Vauhini Vara and Karan Mahajan on When AI Tries to Sound Like Us

Award-winning writers and longtime friends Vauhini Vara and Karan Mahajan join co-hosts Whitney Terrell and V. V. Ganeshananthan to discuss Vara’s recent New Yorker essay “What If Readers Like AI-Generated Fiction?” Vara explains recent research by scientist Tuhin Chakrabarty, who has attempted to fine-tune large language models to produce better writing by feeding them authors’ entire oeuvres. She considers what it means that when Chakrabarty ran the results by some creative writing graduate students, they preferred AI imitations of writers like Junot Diaz, Sigrid Nunez, and Tony Tulathimutte to the writers themselves, or could not tell the difference. She and Mahajan talk about their decades-long connection and familiarity with each other’s writing. They muse on what it means that, when Vara talked Chakrabarty into letting her compete with a large language model, even Mahajan could not separate her original work from what it produced. Mahajan and Vara debate ways in which this technology will and won’t change how literature is written and received, the importance of style, reading as a collective experience, and if there is anything AI will never be able to capture about writing. Vara reads from the essay.

To hear the full episode, subscribe through iTunes, Google Play, Stitcher, Spotify, or your favorite podcast app (include the forward slashes when searching). You can also listen by streaming from the player below. Check out video versions of our interviews on the Fiction/Non/Fiction Instagram account, the Fiction/Non/Fiction YouTube Channel, and our show website: https://www.fnfpodcast.net/ This podcast is produced by V.V. Ganeshananthan and Whitney Terrell.

Vauhini Vara

“What If Readers Like A.I.-Generated Fiction?” | The New Yorker • Searches: Selfhood in the Digital Age • This Is Salvaged • The Immortal King Rao

Karan Mahajan

The Complex • The Association of Small Bombs • Family Planning

Others:

Pedro Paramo by Juan Rulfo • Beloved by Toni Morrison • “In the Penal Colony” by Franz Kafka • Ngugi wa Thiong’o

EXCERPT FROM A CONVERSATION WITH VAUHINI VARA AND KARAN MAHAJAN

V.V. Ganeshananthan: So Vauhini, you mentioned that you weren’t always able to find your own voice in these samples, and then Karan misrecognized you. What was your reaction like? What was the conversation that followed this?

Vauhini Vara: This is actually the first time we’re talking with our voices about it, because afterwards we emailed and texted about it. I think there were a lot of skull emojis and exclamation points involved. It was really funny, because my friend Dana, the tech entrepreneur, she and Karan know each other, and she specifically flagged Karan, like, “I’m sure Karan’s going to get way more of these correct, because not only has he known you forever, but he’s a writer and he teaches at Brown.” My point is that all of the humiliation here goes to Karan, not me.

On one level, I was horrified, obviously. On another level, I was fascinated. I found it intellectually interesting that a large language model could be fed all of my work and then given a description of my work and generate something that sounded passively like me. It had to do largely with the fact that I really pride myself on doing something new with every new project. I could imagine that if AI had tried to imitate something that sounded like my first novel, it would be able to because it had absorbed all of that writing. But it hurt my sense of my own originality compared to my past work, to see that it could actually imitate something in a book that I had not even published yet.

I think I chose passages that I thought would be tricky for people. So when I chose passages of my own to share with my readers, a couple of the passages that I chose were passages that I myself thought could potentially pass for AI. In some ways, the response made me think more deeply about my own work. As I continue to revise this novel, the fact that there’s this specter haunting me of people having already mistaken my work for AI, that’s going to make me think critically about my own work challenge myself to do more to make my own work to not “sound like AI” which in our common language or parlance really means “this sounds unoriginal or cliched.” In a way, it has less to do with AI than what we think of as good work, which is original work that says something in a new way.

