AI writing tells: the paragraph level

This is part three of a five-part series on AI writing tells, descending from the surface to the spine. Part one cleaned the words. Part two fixed the sentences and their rhythm. Now assume both. Clean vocabulary, varied sentences, an ear at work. And the piece still has a machine quality to it, a sort of relentless tidiness, a sense that it is marching. This is the paragraph layer, and it is where prose stops sounding like thinking and starts sounding like an outline that got dressed up in sentences.

The shape of a machine paragraph

Almost every AI paragraph has the same architecture, and once you see it you will see it in nearly everything the machine produces. It opens with a topic sentence that announces what the paragraph is about. It follows with two or three sentences that develop or support that point. It closes with a sentence that restates the point, often with a small flourish, as if wrapping a gift. Then the next paragraph does exactly the same thing. And the next.

This is the five-paragraph essay you were taught in school, applied with machine consistency to every unit of prose. State, support, restate. State, support, restate. It is not wrong, exactly. It is a legitimate way to build a paragraph. The problem is the sameness, the fact that the machine does it every single time, so that a thousand-word piece is just this one shape repeated eight times, and the repetition produces a mechanical evenness that the reader feels as monotony even when every individual paragraph is competent.

Human writing does not do this. A human writer builds paragraphs of wildly different shapes depending on what the paragraph needs to do. Some paragraphs are a single sentence. Some are a long accumulation with no topic sentence at all, where the point only emerges at the end, or never gets stated because the details carry it. Some open with the conclusion and spend the rest of the paragraph undermining it. The variety is not decoration. It is thinking, made visible, because real thought does not move in identical units.

The topic sentence that announces

Start with the opener, because it is the most fixable part. The machine opens paragraphs with a sentence that tells you what the paragraph will say before it says it. “There are several reasons this approach works.” “The benefits of this method are significant.” “Understanding this requires looking at three factors.” These are not sentences. They are table of contents entries that wandered into the prose.

The announcing topic sentence is a tell because it reveals the outline underneath. The machine generated a structure first, a list of points to hit, and then wrote a paragraph for each point that begins by naming the point. The seams show. You can practically see the bullet list the prose was generated from, because every paragraph starts by reading its bullet aloud.

The fix is to cut the announcement and start with the actual content. “There are several reasons this approach works. First, it is cheaper” becomes, simply, “it is cheaper.” You do not need to announce that reasons are coming. Give the reason. The reader will understand it is a reason. When you cut every announcing opener across a piece, the prose stops feeling like a filled-in outline and starts feeling like a continuous thought, because the scaffolding that was showing through has been pulled out.

The closing restatement

The other end of the machine paragraph is just as predictable. It closes by restating what it just said. “Clearly, this approach offers real advantages.” “In this way, the method proves its worth.” “This is why the distinction matters.” The paragraph made its point, developed it, and then, not trusting you to have absorbed it, says it again on the way out.

This is the distrust of the reader showing up at the paragraph level, the same distrust that produced the inflated word and the hedge. The machine cannot end a paragraph on a developing detail and let the point be implied. It has to circle back and confirm. The result is that every paragraph ends on a flat note of summary, the energy draining out in the last sentence as the machine reassures itself that the point landed.

Cut the restatements. A paragraph should end on its strongest, most specific content, not on a summary of that content. When you developed a point well, the reader got it, and restating it insults them and kills the momentum. The hardest part of this fix is psychological, because cutting the closing sentence feels like leaving the paragraph unfinished, and that feeling is exactly the thing to push through. The paragraph that ends on a sharp detail and trusts you to draw the conclusion is the human one. The paragraph that draws the conclusion for you is the machine.

The deadly evenness

Now the hardest paragraph-level tell, the one that has no single fix because it is about the whole rather than the parts. AI paragraphs are all the same length. Read down the left margin of a machine-written piece and you will see it, a column of paragraphs each running four to six sentences, each occupying roughly the same vertical space, a regularity so consistent it looks typeset.

Human writing is ragged. A human writer will spend nine sentences on a paragraph that needs room and then drop a one-sentence paragraph that lands like a punch. The variation is not random. It tracks the importance and the energy of the content. The big idea gets a long, building paragraph. The turn, the reversal, the gut-punch, gets a single line alone, because the white space around a short paragraph is what gives it force. The machine never does this, because it has no sense that some thoughts deserve more room than others. To the machine every point is a point, equal in weight, deserving its standard four-to-six-sentence container.

This is the tell that survives everything else. You can fix every word, every sentence, every topic sentence and restatement, and if every paragraph is still the same length, the piece still reads as machine-made, because the even rhythm of the blocks themselves carries the signal. The fix requires you to think about weight. Which point here is the most important? Give it the longest paragraph. Which line is the turn, the moment everything pivots? Pull it out into its own one-sentence paragraph and let the white space hit. Where are you covering ground that just needs to pass? Compress it into something short. When the paragraph lengths start to vary according to what matters, the prose gains a quality the machine cannot fake: it starts to breathe, speeding up and slowing down, and the reader feels the writer’s hand deciding what deserves space.

