A paragraph is a thought and its development. That’s what I learned at school some thirty years ago. You introduce an idea, stay with it for a while, explain it, perhaps qualify it, and take the reader somewhere. When you’re done, you start a new paragraph.
Spend enough time reading LLM-generated writing and you start recognizing a different rhythm. A sentence makes an observation. New paragraph. Another sentence makes another observation. New paragraph. A third introduces a twist, followed by a rhetorical question in yet another paragraph. Your eyes fly down the page because something new is always waiting below the whitespace. But when you reach the end, what did you learn?
An observation gives you somewhere to go
LLMs are phenomenal idea generators. Give one a topic and it can produce observation after observation, something like this:
AI changes how we work.
That affects how we think.
Education will need to adapt.
Creativity will become more important.
Our definition of expertise will change.
Any one of these could be the premise of an essay. Put enough of them after each other and, superficially, you have one.
Good thinking also requires vertical movement. Why does AI change how we think? What does that actually mean, and what is the mechanism? What evidence do we have? Where does the claim stop being true? An observation gives you somewhere to go. Elaboration is going there, and the paragraph is where the writer stays with a thought long enough to make that journey.
The problem with one-sentence paragraphs comes from what they often leave undone. A single sentence can contain a profound thought, while a twelve-sentence paragraph can say almost nothing. Conciseness remains a virtue, especially when a hundred things compete for our time. Nobody needs a writer taking 2,000 words to say what could have been said in 500. Conciseness removes what a thought doesn’t need. But there’s a difference between being concise and going overboard with it. Atomization abandons the thought before it has been developed.
An outline dressed as an essay
Bullet points have an honest relationship with the reader:
- This happened.
- This matters.
- This is what we should do.
I know what I’m getting: a summary. If I want to understand why those things are true, I’ll need to look elsewhere. One-sentence paragraphs are less honest. Put enough of them after each other and you get the visual grammar of an essay with the intellectual structure of a bullet list. It is an outline written in prose.
That might explain a peculiar experience I’ve increasingly had with AI-generated writing. I can read 2,000 words effortlessly and reach the end feeling that I’ve consumed a lot, yet struggle to explain what I learned. The writing moved too quickly for either of us to stay anywhere. Thought. Observation. Twist. Another thought. Each piece is immediately rewarding, but together they leave little to hold on to.
It reminds me of eating cookies when I’m hungry. I can keep putting them in my mouth and might even end up full. But I’m not satiated because what I consumed wasn’t particularly nourishing. We’re starting to write the same way.
Maybe our attention spans haven’t changed
It’s tempting to blame atomized writing on shrinking attention spans, but I’m not sure that’s right. My kids will happily devour books. Give them the choice of what to watch, however, and they’ll often pick short entertaining YouTube videos over a movie. That seems contradictory only if attention is one generic resource.
I grew up with television and Blockbuster. If I wanted to watch a movie, I watched whatever happened to be on TV or went to the video store and picked something that was hopefully still available. Twenty minutes in, a boring scene didn’t present much of a dilemma. What else was I going to do? Today, abandoning something costs almost nothing. A swipe, click or tap reveals an effectively infinite collection of alternatives, many selected because an algorithm thinks I’ll prefer them.
Maybe our capacity for sustained attention hasn’t changed as much as our threshold for disengagement. A series of experiments published in the Journal of Experimental Psychology let people switch between online videos when they became bored. People expected switching to relieve their boredom. It did the opposite: more switching increased boredom while reducing satisfaction, attention and meaning.
That creates a nasty loop. Something doesn’t immediately hold my attention, so I switch. Switching makes the experience less satisfying, which makes me more inclined to look for something else. We spent centuries striving for more choice, only to discover that unlimited choice may make it harder to enjoy what we’ve already chosen. Creators have adapted accordingly.
Don’t let the reader leave
YouTube creators can see where viewers abandon their videos. MrBeast’s production guidance famously discusses re-engaging viewers at regular intervals. When their attention begins to wander, give them another reason to stay. The optimization makes perfect sense when the objective is entertainment and success is measured in viewers retained. After all, losing the viewer means failure.
Apply the same optimization to writing and you get short paragraphs, new observations, whitespace, surprising statements and another payoff before the current idea demands effort. The reader never has to stay anywhere for long. Before an idea becomes familiar enough to require thought, another arrives to recapture their attention. It might work brilliantly, but what exactly are we optimizing for?
When I write an essay, I’m trying to convey a thought. Getting you to the bottom of the page is only part of that job. I want you to understand what I’m saying and question it. Perhaps you’ll disagree or connect it with something I hadn’t considered. If I’m lucky, you’ll still be thinking about it after you’ve closed the page. None of those outcomes are easy to measure.
What we can measure, however, is whether you clicked, how long you stayed and how far you scrolled. We can see whether you reached the end or clicked something else afterward. Engagement therefore becomes a convenient proxy for value, and proxies have a habit of becoming targets. An engagement system asks whether you finished the article. A writer should care whether the article finished the thought, while a reader should care whether the thought continued after finishing the article.
