Although we are walking on barely trodden ground as we try to meaningfully incorporate GenAI into our works, there are other who have already forged ahead that we can look to. Among them is Vauhini Vara, a journalist, editor, and award winning writer whose The Immortal King Rao was a finalist for the Pulitzer Prize.
Especially interesting to us is her essay “Ghosts” where she writes about dealing with the death of her sister with GPT-3. Her experience writing ‘Ghosts’ and all the surrounding discussions she’s had on it since 2021 are covered in Confessions of a Viral AI Writer.
Confessions of a Viral AI Writer gives us a wide and intimate view of Vara’s interactions with GenAI. She charts out her experiences with GPT3, her discussions in interviews with Sam Altman and fellow writers that turn their nose as the suggestion of using AI.
Her initial approach with GPT-3 was simple. She would start out with a sentence.
“My sister was diagnosed with Ewing sarcoma when I was in my freshman year of high school and she was in her junior year.”
She would then correct it as the AI hallucinated and created its own fiction.
She kept doing this.
She would start iterating and after four or five attempts before GPT-3 “began describing grief in language that felt truer”. “Ghosts” was by far her best received work, even being anthologized in Best American Essays.
Vara’s struggle with writing was always ‘finding the right words’, especially so when it concerned her sister, and here was GPT-3 verbalizing for her, in a way that this writer felt she couldn’t. Not only that, but this award-winning writer even felt that GPT-3 had the best lines of the essay.
“I’d often heard the argument that AI could never write quite like a human precisely because it was a disembodied machine. And yet, here was as nuanced and profound a reference to embodiment as I’d ever read.”
-Vauhini Vara
With her guidance, GPT-3 was filling in the words, finding the sentiments, even the framing that she “could never find”.
What Vara's Experiences Show Us
Vara’s experiences with GPT-3 gives us a much more concrete example of what AI can do for a creative who wants to express themselves.
Thoughts on the literary potential of LLM’s aside, everyone can agree that GPT-3(2020) would not have churned out an essay worthy of being anthologized for any other reason than its novelty. It was Vara’s intentions, her prior taste as a reader, her experiences with her sister that she knew she wanted to capture, her experience as a publishing writer and editor which provided the structure and framing, all of which set the direction for “Ghosts”.
Vara’s candidness as she presents her thoughts, her process, and her appreciation for what GPT-3 captured gives an impression that she molded it to be an extension of herself.
In her case, the realization was that she “could use AI to do some of the most essential labor of a writer—to come up with the right words.”
An interesting point that she raises though is that ChatGPT’s responses have grown increasingly more cliché and sterile since the version that she used to write “Ghosts”, even with excessive prompting and interating. Her point is that Chat-GPT first and foremost is a chatbot that isn’t meant to be off-putting or challenging. In OpenAI’s very intentional effort to ChatGPT make more accessible and safe for cooperate use, as it becomes more friendly, and the targeted use cases for it become more particular, it may consequently become more uncreative. It seems that it simply may no longer be capable of the kind of language formation that made it so useful for Vara. This will be something else to consider going forward, not just for ChatGPT but for any other LLM.
Charting Your Start with AI in Writing
With Vara, we can outline and break down the most intuitive approach to using GenAI for creative writing:
Use it whenever you’re stuck, and be willing to work with it.
Go one sentence at a time. See how and where the LLM takes it. Be patient in directing it.
Fall into that flow, a process of creating, inputting, and then whittling where your vision dictates the direction that you and the AI you are working with go.
And most importantly, recognize that you can try to fit a hat onto an AI model, but as impressive as it is in its capabilities, it’s only going to be able to act how you want in so many roles. In that case, don’t write off GenAI as a whole. Consider another option, or by that point maybe consider asking a friend 😉