Something keeps happening to me when I read online. I’ve started calling it “the flinch”. It’s faster than I am, my body reacting before I can name what it’s reacting to.
This time it’s a tutorial I clicked on for something I actually want to learn. The piece is fine, the grammar’s clean, the argument is sensible, and yet something keeps pulling my eyes back to the top of the page to make sure a real person wrote this.
Three paragraphs in, I figure out what’s bothering me. Every sentence is shaped like it wants me to do or feel something. Not a specific thing. Just… something.
A more common term these days for this kind of prose is “slop”. Specifically “AI slop”, but I wanted to try and explore it a little more deeply than just labeling it and moving in.
This used to be easy to spot, but now, even for a millenial who has honed his skill in avoiding fake download buttons, automatically ignoring sponsored search results, and having lightning quick reflexes to close popup ads…that millenial is suddenly finding himself assaulted from all angles by it, in places he doesn’t even expect.
So what exactly does it look like?
A coworker who’d never said “delve” in their life would suddenly start delving into everything.
Or suddenly, everyone is speaking like they’ve discovered the fountain of youth or something. Tapestry. Unlock. Robust. Landscape. Shift.
People scattered em-dashes through their emails as though every dash were an FSA dollar and December 31st was tomorrow. (My take on the mark is that it isn’t really for anything, but at least when humans abuse it they’re enjoying themselves.)
Then people got smarter as they started to get called out. They told their models not to use “delve” or “landscape” or “robust.” They asked for em-dashes to be stripped out. A few even started telling the model to add deliberate typos. Except they still get caught all the time and can’t understand why, and it’s because they have a foundational gap in their own understanding of communication.
An LLM is an amplifier. Point it at a clear signal and you get that same signal, but louder, clearer, more well defined.
But point it at noise and you get louder more polished noise, noise that passes for signal without actually carrying much more value.
If you knew what you wanted to say, the model’s the fastest way you’ve ever had to say it, and if you didn’t, you just outsourced a sentence and got one back with extra paint on it. Either way the page fills itself, but the model has no idea what the page was for, because surprise, that was supposed to be your job.
These days, almost every sentence I read online is trying to convert me to do or feel or buy something. After twenty years of that, I flinch on contact. AI didn’t create the flinch, but it’s absolutely ensured that there’s nowhere left to read without triggering it.
the shape of an ask
A sentence that wants to convert me looks almost the same every time. Here’s one out in the open.
“Embarking on a Python journey can feel overwhelming, but with the right tools and mindset, you can transform your workflow and unlock your true potential as a developer. Let’s dive into the essential libraries every modern developer should know.”
Embarking. Transform. Unlock. True potential. Let’s dive in.
Notice what it does. Stakes blown up to “true potential” for a list of libraries. Reader flattered as “modern developer” before doing anything. Pulled forward by a hyperventilating “let’s dive in.” Bravo.
A normal, sane human writing this would just say what they meant:
“Here are five Python libraries I use almost every week.”
The next example is where this stops being just a tone problem and starts being something stranger (and in my opinion, much, much worse).
“I am so deeply sorry for your loss. Losing a loved one is one of the most difficult experiences anyone can face, and I want you to know that your feelings are completely valid. Please remember that you are not alone, and reaching out for support during this challenging time is one of the strongest things you can do.”
But this is a condolence message. Someone died. Grief is supposed to be the one place the ad-copy voice can’t reach. And yet people will send this exact paragraph without hesitation, certain they’ve said something meaningful.
“Most difficult experiences” inflates a death. Oh gee thanks, I thought this was going to be easy.
“Your feelings are completely valid” flatters before agreeing. Phew, I was about to apologize for them!
“Reaching out for support” is a CTA (call-to-action) aimed at someone in mourning. Great idea, let me file a ticket.
It resembles the same shape as the marketing example, because it’s the only shape the model knows.
What a real, genuine person actually sends is something like: “I heard about your dad. I’m sorry. Tell me if you want company or want to be left alone. Either’s fine.”
