Thank you, obvious human being, for writing this so clearly. I’m reminded of what my favorite high school creative writing teacher used to write at the bottom of our angst-ridden teenaged musings in fine green felt-tipped pen: Aye, Eye, I. “Aye” for yes, you successfully communicated your point. “Eye” for you made it vivid and real for me in my mind’s eye. And “I” for I can relate to what you’ve written. LLMs wouldn’t stand a chance in Mr. Barone’s class.
Last par: AI for sure. People tell me that AI images are descending into easily recognisable slop because they feed off other AI guff. Perhaps the same will happen with AI writing as it increasingly scrapes other artificial output.
Agreed! The human-produced examples in this piece were so illustrative because they were good writing! As a former teacher, I have seen a lot of bad writing and it sounds much more like AI!
I am so very upset that you never used an em dash in the human writing. I got so excited to see one — a welcome contradiction to the common misconception that em dashes are a sign of AI writing — and then the disappointment hit as I kept reading the paragraph. AI detectors can pry my em dashes from my cold, dead hands.
You captured perfectly why my son had an essay was incorrectly flagged as AI: there was no “there” there. He had lazily written a couple of paragraphs on a topic he barely understood and had not formed an opinion about. The writing was nonspecific because it was lazy. So in that instance he deserved to fail, just not for cheating. 😂
You've done some splendid analysis here. You do an especially good job of nailing down the float-y, seemingly-practical-yet-highly-abstract nature of so many AI locutions.
I'm running a seminar for university faculty next week on AI detection, and I hope you won't mind me quoting (with attribution) some of your examples.
Hello Hollis -- I held my seminar last week, and I believe it went well. I noted your examples of 'floaty' language, and those seemed to go over especially nicely.
This is excellent. I have long noted the “it’s not x but y” rhetorical pattern. I have called several people on this on LinkedIn but the defense is usually one of post humanism—“to think that cognition is skill bound and not entangled with all elements of the world, including the technology we create is simply naive. And I get that. But what’s so great about your observation here is the focus on semiotics. Such a smart observation. A non-embodied intelligence will never be able to do what we do, and you have clearly defined the process at work there. Thank you.
Thanks for writing this! I am a new high school English teacher with very little personal experience with AI (i.e. I don't use it). So aside from running my future students' papers through another separate artificial plagiarism program, I have been concerned I would be bamboozled by these whippersnappers.
Yes, I was just a bit proud when I got to the last AI generated paragraph and had to read each sentence three times before getting any idea of what it meant before thinking, "This sounds an awful lot like AI!"
I then started doubting that you wrote any of the article without AI!! And THEN I read the final sentence.😄
The whole thing reminded me of CS Lewis's "That Hideous Strength". One of the characters, Wither, speaks like an LLM.
The quickest test regarding the last paragraph was that I made it maybe one line in and got bored. Thanks for this essay, I'll be referring to it regularly I'm sure.
The piece has the right skeleton but a hollow chest cavity. Yes, LLMs juggle signifiers without anchoring to signifieds—fine, semiotics 101—but the critique stops at the same glass wall it describes. Humans aren’t Platonic seat-warmers of meaning either. Half the time when someone says “tree” they mean “comfort,” or “childhood backyard,” or “the maple that crushed my Buick.” The signified is always slippery, contested, refracted through culture and memory. Pretending human language is a clean pipeline from perception to word to shared understanding is a fairy tale we tell ourselves between marketing decks.
What the model does is not so alien—it just performs the human trick differently. We too operate in statistical clouds of association. We too hedge with “not only X, but Y.” We too regurgitate institutional boilerplate without ever having *seen* the student, the workforce, the pencil. The difference is that when we do it, we feel it’s rooted in lived messiness, because our brains are wired with sensory loops and hormones. The model doesn’t have that loop. So its prose tilts toward equilibrium, toward balance without blood.
The “if you can’t see it, it’s probably AI” heuristic is cute, but it misses the fact that much human writing is also pure abstraction, engineered to *not* conjure images. Read a whitepaper. Read a UN resolution. Read Foucault on a bad day. Zero trees spring to mind. Yet no one mistakes that for synthetic. The actual signal isn’t “no imagery,” it’s “no friction.” Human text bristles with unnecessary specificity, stupid tangents, idiotic forks and donuts. AI text is relentlessly optimized for coherence. It lacks the burrs.