Those definitions lead to very different writing. A 2,000-word article composed of tiny observations might perform extraordinarily well according to the first. You reached the bottom, so the metric says it worked. Three days later, you may barely remember what it said.
Reading gives you control
Reading is a peculiar medium. Video moves at the video’s pace, and audio moves at the speaker’s pace. You can change the playback speed, pause or rewind, but you’re still manipulating a fundamentally temporal medium. Writing waits for you. You can race through something familiar, stop for five minutes over one sentence, reread a paragraph or jump back to something you remember seeing earlier. Text lets the reader control the bitrate.
Developed paragraphs help you use that control. A paragraph tells you that its sentences belong together. If you’re familiar with the argument, you can recognize where it’s going and skip ahead without losing much. If the idea is new, you can slow down and let the writer take you through it. Ironically, atomizing prose in the name of scannability can make the argument harder to scan. Every sentence demands inspection because the structure no longer tells you which observations belong to the same thought. The eye moves faster while the mind gets less help navigating.
More importantly, reading involves more than transferring information from the page into storage. When I tell you that abundance lowers our threshold for disengagement, questions start forming in your head. Does it? Why would it? Does that match my experience? What about people watching three-hour interviews, or kids who won’t sit through a movie but happily read a 400-page book? Those questions are part of reading, which is why I’ve been experimenting with giving readers room to capture their thoughts.
The next sentences can think alongside you. They can anticipate objections, provide evidence and qualify the claim. They can lead both of us somewhere neither the original observation nor your immediate reaction could reach alone. When I introduce another interesting observation instead, I interrupt that process. Repeating the pattern makes atomized prose feel like a writer continually interrupting the reader’s thinking with another thought.
Some friction helps us think
We’ve spent years removing friction from digital experiences, and usually for good reason. Writing shouldn’t be difficult merely to prove that somebody worked hard on it. Bloated prose isn’t thoughtful prose, and a twelve-sentence paragraph that should have been three is still a bad paragraph. But treating all friction as waste ignores the work that understanding requires.
Sometimes you need to hold an unresolved idea in your head while following an argument. You remember a qualification, reconcile conflicting evidence and discover halfway through a paragraph that what seemed obvious in its first sentence isn’t obvious at all. That takes effort, and effort gives you an opportunity to think. There is a difference between making a thought unnecessarily difficult to understand and making it so effortless to consume that the reader never needs to understand it.
Research into digital reading offers some reason to take this seriously. Studies of scanning-oriented reading have found that readers can retain the gist while losing detail. Other research suggests that constantly switching between content can increase boredom rather than relieve it. None of this proves that a two-sentence paragraph makes us stupid. Meaningful segmentation can improve comprehension, and short paragraphs can be exactly the right tool. The concern begins when segmentation becomes atomization and breaking up writing to expose its structure starts breaking up the thinking itself.
Then we gave the job to machines
LLMs inherited a world already obsessed with clarity, conciseness, scannability and engagement. They are exceptionally good at taking those preferences to their logical conclusion. Left to its own devices, an LLM will happily produce observation after observation, each surrounded by whitespace and offering another reason to keep reading. And increasingly, we publish that writing rather than leaving it in chat windows.
People who might once have written a post, report or email themselves now give their thoughts to an LLM and ask it to produce finished prose. Sometimes they edit the output extensively. Sometimes they don’t. The default writing tendencies of these systems therefore matter. Machine-written sentences aren’t inherently inferior to human-written ones. What matters is whether somebody did the thinking those sentences claim to contain.
Imagine a researcher starting with a handful of findings and observations. Academic convention expects a paper, so an LLM expands those points into thousands of words of respectable academic prose. A reader encounters the paper and asks another LLM to summarize it. Findings in. Paper out. Paper in. Findings out. We’ve gone full circle, spending thousands of words and two rounds of inference to return to roughly where we started. AI already makes it easy to shift the cost of communication from the writer to the reader. Summarizing the expanded text completes the absurdity.
The important part used to happen in the middle. Elaborating the findings forced the researcher to explain why one followed from another and confront contradictory evidence. It exposed gaps, qualified claims and might have shown that the original conclusion wasn’t quite right. Writing represented the thinking, but it also produced it.
Stay with the thought
Conciseness matters more than ever. Wasting readers’ time is disrespectful when they have a hundred other things competing for their attention. Respecting their time means determining which thoughts matter, expressing them clearly and developing them enough that the reader can understand why they matter. A faster succession of unfinished thoughts doesn’t serve them.
We can question, imagine and disagree. We can connect an argument with our own experience and change our minds. Writing should give us room to do that. The paragraph has quietly supported this process all along by asking the writer to stay. Here’s your thought. Don’t rush to the next interesting observation. Explain this one. Why do you believe it? What follows from it, and where does it break? While the writer stays, the reader gets to stay too.
Our writing has become very good at keeping people moving until they finish the prose. Finishing the reading was never the point. The point was to finish the thinking.