A machine paragraph, taken apart

Let me show you the whole architecture at once, because seeing it dissected once makes it impossible to unsee. Here is a paragraph the machine would happily produce, clean at the word and sentence level, and rotten at the paragraph level:

“There are several reasons why consistency matters for new writers. First, writing every day builds the habit, and habits are what carry you through the days when motivation is gone. Second, regular practice improves your craft faster than occasional bursts of effort. Third, a steady output gives you more material to work with and learn from. In these ways, consistency proves to be one of the most important factors in a writer’s development.”

Read it and it sounds fine, which is the problem. Now watch it come apart. The first sentence announces the paragraph: “there are several reasons.” It is a bullet point read aloud. The body marches through “first, second, third,” exposing the outline underneath, three items because three is the machine’s number. The closing sentence, “in these ways, consistency proves to be one of the most important factors,” restates the opening, wrapping the gift, telling you what you just read. The whole thing is state, support, restate, with the support itself rigidly enumerated. It is an outline in a costume.

Here is the same content, rebuilt by a human hand: “Write every day. Not because daily writing is sacred, but because the habit is what carries you through the mornings when you have nothing and motivation is gone, and those mornings are most of them. The daily writer also simply gets more reps, and reps are how the craft moves, faster than any amount of inspired weekend bursts. You end up with more pages, more mistakes, more things to learn from. The occasional writer is always starting cold. The daily writer is always already warm.”

Notice what changed. The announcing opener is gone, replaced by a flat command that drops you straight into content. The “first, second, third” scaffolding is gone, the points now flowing into each other instead of being numbered off. The restatement is gone, the paragraph ending instead on a sharp contrast, “always starting cold” against “always already warm,” that makes the point without summarizing it. And the rhythm varies, a short opening command, a long winding middle sentence, a clipped close. Same information. Completely different object. One is an outline wearing prose. The other is a person thinking on the page.

The wall of narration

There is a failure that only shows up when you let the machine run unconstrained, and it is the most visible of all of them, visible from across the room before you read a word: the wall. The machine, left alone, builds enormous paragraphs, hundreds of words each, with no break and no dialogue and no white space, one continuous block of narration after another, because its default mode is to narrate continuously and it has no instinct for when to break.

A human writer breaks constantly. A paragraph ends because a beat ended, because the camera moved, because a line of dialogue interrupts, because the reader needs a breath, because a short paragraph after a long one lands like a punch. White space is a tool, and a page of varied paragraph lengths with dialogue cutting through it is doing half its work through shape alone, before a single word is read. The machine does none of this. It pours, and the page comes out as a slab, and the slab is exhausting to read in a way the reader feels immediately even if they cannot name it, because there is nowhere to rest and nothing changes pace and the eye has no handholds.

This compounds with the retreat into narration covered in the scene article. The machine avoids dialogue, which would have broken the paragraphs naturally, and so the absence of dialogue and the wall of narration are the same problem seen twice: nothing interrupts the flow, because the machine never wanted to interrupt it, because interrupting means staging a scene instead of narrating one. The fix is physical as much as verbal. Break the walls. Find the places a scene was buried inside a slab of narration and pull it out into dialogue and beats with white space around them. Vary the paragraph lengths violently, a one-line paragraph against a long one. The page should look uneven before it reads well, because evenness, here as everywhere with the machine, is the tell.

The transition that announces itself

One more, related to the announcing topic sentence but worth its own treatment. The machine connects paragraphs with explicit transition words, and it leans on a small set of them. “Furthermore.” “Moreover.” “Additionally.” “However.” “In addition.” Each new paragraph begins by formally signaling its logical relationship to the previous one, like a debater announcing “my second point.”

These transitions are a tell for the same reason the topic sentences are: they expose the outline. “Furthermore” means “here is the next bullet in my list.” “However” means “here is the counterpoint bullet.” The machine signposts the structure because the structure is all it has, a sequence of points it is walking through in order, and the signposts are how it keeps its place.

Human writers transition through content, not through labels. Instead of “Furthermore, the cost is high,” a human picks up a thread from the previous paragraph and pulls it forward, so the connection is carried by the ideas themselves rather than announced by a word. The fix is to cut the transition words and force the paragraphs to connect through their actual content. This is harder than it sounds, because sometimes when you cut the “however” you discover the two paragraphs were not actually connected by anything except the label, and now you have to do the real work of making them follow from each other. That discovery is the point. The transition word was hiding the fact that the logic was not really there.

