Everyone gets better. Everything gets the same.

In 2024, researchers Doshi and Hauser published something unsettling in Science Advances. They took a dataset of creative writing samples—stories, mostly—and had human raters evaluate them on multiple dimensions: originality, quality of prose, enjoyment, creativity. Half were written by humans alone. Half were written by humans using AI as a collaborative tool.

The results landed neatly in two categories, depending on which way you looked.

Look at individual pieces of writing, and the story is unambiguous: AI made them better. Stories written with AI assistance scored higher on conventional measures of quality and creativity. They were more engaging. More polished. They took greater narrative risks. Readers preferred them. The individual writer, armed with AI, had become more capable—more creative, even—than the individual writer working alone.

But then zoom out. Look at the corpus as a whole. Look at the collective distribution of all the stories together.

The AI-assisted stories, while individually superior, were more similar to each other. Their variance decreased. Where human-only stories sprawled across a wide range of narrative approaches, tones, structures, and stylistic choices—some brilliant, some clumsy, some genuinely strange—the AI-assisted stories clustered. They converged. The diversity of expression collapsed even as the quality of individual expression rose.

This is the paradox that neither a simple "AI is good" nor "AI is bad" framework quite captures: the same tool that elevates the individual writer suppresses the collective voice.

The Homogenization Effect

To understand why this happens, it helps to think about what's actually occurring when a writer uses these tools. An AI model is trained on the aggregate of human writing—millions of texts, normalized and processed into statistical patterns. When you query it, you're not asking for an individual voice or a unique approach. You're sampling from a probability distribution that represents "what usually comes next" across all human writing ever fed into the system.

This is immensely useful. It's like having access to every writing decision ever made, weighted by frequency. When a writer feels stuck, when they're not sure what comes next, the model suggests paths forward—usually the most statistically probable paths, the ones most consonant with how writing "usually goes." The writer feels liberated. They have velocity. They can move forward.

But there's a subtle cost embedded in that liberation. The most statistically probable paths are probable precisely because many writers have already taken them. The model, by definition, pulls toward the mean. It suggests what's typical, what's worked, what hundreds or thousands of similar writers have done before.

A human writer working alone, without this gravitational pull toward the center, might take the weird path. They might use a word no one uses anymore. They might structure a paragraph in a way that feels broken until it suddenly lands. They might write sentences that are too long, or rhythms that are off-kilter, because the constraints of their own mind produced something idiosyncratic. These idiosyncrasies—the fact that you can recognize a voice instantly, that a writer has unmistakable fingerprints—these emerge from the places where human writers deviate from the statistically probable.

AI doesn't create deviations. It reduces them. Or rather: it creates deviations, but deviations that are statistically palatable, deviations that have been validated by thousands of prior examples. The creative choices it suggests are creative within a narrow range of creativity that the training data has already explored.

The result: every writer using the same tools trained on the same data, guided by the same statistical probabilities, produces text that is individually better and collectively more uniform. The landscape of writing becomes smoother. The peaks get higher. The valleys fill in. Everything optimizes toward a well-defined center.

Style as a Function of Constraint

There's a concept in literary criticism called defamiliarization—the idea that art's primary job is to make the world strange again, to defamiliarize what has become too familiar through habit and routine. A style is interesting precisely because it deviates from the norm. Faulkner's syntax is difficult because it breaks how we expect sentences to move. Toni Morrison's prose is distinctive because it carves its own path. McCarthy's dialogue is strange because it doesn't behave like dialogue "should."

These styles didn't emerge from trying to be better writers in the conventional sense. They emerged from writers working within severe constraints—economic constraints, social constraints, the constraint of what was achievable given their particular tools and circumstances. Writers develop voice not when they have infinite freedom, but when they have to create within limitation. The constraint forces invention.

AI removes constraint. Or rather, it replaces the constraint of reality with a different kind of constraint: the constraint of statistical likelihood. And in doing so, it inadvertently replaces the strange with the familiar. It defamiliarizes the defamiliarization.

A writer's quirks—their pet vocabulary, their structural obsessions, the rhythmic patterns that come naturally to them—are often the most interesting thing about their work. These quirks feel like bugs when you're learning to write, when you're trying to be "good." You sand them down, erase them, try to write like the writers you admire. But over time, if you stick with it, those bugs become features. They become voice.

An AI model, averaging across millions of examples of writers doing exactly this—sanding themselves down toward excellence—learns the sanded-down version. It learns what "good writing" looks like after the author has already removed all the idiosyncrasies. It learns not the voice, but the ghost of the voice. Then it applies that ghost-voice uniformly to everyone who uses it.

