AI Boosts Your Creativity, But Makes Everyone Write the Same Story
Everyone is writing better. But when you look closer, why does every piece look exactly the same?
This isn't just a complaint from a writing community; it's the conclusion of an experiment published in Science Advances. Researchers Anil Doshi and Oliver Hauser recruited about 300 writers and randomly assigned them to three groups: one wrote short stories entirely on their own, one could get a single story idea from GPT-4, and the third could get five. After the writing was finished, 600 judges rated the stories independently.
The results were impressive, yet unsettling.
Writers who received AI ideas had their stories rated as more creative, more readable, and more interesting. Those who were originally weaker writers benefited the most, seeing a 10% to 11% increase in creativity scores and a staggering 22% to 26% boost in readability. It sounds like a happy ending for everyone.
But then, the researchers did something else: they compared all the AI-assisted stories against each other.
The similarity between them was significantly higher than that of the purely human-written stories. The individual improved, but the collective narrowed. Everyone’s work was upgraded, but they were all upgraded in the exact same direction.
A Social Dilemma of Creativity
Doshi and Hauser used a precise analogy: this is a social dilemma.
The classic example of a social dilemma is the tragedy of the commons. Every shepherd benefits from adding one more sheep to the public pasture. But if everyone adds one more, the pasture collapses. The logic of AI-assisted writing is identical: it is perfectly rational for every writer to use AI to improve their work; but when everyone does so, the diversity of the entire creative ecosystem begins to wither.
This is the exact opposite of the AI narrative we usually hear. The mainstream story is that AI lowers the barrier to entry, allowing more people to tell their stories, making the world more diverse as a result.
The data says otherwise.
The problem lies in how AI generates ideas. Large language models learn statistical patterns from massive amounts of text, and they output the most "likely" result rather than the most "unique" one. When a hundred people ask the same model for "a story idea about loss," the material they receive naturally clusters around a certain aesthetic center point. Perhaps it’s an unposted letter, a memory of a rainy day, or a childhood scene. Each idea looks good in isolation, but when viewed together, you realize they share a certain temperament, a narrative intuition, and a specific definition of what a "good story" is.
There is a technical term for this phenomenon called mode collapse. In the field of machine learning, it refers to a generative model gradually losing output diversity and starting to produce similar results repeatedly. A term originally used to describe the degradation of model training is now accurately predicting the trajectory of collective human creativity.
"Creative Scars": What Happens When AI is Removed?
If homogenization only happened while using AI, there would at least be a solution: just use it less.
But it’s not that simple.
A seven-day experiment conducted by Yiyong Zhou and others tracked the creative performance of 61 college students. For the first five days, the experimental group used ChatGPT to assist with creative tasks. On the seventh day, everyone completed the same task independently without AI. They were tested again thirty days later.
The results had two layers. The first was expected: after removing AI, the experimental group's creativity scores dropped back to baseline levels. Five days of AI assistance didn't "teach" them to be more creative. It’s like a crutch effect—your posture looks great while you're leaning on it, but once the crutch is gone, you're back to your original gait.
The second layer was more disturbing: the homogenization did not fade. Even when they were no longer using AI, the similarity between the content written by the experimental group continued to climb. The researchers called this a "creative scar." The AI was gone, but the aesthetic imprint it left behind remained. Those thought patterns shaped by AI seemed to have been internalized as the writers' own habits.
This is much more serious than "AI makes work mediocre." It suggests that even if you only use AI briefly, its impact on your creative intuition may be lasting. You won't notice this influence because what you write feels "good." The problem is that the writer at the next desk also thinks what they wrote is "good." And your versions of "good" are becoming increasingly identical.
What Exactly is Being Flattened?
Saying "AI makes stories similar" is one thing; figuring out "how they are similar" is another.
Doshi and Hauser's research pointed out that AI-assisted stories show a tendency toward convergence in structure, turning points, and even emotional arcs. This doesn't mean every story is identical (that would be too easy to spot), but rather that they share a more hidden commonality: a sense of predictable rhythm.
When you read one good AI-assisted story, you feel it is smooth, rhythmic, and has the right twists at the right times. Read ten, and you start to feel that the "right times" for every story seem to be in roughly the same places. By the twentieth, you can already guess where the twist will be. Not because the plots are the same, but because the underlying narrative logic follows the same implicit template of "what constitutes a good story."
Language models learn the statistical distribution of group preferences. They know which nodes usually move readers and what kind of endings receive high ratings. The suggestions they provide naturally lean toward these "safe zones."
The result is that AI is great at helping you avoid being "bad," but it is also helping you avoid being "weird."