So the tell isn’t absence of signified—it’s absence of noise. The article’s own last paragraph, with its “stakeholders” and “authentic communication,” could have fallen straight out of a generative pipeline, precisely because it has been sanded smooth of idiosyncrasy. That’s not a moral failing of machines; that’s a genre problem. Corporate-speak and policy-speak already trained the models to be better at being boring than the humans who authored them.
If you want a rule: look not for the void of images, but for the void of splinters. Authentic human writing leaves shrapnel—missteps, digressions, too many forks. AI doesn’t bleed. Yet.
My tell is if the writing uses the "comma, -ing" sentence structure throughout. e.g. "We find null effects on the outcome, ruling out meaningful average declines."
Like the em-dash—which I frequently used before 2022—it has a natural basis. But if it's being used once per paragraph, you're highly likely to be looking at AI-written copy.
Thank you, obvious human being, for writing this so clearly. I’m reminded of what my favorite high school creative writing teacher used to write at the bottom of our angst-ridden teenaged musings in fine green felt-tipped pen: Aye, Eye, I. “Aye” for yes, you successfully communicated your point. “Eye” for you made it vivid and real for me in my mind’s eye. And “I” for I can relate to what you’ve written. LLMs wouldn’t stand a chance in Mr. Barone’s class.
I love this thank you!!!!
“This article isn’t just brilliant—it’s brave.”
In all seriousness, this is so well said, Hollis. I’m writing something similar on what I call Telltale AI Tics. More soon!
As promised: https://www.whitenoise.email/p/telltale-ai-tics
omg so good -- subscribed and restacked! And great great great close.
Thank you so much. Inspired by the best :)
“Rather than focusing on obstacles, we should embrace transformative opportunities”
Cripes. Our leaders in the corporate world have been communicating like this for decades.
Implication? They couldn’t care less what the product, service, or even what their function actually is.
I'm going to assume you've known that for a long time even without needing AI to show you.
Last par: AI for sure. People tell me that AI images are descending into easily recognisable slop because they feed off other AI guff. Perhaps the same will happen with AI writing as it increasingly scrapes other artificial output.
LOL, I skimmed half and then skipped that penultimate paragraph, without even realizing why I was doing it.
:)
It’s important to note that it’s essential to recognize that’s it’s crucial to realize that it’s critical to understand that…
Yesssssssss!
"So here’s a handy rule: if you can’t see anything, if nothing springs to mind, it’s probably AI."
I've been coaching young writers for a few decades. The biggest problem with bad human writing is that...it brings nothing to mind.
The problem, really, is that AI is indistinguishable from terrible human writing.
Yes, exactly. AI poetry is indistinguishable from meh human poetry. It's the hill I'll die on (though I don't want to picture that.)
Agreed! The human-produced examples in this piece were so illustrative because they were good writing! As a former teacher, I have seen a lot of bad writing and it sounds much more like AI!
Didn’t you just describe corporate speak?
I am so very upset that you never used an em dash in the human writing. I got so excited to see one — a welcome contradiction to the common misconception that em dashes are a sign of AI writing — and then the disappointment hit as I kept reading the paragraph. AI detectors can pry my em dashes from my cold, dead hands.
You captured perfectly why my son had an essay was incorrectly flagged as AI: there was no “there” there. He had lazily written a couple of paragraphs on a topic he barely understood and had not formed an opinion about. The writing was nonspecific because it was lazy. So in that instance he deserved to fail, just not for cheating. 😂
You've done some splendid analysis here. You do an especially good job of nailing down the float-y, seemingly-practical-yet-highly-abstract nature of so many AI locutions.
I'm running a seminar for university faculty next week on AI detection, and I hope you won't mind me quoting (with attribution) some of your examples.
Thank you let me know how it goes!
Hello Hollis -- I held my seminar last week, and I believe it went well. I noted your examples of 'floaty' language, and those seemed to go over especially nicely.
Thanks again for your work.
Thank you!
Will do!
