Sam Kriss on AI Writing

Wonderful essay at the NY Times. Excerpts:

I remember encountering a particularly telling example [of overfitting] shortly after ChatGPT launched. One of the tasks I gave the machine was to write a screenplay for a classic episode of “The Simpsons.” I wanted to see if it could be funny; it could not. (Still can’t.) So I specified: I wanted an extremely funny episode of “The Simpsons,” with lots of jokes. It did not deliver jokes. Instead, its screenplay consisted of the Simpsons tickling one another. First Homer tickles Bart, and Bart laughs, and then Bart tickles Lisa, and Lisa laughs, and then Lisa tickles Marge.

It’s not hard to work out what probably happened here. Somewhere in its web of associations, the machine had made a connection: Jokes are what make people laugh, tickling makes people laugh, therefore talking about tickling is the equivalent of telling a joke. That was an early model; they don’t do this anymore. But the same basic structure governs essentially everything they write.

On AI word choice:

A.I.s do not have the same vocabulary as humans. There are words they use a lot more than we do. If you ask any A.I. to write a science-fiction story for you, it has an uncanny habit of naming the protagonist Elara Voss. Male characters are, more often than not, called Kael. There are now hundreds of self-published books on Amazon featuring Elara Voss or Elena Voss; before 2023, there was not a single one. What most people have noticed, though, is “delve.”

A.I.s really do like the verb “delve.” This one is mathematically measurable: Researchers have looked at which words started appearing more frequently in abstracts on PubMed, a database of papers in the biomedical sciences, ever since we turned over a good chunk of all writing to the machines.  . . .  According to the data, post-ChatGPT papers lean more on words like “underscore,” “highlight” and “showcase” than pre-ChatGPT papers do. There have been multiple studies like this, and they’ve found that A.I.s like gesturing at complexity (“intricate” and “tapestry” have surged since 2022), as well as precision and speed: “swift,” “meticulous,” “adept.” But “delve” — in particular the conjugation “delves” — is an extreme case. In 2022, the word appeared in roughly one in every 10,000 abstracts collected in PubMed. By 2024, usage had shot up by 2,700 percent.

Amazing. I suppose I am the odd human here, because I like to write. I would never use an AI to write for me, because for me choosing my own words is the point.

But it certainly is fascinating that without anybody intending it ChatBots have all ended up sharing the same tics and word choices.