The model has no idea that that’s all somebody needs to hear. Could you make it produce something that short? With the right prompting and guardrails, absolutely. I spend half my time tuning LLM output and the other half throwing it out. But that assumes you think something’s wrong. The bar for “impressive” has dropped to “vaguely intelligent-sounding,” and this clears it without trying. Most people would simply stop evaluating there.
Ironically, the message which contains the fewest words actually holds the most meaning.
everything, everywhere, all at once
Persuasive prose isn’t new. Marketing’s been full of engagement bait since forever. For god’s sake, “copywriter” was and technically still is a real job. What’s different is volume, and the fact that nowhere is off-limits anymore.
A single ad in a magazine is easy to ignore. A model producing persuasion-shaped paragraphs faster than I can scroll, into every app I open? Yeah sure, let me just grow another set of eyes. Ad space used to be somewhere specific, but now I’m drowning in it the moment I unlock my phone.
Just ponder for a moment the context in which you come into contact with this stuff. This sort of persuasive cadence used to live in specific places, like billboards, spam folders, and the middle ten minutes of YouTube videos (funny how so many YT videos are coincidentally just a few seconds longer than the minimum duration needed to monetize, huh?).
Now language models mix it into code comments, condolence notes, tutorial intros, customer service replies, slack posts, bedtime stories, whatever. There’s no place left where I can let my guard down, no place left sacred.
Nowhere is this more prevalant than on Twitter and LinkedIn, where they pay you for engagement now, so the comment section is just somebody’s small business.
Somebody posts something that nobody’s arguing against, and like clockwork, ten replies show up calling the point so real, deeply valid, a real unlock, and yet not one godforsaken comment seems adds anything to the conversation.
Because they’re not supposed to add anything. They’re just trying to look like the reply that does, because that’s the reply that gets engagement, and engagement is the actual product. Their “personal take” on something is their brand. Same shape as the AI prose, same shape as an ad, leaning toward an audience that didn’t ask and has long since stopped expecting better.
It’s just literary junk food, except not even the good kind, more like the knockoff brand at a gas station that looks sort of right and tastes like nothing.
I don’t read the way I used to. I used to sink into prose, let it carry me. Now I’m scanning for the angle before I’ve finished the first paragraph. Is this trying to sell me something? Convert me? The suspicion fires before I’ve even decided if the content is worth my time.
Hell, even as I write and iterate on this essay that you’re reading, I’m sitting here thinking “wow, this kinda sounds like AI, doesn’t it?”. Am I trying to sell you on something? Am I trying to persuade you? I’m getting self-doubt just reading this sentence.
leaning inward
What about the stuff that doesn’t want anything from me? Where is that stuff?
Well… it’s still out there. Stack Overflow answers from someone who’s debugged this thing eight times and isn’t selling a course. Old technical books where the publisher just got out of the way. Reddit comments from people with unhealthy but extremely useful obsessions with finding the perfect washing machine. Opinionated writing from back when not everyone was building a personal brand.
So what actually separates the slop from the good stuff?
A persuasive piece leans outward, toward its audience, trying to get them to do or feel or believe something.
A sincere piece leans inward, into its subject, trying to understand it.
Leaning outward is selling. Leaning inward is thinking.
Once I know the direction, the prose is easy to find if I look for it, but it’s never going to land in my feed for me, because algorithms boost whatever leans outward.
The inward voice doesn’t try to climb the algorithm, it just sits there and waits for someone to need it. Which, incidentally, is why this site has no SEO, no trackers, no ads, no social buttons, no comments. That stuff is outward-facing, designed to grab your attention. This site just…is.
A blog that leans inward isn’t about you. If you’ve been paying attention to this essay, you know I mean that fondly.
what it costs
The flinch doesn’t discriminate, though. Some genuinely useful, well-meaning writing brushes against that persuasive cadence by accident, and I’ll flinch at it too. Honest enthusiasm gets caught in the crossfire.
That’s the price of twenty years of being sold to, mercilessly, in every direction.
So next time a paragraph makes my body recoil before my brain catches up, the question I want to ask is what the sentence wants from me.
That question’s older than any large language model, and it’ll outlive every one of them.