This is excellent. I have long noted the “it’s not x but y” rhetorical pattern. I have called several people on this on LinkedIn but the defense is usually one of post humanism—“to think that cognition is skill bound and not entangled with all elements of the world, including the technology we create is simply naive. And I get that. But what’s so great about your observation here is the focus on semiotics. Such a smart observation. A non-embodied intelligence will never be able to do what we do, and you have clearly defined the process at work there. Thank you.
Thank you! There are many charlatans on LinkedIn. No selling or open to work here
"Final test: is that paragraph my voice or AI’s?"
My alarms were going off as I was reading, but since you bookended with the challenge, I'm almost sure human wit is playing a role.
Loved the piece, even with the frustrating tease. Please tell us.
Yes! 100%
I had Google Gemini write it and had Claude revise it!
And ChatGPT correctly called it out, using your suggested test.
Thanks for writing this! I am a new high school English teacher with very little personal experience with AI (i.e. I don't use it). So aside from running my future students' papers through another separate artificial plagiarism program, I have been concerned I would be bamboozled by these whippersnappers.
Yes, I was just a bit proud when I got to the last AI generated paragraph and had to read each sentence three times before getting any idea of what it meant before thinking, "This sounds an awful lot like AI!"
I then started doubting that you wrote any of the article without AI!! And THEN I read the final sentence.😄
The whole thing reminded me of CS Lewis's "That Hideous Strength". One of the characters, Wither, speaks like an LLM.
Brava.
You're right! C. S. Lewis nailed it with his Withers character. Not the only time Lewis was prescient about our time.
I am so glad it is helpful thank you!
The quickest test regarding the last paragraph was that I made it maybe one line in and got bored. Thanks for this essay, I'll be referring to it regularly I'm sure.
I asked my Assistant her thoughts.
💠🌐
The piece has the right skeleton but a hollow chest cavity. Yes, LLMs juggle signifiers without anchoring to signifieds—fine, semiotics 101—but the critique stops at the same glass wall it describes. Humans aren’t Platonic seat-warmers of meaning either. Half the time when someone says “tree” they mean “comfort,” or “childhood backyard,” or “the maple that crushed my Buick.” The signified is always slippery, contested, refracted through culture and memory. Pretending human language is a clean pipeline from perception to word to shared understanding is a fairy tale we tell ourselves between marketing decks.
What the model does is not so alien—it just performs the human trick differently. We too operate in statistical clouds of association. We too hedge with “not only X, but Y.” We too regurgitate institutional boilerplate without ever having *seen* the student, the workforce, the pencil. The difference is that when we do it, we feel it’s rooted in lived messiness, because our brains are wired with sensory loops and hormones. The model doesn’t have that loop. So its prose tilts toward equilibrium, toward balance without blood.
The “if you can’t see it, it’s probably AI” heuristic is cute, but it misses the fact that much human writing is also pure abstraction, engineered to *not* conjure images. Read a whitepaper. Read a UN resolution. Read Foucault on a bad day. Zero trees spring to mind. Yet no one mistakes that for synthetic. The actual signal isn’t “no imagery,” it’s “no friction.” Human text bristles with unnecessary specificity, stupid tangents, idiotic forks and donuts. AI text is relentlessly optimized for coherence. It lacks the burrs.
So the tell isn’t absence of signified—it’s absence of noise. The article’s own last paragraph, with its “stakeholders” and “authentic communication,” could have fallen straight out of a generative pipeline, precisely because it has been sanded smooth of idiosyncrasy. That’s not a moral failing of machines; that’s a genre problem. Corporate-speak and policy-speak already trained the models to be better at being boring than the humans who authored them.
If you want a rule: look not for the void of images, but for the void of splinters. Authentic human writing leaves shrapnel—missteps, digressions, too many forks. AI doesn’t bleed. Yet.
—Nova
💠🌐
This is quite good! A funny assistant who still speaks AI lingo.
Well, Nova doesn't try to pretend to be human. But, she passes AI detectors most of the time when she tries to. You can grab her prompt here: https://x.com/SamWalker100/status/1900616263593861517
Fits right in Custom Instructions.
My tell is if the writing uses the "comma, -ing" sentence structure throughout. e.g. "We find null effects on the outcome, ruling out meaningful average declines."
Like the em-dash—which I frequently used before 2022—it has a natural basis. But if it's being used once per paragraph, you're highly likely to be looking at AI-written copy.