The Monoculture Question

In agriculture, a monoculture is efficient and productive in the short term. One crop, optimized for yield, takes up all available space. Output per acre rises. Economics work. Then a disease arrives, or an insect, or a climate shift, and the monoculture collapses entirely. There was no diversity to fall back on. No hedging. No backup plan in the genetic variation that real agriculture maintains.

A literary monoculture would look like this: millions of writers, each individually producing what reads as good writing, each using similar tools, each guided by similar statistical probabilities. The mean quality rises. Reading becomes easier. Everything is competent, serviceable, well-constructed. And if you step back—if you read across hundreds or thousands of stories—you begin to notice that they sound the same. The narrative turns are familiar. The emotional beats hit when you expect them. The vocabulary is consistent. The sentence structures echo each other.

No single story is bad. But the corpus itself becomes sterile.

The strange part is that this could happen almost invisibly. Each reader, encountering individual stories, would find them well-written and engaging. The problem wouldn't announce itself. You wouldn't notice until you started comparing across the landscape—and most readers don't. They read one thing at a time. The monoculture would feel like just how things are.

This is the frog-in-slowly-heating-water scenario. There's no single moment where you can point and say, "There, that's when things changed." It's incremental. It's a shift in the distribution, not a collapse. And by the time you notice it—if you notice it—the genetic diversity of human expression has already been substantially narrowed.

The Precedent

Auto-tune arrived in music around 1997. Before it, a singer was bound by the limits of human physiology. You could hit the note or you couldn't. This created a particular kind of beauty in imperfection—the quiver of vibrato, the slight flatness that gave a voice character, the moments where a singer was reaching and almost failing. There was an entire dimension of music that existed in the gap between intention and execution.

Auto-tune removed that gap. Suddenly, every vocalist could be perfect. Pitches snapped into place. Performances became flawless. And in becoming flawless, they became interchangeable.

The quality of individual performances rose measurably. A singer using auto-tune could deliver a technically better vocal than they otherwise could. But listen to the radio, listen to what's released now versus what was released in 1996, and something is missing. Not quality. Texture. Character. The specific way that imperfection used to convey humanity.

Music didn't stop being good. It just stopped being surprising. The voice became a neutral delivery system for a melody rather than a unique instrument in itself.

This isn't a cautionary tale about technology corrupting an art form. It's an observation about what happens when you optimize a constraint away. The constraint wasn't the problem—the constraint was the source of character. Remove it, and you get better mechanics. You lose everything else.

The Shape of the Question

None of this is a problem that AI "caused" in the sense of creating something new and harmful. Optimization toward a mean is a property of any tool, any process, any system designed to make things better in measurable ways. You can see it in education when standardized testing becomes the metric for success—test scores rise, but education becomes narrower. You can see it in social media when engagement algorithms optimize for what keeps people clicking—individual engagement rises, but the range of what gets expressed shrinks.

The question, laid bare, is: do we value the mean or the margins?

Do we want writing to be better in aggregate—more polished, more engaging, more technically proficient? Or do we want writing to have more range, more strangeness, more of the stuff that only emerges when people struggle within constraints?

These aren't the same thing. They may be opposed.

A society could rationally choose the mean. Competent, accessible, well-written communication is genuinely useful. It's better than the alternative in many practical contexts. A manual written by AI is probably clearer than one written by a struggling human trying to figure out how to explain something. An email is better when it's more polished. A proposal is stronger when it's more professional.

But cumulative choice toward the mean, made by millions of individual writers using the same tools, adds up to something else: a narrowing of the possibility space itself. The distribution of human expression tightens. The weird gets rarer. The idiosyncratic is smoothed away not by force but by sheer efficiency. Why write a strange sentence when a normal sentence is available and better?

The Doshi and Hauser study didn't measure this directly. It just measured the variance in a dataset. But that variance is the measure of something real: how much room there is in the world for different ways of being, different ways of expressing, different ways of thinking that don't align with the statistical center.

What Gets Lost

When everyone gets better, we notice. There are new books, and they're well-made. There are more writers, and they're all more competent. The baseline rises. This is visible and good.

When everything gets the same, we notice something else: that we can't quite remember the last time we read something that surprised us. We can't remember a voice we couldn't forget. We can't remember writing that made us uncomfortable because it was shaped by a person, not smoothed by a process.

We'd notice it in aggregate, if we looked. We probably wouldn't notice it story by story, piece by piece, word by word. Each individual thing would still be good. The problem is the pattern, and patterns are hard to see when you're inside them.

The eerie part isn't that AI is ruining writing. It's that AI might be perfecting writing in a way that removes what made writing matter in the first place: the sense that a human being—with all their particular strangeness, their specific constraints, their individual quirks—had something to say that only they could say.