The paragraph that has no topic sentence at all

Here is a move the machine almost never makes, and learning to make it is one of the fastest ways to break the machine shape. Write a paragraph that does not state its point anywhere. Not at the top, not at the bottom. The point emerges from the details, or hangs just past the edge of them, and the reader assembles it without ever being handed it.

This is how a lot of the best prose works, and the machine cannot do it, because to the machine a paragraph without a stated point is a paragraph with a hole in it. It needs the topic sentence the way it needs the restatement, as a guarantee that the meaning got delivered. So it always plants the point somewhere explicit, and the explicitness is part of what makes the prose feel machine-built. Everything is labeled. Nothing is left for the reader to find.

Try this as an exercise on your own machine-drafted prose. Take a paragraph and delete its topic sentence, the one that states the point. Then delete its closing restatement. What is left is just the content, the details and the development, and very often you will find that the point is still completely clear, carried by the material itself, and the paragraph is now stronger because the reader gets to arrive at the meaning instead of being marched to it. If the point is not clear once the labels are gone, that tells you something too: it tells you the content was thin, that the paragraph was leaning on its topic sentence to do work the details should have done. Either way, deleting the labels shows you the truth of the paragraph, and the machine never deletes the labels, so it never sees that truth.

The recap paragraph between sections

Watch what the machine does at the seams between sections, where one part of a piece ends and the next begins. It recaps. Before moving on, it writes a sentence or a short paragraph summarizing what the previous section established, then a sentence previewing what the next one will cover. “Now that we have looked at the causes, let us turn to the effects.” “Having established why this matters, we can examine how it works.” These transitional recaps are pure connective scaffolding, and the machine builds them at every seam because it learned that good structure guides the reader between parts.

The recap is a tell for the same reason the announcing topic sentence is. It exposes the outline by narrating the outline’s structure out loud. “Now that we have looked at the causes, let us turn to the effects” is the machine reading its own table of contents to you. A human writer trusts the section break itself, the white space and the new heading, to signal the transition, and simply starts the next section in motion. The recap is not just unnecessary, it is a small insult, because it assumes the reader cannot tell that a new section has begun or remember what the last one said thirty seconds ago.

Cut the recaps entirely. End each section on its own strongest point and start the next one in the middle of its own content. The break does the work. When you strip the connective recaps from a piece, it gets noticeably faster and less condescending, because you have removed the running commentary the machine kept inserting between the parts, the voice that kept stopping to explain the structure of the thing instead of just letting the thing happen.

Why this is the layer people miss

The paragraph layer is the one most writers never even think about, and that is exactly why it is such a reliable tell. People think about words, because there are lists. People can be taught to think about sentences, because the constructions are nameable. But the paragraph as a unit, the shape of it, the length of it, the way it opens and closes and connects to its neighbors, this is invisible to most writers, who experience paragraphs as just “where you hit enter.” So they fix the words and the sentences and leave the paragraph architecture completely untouched, and the machine signal sails right through.

An editor sees paragraphs. It is one of the things that separates an editor’s eye from a writer’s. The editor reads the shape of the blocks, the rhythm of long against short, the places where the structure shows through, and a machine-written piece is loud at this level even when it is quiet at every level above. So if you want to clean prose past the point where a careful reader catches it, you have to learn to see paragraphs the way an editor does, as objects with shape and weight, not just as containers for sentences.

There is a simple trick for learning to see them, and it works fast. Stop reading the words and look at the shape of the page. Squint, almost, so the text blurs and you are seeing blocks rather than sentences. Machine prose, seen this way, is a stack of nearly identical rectangles, evenly sized, evenly spaced, marching down the page. Human prose is ragged, a tall block, then a single line, then a medium block, then another single line, an irregular skyline rather than a row of identical buildings. You can spot machine writing from across a room with this trick, without reading a word, because the evenness of the blocks is visible at a glance once you know to look for it.

Use the same trick on your own drafts. Pull back, blur your focus, look at the silhouette. If it is a stack of even rectangles, the paragraph architecture is machine-shaped no matter how good the sentences are, and you have work to do: break some blocks, pull some lines out alone, let some paragraphs run long and others stop short. You are sculpting the silhouette, deliberately making it ragged, because raggedness is the visual signature of a human deciding moment by moment how much room each thought deserves. The even skyline is the machine giving every thought the same room because it cannot tell them apart.

And even when you have done that, when the words are clean and the sentences sing and the paragraphs breathe with real variation, there is a deeper problem waiting, one that only shows up when you stop explaining things and start trying to render actual lived experience. That is the scene and voice layer, and it is where the machine’s real limitation, its total absence of having ever been alive, finally becomes impossible to hide. That is part four.

Next in the series: Part four, the scene and voice level, where the machine flinches from experience. Or revisit part two or the series hub.