetime development, an amalgam of their visual vocabulary, interpretations of influences, and their beliefs. It’s everything. To have work come out from a mindless machine that looks just like mine, but without the struggle of developing that voice, is a slap in the face.
Nguyen draws an interesting comparison between the invention of the camera and its threat to painting as a medium. The medium remained, but the technology has given rise to new types of art and artistic movements. ChatGPT might streamline creative writing for the average person, similar to how smartphones have made more people casual photographers. But regardless of the creative work, we care deeply about authorship. “I want to know which author I am reading,” Nguyen concludes her latest essay. “Even if they are dead—even if they were never alive at all.”
Our participation as artists in both existential and practical conversations around AI is essential to shaping whatever comes next. I want to keep an open mind to the possibilities, because it’s too simplistic to just villainize technology without attempting to understand it. But I’ve also been reflecting on why I care so deeply about art made by humans. At the heart of it, I find art created by humans interesting because I find humans interesting. It’s why I love reading author interviews and seeing authors’ writing spaces.
Creativity, after all, inherently resists optimization. Cartoonist and illustrator Rosemary Valero-O’Connell asks, “What is the point of removing the human element from human expression?” What joy or surprise is left if you automate the parts that are the most rejuvenating and life-affirming?
- Brandon Taylor’s writing advice on why it’s useful to write ‘the boring draft:’ “When I tell authors to try to bore me, to dedicate space and time to the listing of details about the circumstances of their characters, I am trying to get them to stop thinking in terms of psychology and abstraction. I am trying to get them to move away from their very excellent and refined taste, to forget the examples of the books they have read, and to focus instead on the surface of their story.”
- Check out Hana Lee’s “When do I earn out?” tool for trade pub authors—a way to calculate how many copies you must sell to earn out your advance. Lincoln Michel provides some additional context on the math and dispels some publishing myths.
- Sign up for Cleaver Magazine’s Summer Writing School. There are weekly workshops as well as one-day masterclasses, all available online. Sign up for individual masterclasses or get a 4-class bundle.
- Nancy Reddy on Jenny Odell’s books How to Do Nothing and Saving Time, and creating space for your creative spirit.
- Author and publishing expert Courtney Maum is starting a new line of writing retreats called Turning Points, “a mentor-led blend of craft and business instruction.” Applications are open for fiction and creative nonfiction for the New Mexico retreat in October.
Recent reads & other media
Last weekend was very rainy so I curled up with a cozy mystery called The Golden Spoon by Jessa Maxwell. It dares to imagine “What if Great British Bake Off but someone was murdered?” and I finished it in two sittings. It departs enough from GBBO (it’s set in America, the Mary Berry and Paul Hollywood-inspired judges hate each other) but casts the character archetypes of the baking show well and layers many twists and turns.
Otherwise, I’ve been trying a new thing where I try to read longer short stories (20+ min reads) before I sit down to write. Recent stories have included: “Background” by Elaine Hsieh Chou, “Peking Duck” by Ling Ma, “Silent All These Years” by Katie Devine, “Fable” by Charles Yu, and “My Friend Juniper” by Jemimah Wei.
Note: Book links are connected to my Bookshop affiliate page. If you purchase a book from there, you'll be supporting my work and local independent bookstores!
~ meme myself and i ~
Inflation continues to be extremely bad. Petey the desk dog. How people walk in museums. When someone asks me how I use dashes in my writing. This brought me physical pain. POV: your company sends another “Organizational Update” email.