And "weirdness" is precisely one of the most precious things in literature. Kafka’s The Metamorphosis begins by turning the protagonist into a bug; an editor receiving that manuscript probably wouldn't think it was a "good" opening. Otsuichi writes about tenderness with a tone so cold it borders on cruel—something that would be penalized under any standard of "making a story more engaging." These works are important precisely because they deviate from the statistically optimal path.
AI won't suggest you turn your protagonist into a bug.
(Well, if you ask repeatedly, it might. But it will add a hopeful internal monologue to the bug, because that is the statistically "better" way to handle it.)
Who is Affected Most? The Answer Might Surprise You
There is an easily overlooked detail in Doshi and Hauser's study: AI had almost no improvement effect on writers who were already highly creative.
In other words, for those who already know what they want to say and how to say it, AI ideas aren't very useful. The biggest beneficiaries are those who were originally struggling with creativity. AI pulled their performance up to a level close to that of the high-creativity writers.
On the surface, this is democratization. The gap in writing ability is being leveled.
But think about it carefully: is it the gap that's being leveled, or the difference?
High-creativity writers perform well because they have unique perspectives, unconventional intuitions, and the courage to make counter-intuitive narrative decisions. What low-creativity writers improve with AI is structural integrity, narrative flow, and the precision of twists. These are improvements at the level of "craft," which are valuable, but they are improvements moving in the same direction.
The result: the bottom is pulled up, the top isn't pushed higher, and the middle becomes more crowded. Everyone writes a "pretty good" story, but "different" stories don't increase.
It’s like a school raising all students' test scores to above 80. The parents are happy and the average looks great, but you can no longer find those eccentric geniuses who scored 40 (who might have scored 40 because they drew a full-page comic on the back of the exam).
Writing "Crooked" on Purpose in the AI Era
If you are already using AI to assist your writing, or are considering it, these research results shouldn't make you abandon AI entirely. That would be an overreaction. AI’s value in organizing drafts, checking logic, and polishing language is real.
But you need to be aware of a risk: every suggestion AI gives you is pushing you toward the middle. Every time you accept an AI edit without questioning it, your work moves one step closer to that statistical "sweet spot." One step won't matter. Twenty steps, and you are standing in the exact same spot as twenty other writers who accepted the same suggestions.
So, what can you do specifically?
Finish writing before asking AI. This is the most critical rule. If you ask AI for ideas during the brainstorming phase, your story’s DNA is contaminated from the zygote stage (sorry, the metaphor is a bit gross, but it’s accurate). Let your own intuition go the full distance first, and write that draft that might be terrible and messy, but is entirely yours. Then use AI as an editor to handle language and structural issues.
Reject the AI's first suggestion. Make it a habit. The first option AI gives you is almost certainly the most "mainstream" statistical option. Look at it, understand why it’s appealing, and then demand a different direction. Or better yet: extract the fragments you find interesting from the AI's suggestions and recombine them in your own way.
Write periodically in a completely AI-free environment. The warning from the "creative scar" study applies here. If your creative habits have been shaped by AI for a while, you need to consciously return to a purely manual state, even if what you write is much rougher. That roughness might hide the voice you were starting to forget.
Let AI find your flaws, but don't let it make your decisions. The design behind Slima’s AI Beta Reader follows a principle: AI is great at spotting problems (pacing issues, vague character motivations, lack of sensory detail), but the way to solve those problems should be decided by the writer. If you hand over both the diagnosis and the prescription to AI, you get a standardized treatment plan. Sometimes, your "flaw" is exactly what makes your style.
Returning to That Unsettling Number
There is a fact in Doshi and Hauser's study that is easily drowned out by optimistic narratives: AI did not significantly improve the quality of work for high-creativity writers.
What does this mean? It might mean that true creativity has a core that AI cannot reach. That core isn't in the fluency of language, the ingenuity of structure, or the unexpectedness of a twist. It’s somewhere deeper—in how a writer views the world, what they choose to notice, and what they decide is worth writing down.
AI can help you polish a B-grade story into a B+ or an A-. But it cannot upgrade a B-grade perspective into an A-grade perspective, because perspective isn't a language problem—it’s a life experience problem.
Back to the study at the beginning: everyone wrote better. All the stories were smoother, more structured, and more "professional." But after reading them, the 600 judges felt a subtle similarity among those AI-assisted stories.
That similarity wasn't at the level of vocabulary, plot, or even style. It was in a more hidden place. In the consensus of the stories on "what makes a good ending." In the default setting of "how a character should react" when facing conflict. In an imperceptible but very real standardized imagination of what a "good story" is.
To be honest, for a writer, the most brutal revelation of this research isn't that "AI will make you mediocre." It’s that "AI will make you good, but that 'good' is mass-produced." You need to decide for yourself: do you want to be "good," or do you want to be "yours"?
Sometimes the two overlap, and sometimes they don't. Knowing when they don't might be the most important judgment a writer needs to cultivate in the age of AI.