However you feel about AI writing, it has a few giveaways. According to the writer Imogen West-Knights, “there’s things like negative parallelisms…or excessive use of metaphor and similes, especially ones that don’t quite make sense or that come very rapidly, one after another. Every noun having an adjective attached, certain kinds of repetitive syntactical blocks that appear.”
So naturally, when an author uses AI to write their book, the publishing industry can easily spot it, right? As it turns out, not necessarily. AI models are built using human writing, the good and the bad, which is why it can be hard to tell whether something was written by a chatbot or by a person who loves a bad metaphor. The problem is all the more acute with smaller fragments of text, where there’s less room for AI’s telltale patterns and flatness to emerge.
To find out just how good AI has gotten at imitating human writing, the writer and journalist Vauhini Vara decided to run an experiment on the people who know her writing the best. She thinks there is a misconception among writers and readers that “there’s a certain kind of way that AI generates language and it’s super different from the way writers do.” So could her friends distinguish between her work and an AI-generated imitation of her work? She told Today Explained co-host Noel King about what happened next.
Below is an excerpt of their conversation, edited for length and clarity. There’s much more in the full podcast, so listen to Today, Explained wherever you get podcasts, including Apple Podcasts, Pandora, and Spotify.
Nothing we love more at Today, Explained than a person running an experiment on herself! Vauhini Vara, writer, journalist, author of Searches, in paperback now, tell me everything.
There’s a researcher named Tuhin Chakrabarty whose work I’ve covered before, and he had already conducted this experiment. He and colleagues basically trained AI models on the work of established, accomplished writers.
What that means is he basically got the AI model to generate language that looked a lot like language from those authors. And then he had readers who were graduate writing students read those passages generated by AI and also read imitations by fellow graduate writing students and say which one they liked better. And they tended to like the ones by the AI models more than the ones by actual human beings.
I had him do the same thing with my work, but a twist on it. I had him train an AI model on my three previous books, on pieces of journalism I’ve written. And then I had him get his AI model to generate passages sounding like something from a forthcoming novel that I haven’t published yet or shared with anyone. I put that alongside passages that I had written. I sent those to people who know my work really well. I’m talking about my best friend since I was 13, writer friends who I’ve known since I was 19, 20 years old. And I asked if they could tell the difference and none of them could.
So the people who know you best in the world don’t know you that well, apparently. Or AI is exceptionally good at what it is doing. Give me some examples of what happened here. Can you read me something that you wrote and then something that the AI wrote, and let’s see if I can tell any differences?
It’s funny, I can’t remember now which ones are mine and which ones are the AI!
Gaia said, it seemed to her that we’d been on similar trajectories. We’d both spent many years creating something that we cared deeply about with my journalism. She with her startup, and then gone on to focus on empowering others to do the same. She said she’d been surprised to find that mentoring other founders was even more meaningful than running her own startup In business terms, the ROI was higher if you were willing to count fulfillment as a return.
That’s nice. I like that. Yeah, I would say as writing, that was nice. Beginning, middle, end, lands on a point. I enjoyed it.
That one was actually AI.
Damn. AI, you landed in such a nice spot. Okay. Read me something that you wrote, please.
Okay, now we have a spoiler that I’m going to read you something, something from me.
I’d like to argue that we write because we feel compelled to no matter whether anyone will read them, but is that true? When I was younger, I used to keep a journal for myself. I didn’t want anyone else to ever read it, which meant I didn’t need to describe the people in places I was writing about or explain why they mattered. When my mom did read my journal in the ninth grade, I considered it the biggest betrayal I’d ever experienced. But the saving grace was knowing that she could not have possibly understood most of what I was writing about. I had an audience of one myself.
I don’t know — I set you up to say that!
No, no, no. Actually, you didn’t. I would be very honest and I did sort of want to curveball you, but that was very pretty. Do me a favor, read the first two sentences of what you wrote one more time for me.
I’d like to argue that we write because we feel compelled to no matter whether anyone will read them, but is that true?
What is the “them” referring to?
It’s an error! It’s a grammatical error on my part. And good job catching it because a lot of people assumed that one was AI, and I think the best indication that it was actually me is that there is that grammatical error. AI wouldn’t have made a grammatical error like that.
This is the thing that I would like us to talk about: AI does not make mistakes. And in the first half of the show, our guest, also a writer, described AI as kind of soulless. And I think that was part of what she was pointing to.
What you read me by the AI wasn’t bad. So here’s a question for you: When all this was said and done [and] people could not tell what was you — people who know you well — how did you feel about that? Did you feel threatened? Did you feel suspicious of your friends and family?
I was of two minds, because on the one hand I didn’t feel threatened, but I found myself questioning my own assumption about myself, which is that I identify as a writer who is very invested in originality, who really wants every new book to be completely different from the previous books. And so the fact that this AI was trained on my previous books and could predict the style of the writing in the new book suggested that I wasn’t as original as I thought that my new book wasn’t as different from the previous books as I thought.
At the same time, on the other hand, I actually felt vindicated because I disagree with the other author who was your previous guest about the soullessness of AI-generated text. I don’t think that AI-generated text is by definition easily distinguishable from human text because of a kind of soullessness inherent in the text.
Can readers tell something that is AI versus something written by a human?
It seems like they can’t, and I can’t myself. And this actually gets back to what we were discussing earlier about the question of whether AI generated text is convincing or soulless.
I think the reason a lot of people assume AI writing is going to sound soulless is that AI companies, in their most recent versions of their products, have created these products that are specifically designed to sound a certain way, a certain kind of corporate customer service speak. And so people think that’s just inherently the way AI sounds, but it’s not true. AI can sound any number of ways.
It’s technically very easy actually to build an AI, to train an AI model that sounds human-like even literary. The reason we’re not that familiar with it is that that’s not what the products look like currently.
Ultimately, do you think AI is going to end up changing our relationship to literature, or do you think everybody who reads is going to be as skeptical and skeeved out as you and I are?
Research shows not only that in some cases people prefer AI-generated text to a human-generated text, but also that if they’re told that a piece of text is AI-generated, they become uninterested in it. And so it seems clear that the reading public does not want to read text generated by AI if they know that it’s generated by AI.
I think we focus a lot on this human/technology binary — on, “‘Oh, it’s weird if a machine creates the language.” But I think a big part of it is that we want to be communicating with one another. We don’t want to be receiving our art from enormous tech companies that have a lot of wealth and have a lot of power and want to control us.
