ChatGPT’s favorite words & punctuation

The Medium Newsletter
The Medium Blog
Published in
Sent as a

Newsletter

3 min readDec 11, 2024

👋 20 more days until 2025
Issue #226: Nikki Giovanni. Sora. Luigi.

Following up on last week’s issue re: two years of ChatGPT, I’ve been thinking about something. How do you know, on a word or sentence level, when someone’s used it? What are the tells? What separates “this is kinda formulaic writing” from “this is definitely AI”?

As you may remember from a previous newsletter, GPT loves to “delve.” Its preference for overly formal, vaguely British language is a well-documented result of being trained by workers in Nigeria, where “delve” is used in a business context far more than in the U.S. or anywhere else. A few more tells, via Jordan Gibbs who analyzed 1 million words from the machine and published his results on Medium:

  • Tirelessly
  • Cannot
  • Reimagined
  • Intertwine
  • Intricate
  • Tapestry
  • Expanse
  • Kaleidoscopic

A more recent analysis corroborates these findings, and adds a few more words. Pivotal. Vital. Comprehensive. ChatGPT also imitates certain syntax patterns common among smart-sounding writing, like em dashes and colons in titles.

There’s a deeper pattern I’ve noticed, though: ChatGPT can’t write well in first-person. By “well” I don’t just mean clearly. I mean the type of first-person writing where you feel like there’s a real human behind the words. (This is a great example of writing that felt 100% human to me.) One way humans do that — show other people we’re human, not automatons — is by telling jokes, but ChatGPT kind of fails at being funny. I mean, it’s fine at dad jokes (one-liners, puns, predictable humor) but it’s terrible at a type of humor that is uniquely human: subtext. Humor that says more than what words alone convey. An example: Sarah Cooper’s 10 tricks to appear smart in meetings, which communicates nuanced subtext about corporate conformity.

Journalist Will Lockett writes: “As ChatGPT doesn’t actually know what it is writing, it can’t have the self-aware, helicopter view of writing needed to create great subtext.” ChatGPT might be able to riff on a theme, but it doesn’t have the societal and self-awareness to come up with those themes in the first place.

Harris Sockel

🖋️ Also today: Nikki Giovanni. Sora. Luigi.

  • Nikki Giovanni — poet, teacher, activist, and Grammy nominee — passed away on Monday. She was 81. At the peak of her career, in the ’60s and ’70s, Giovanni openly admitted that most people don’t read poetry, so she recorded her 1971 debut album, Truth Is on Its Way, with a Gospel choir. “At the time,” writes Duke Professor Mark Anthony Neal, “I didn’t fully understand the genius of Giovanni’s vision — she was blatantly trying to bring the profane in conversation with the sacred, two decades before Kirk Franklin and later Kanye West would bring ghetto theodicy to the top of the pop charts.”
  • OpenAI launched a new tool, Sora, that allows you to input a few sentences and generate a video. We can all be filmmakers now, even without a camera. (Maybe we’ll feed a few of these newsletters to Sora so you can… watch them?)
  • Luigi Mangione — suspected UnitedHealth CEO assassin — emailed a friend earlier this year to rail against the lack of civil disobedience in Japan, which seems foreboding (to me).

✨ Your daily dose of practical wisdom: on creativity

A gem from one of my favorite books (which I really need to read again), Oliver Burkeman’s Four Thousand Weeks: “When you no longer demand perfection from your creative work, your relationships, or anything else, that’s when you’re free to plunge energetically into them.”

Deepen your understanding every day with the Medium Newsletter. Sign up here.

Edited and produced by Scott Lamb & Carly Rose Gillis

Questions, feedback, or story suggestions? Email us: tips@medium.com

Like what you see in this newsletter but not already a Medium member? Read without limits or ads, fund great writers, and join a community that believes in human storytelling.

--

--

Responses (59)