VVG: What Karan said about specificity is also interesting. I had an experience relatively recently where someone wrote to me who teaches my work. They were like “I teach your work, and here I’m attaching papers by students who wrote about your work, and I’m so proud of these students.” One of the students’ papers had a citation from Brotherless Night where I read the sentence, and I was like, “That sounds like me. And also, I don’t know if I wrote that sentence, because I don’t remember that sentence.” I ran a search in my PDF of the book, and was like, “I did not write that sentence.” I didn’t write back to the professor for a while because I didn’t know what to do and he was so excited. Then he wrote to me again, and finally I caved and  wrote back to him. I was like, “I’m so sorry. I think that this is maybe a thing that happened in your class that you didn’t know about.” I think that’s an adjacent feeling.

I can also think of moments when I have caught plagiarism. I went to college and then subsequently worked in an office with Ross Douthat, who also used to work for Spark Notes. I had students who plagiarized from Spark Notes that Ross Douthat had written, and because I knew his work so well, I caught them. Today that probably wouldn’t happen. Voice is this thing—we’re so proud of our voices. And this fine-tuning is also about putting your entire body of work, the entirety of your voice, into this machine, and then the voice is the thing that it is claiming to imitate really well and maybe it’s doing it.

Whitney Terrell: I get that. On the other hand, I feel like my style changes all the time. And also, there are some writers for whom style is super important, and they spend a lot of time, maybe even preening over it. Somebody like Cormac McCarthy spends a lot of time thinking about how his sentences sound and what vocabulary he’s using. With other writers I don’t feel like style is that crucial to them. This story that they’re telling is more important. If I thought about— Karan, how have you thought about your books? Do you think of them as style-heavy or not? I feel like my first novel was style-heavy. I was trying to imitate Cormac McCarthy and his sentences and some of that comes into the book. And then with my next book, I realized, I can’t tell that story that way. It’s too overbearing. And it was much simpler. I don’t know how an AI model would merge those two different approaches to style, and I assume that my approach to style will change in the next book that I write.

Karan Mahajan: It’s a really good question about style, because I think it tends to matter more to authors sometimes than readers. And I’ll tell you why I’m saying that. I can imagine two books by Vauhini, one that is actually written by her, and one that is AI-generated. And the one by Vauhini is amazing in her voice, but maybe doesn’t have a good plot, let’s just say, for the sake of argument. And the other one is less voice-y, but actually delivers a series of events that are far more compelling, right? Maybe I go for that at the end of the day. Maybe the slight reduction in style wouldn’t bother me to the same extent. And you can always just say, “Oh, this is a work in which Vauhini has dialed down her style. Obviously, this happens with late style with most authors, a distillation occurs.” So I see this as being a thing we’re going to have to deal with a lot moving forward. I don’t see this going away on the upside.

I was thinking about how it would be if I was watching a Studio Ghibli film, for example, that had been partly AI-generated. I wouldn’t think of it as being that different from the production that happens on a pop song where there are certain things that can’t actually be done with instruments that are subbed in later on. Maybe the same would be true for a book, like the ideal melding would be if Vauhini wrote a book and she was like, there’s certain sections here that I don’t particularly care for yet, and I’m going to use AI to help me out. In the end, I wouldn’t know, and maybe I wouldn’t care. So I’m curious what Vauhini would think of that too.

VV: I think about that in the context of your work, Karan, which I think of as being very stylish, or work in which style is really important. At least in my reading of your work, this has always been the case for the past, like 20 years that I’ve been reading your work. But I think there’s a really strong relationship between style and consciousness and the broader politics in which the narrative consciousness operates, and the conversation that’s taking place between the narrative consciousness and an imagined addressee. All of which is to say that I think sometimes when we talk about style, we talk about it as if it’s this component of writing that can be siloed and divorced from all the other components, like plot or character or theme, when in fact, the decisions we make about the kind of language to use—which is what we’re talking about when we talk about style—are so intricately bound up with like narrative consciousness and community consciousness and the broader politics of any story. That feels relevant to point out too.

Transcribed by Otter.ai. Condensed and edited by Rebecca Kilroy. Photograph of Vauhini Vara by Brigid McAuliffe.