The Beige Drift: Why AI-Assisted Manuscripts Are Starting to Sound Alike

The flattening is measurable, it has a vocabulary, and it gets into a manuscript one accepted suggestion at a time.

I reread a chapter last month that I was proud of when I drafted it. The scene worked. The pacing worked. And a paragraph in the middle stopped me cold, because it was competent, polished, and could have been written by anyone. Not by a worse writer. By anyone. The sentences had that even, agreeable rhythm you start to recognize after a while, the way you recognize a stock photo. I had accepted three suggestions in a row that afternoon, and the page kept the receipt.

If you write with any kind of assistance in 2026 - autocomplete in your document editor, a chatbot you paste scenes into, a drafting tool - you have probably felt some version of that pause. Something in the prose is yours, and something isn't, and the something that isn't sounds suspiciously like everything else being published right now.

This series is about that problem. Not whether you should write with AI - plenty of working authors do, openly and well - but what the assistance does to the one asset a writer cannot replace: the way you, specifically, put sentences together. This first part is the diagnosis. The next part is what sameness actually costs a book. The last part is what protecting a voice looks like in practice.

The Flattening Is Measurable

Start with the study that should be pinned above every writer's desk. Researchers at University College London and the University of Exeter had hundreds of people write short stories, some unaided and some with story ideas from a large language model. Published in Science Advances, the result cuts both ways. Individually, the AI-assisted stories scored as more creative and better written - especially for the writers who scored lowest on their own. Collectively, those same stories were significantly more similar to one another than the unaided ones. Each writer felt lifted. The group converged.

The researchers called it a social dilemma: what helps any single writer in the moment quietly drains the pool everyone draws from. A reader browsing a shelf of individually improved, collectively interchangeable books does not experience the improvement. They experience the interchangeability.

A second experiment, presented at ICLR 2024, found the same pattern from a different angle. When people co-wrote essays with an instruction-tuned model - the polite, helpful kind that powers most writing assistants - the diversity of what different authors produced dropped measurably. Lexical diversity, down. Ideas, converging. And here is the detail that matters for working writers: the drop came mainly from the model's contributions, not from the humans changing how they wrote. The assistant did not make the writers similar. It added similar material to all of them, and the writers kept it.

The assistant did not make the writers similar. It added similar material to all of them, and the writers kept it.

Even the vocabulary of the flattening has been counted. A team analyzing more than fifteen million biomedical abstracts found that after late 2022, certain words spiked far beyond their historical baseline - delve, intricate, pivotal, meticulously, showcasing - and estimated that by 2024 at least 13.5 percent of abstracts - roughly one in seven - showed the fingerprints of a language model. Scientists are not novelists, but the mechanism travels: when millions of people accept suggestions from the same few models, the models' favorite words become everyone's words. A de facto house style, except the house is the entire internet.

What Sameness Sounds Like on the Page

Numbers are one thing. What does the flattening actually sound like in a manuscript?

It sounds even. The model's job, at every step, is to produce the most plausible next stretch of text - and plausible, statistically, means average. Not average as an insult; average as arithmetic. The suggestion engine is a gravity well centered on the middle of everything it has read. Left alone, it pulls toward sentences of moderate length, emotions named rather than staged, metaphors that have already survived a million uses, and a tone of mild, competent warmth. Every rough edge that would have marked the page as yours is precisely what the machinery is built to sand off.

None of the right-hand column is bad writing. That is what makes it dangerous. A clumsy sentence announces itself; you fix it. An average sentence reads fine, so it stays. Sameness does not enter a manuscript through your failures. It enters through your accepted suggestions - each one fine, each one a little less you, compounding at the rate of a dozen small surrenders per session.

Why Writers Feel It Before They Can Name It

Voice, as editors keep saying, is the one thing that cannot be added in revision. Craft teachers have been blunter still: agents can fix structure and pacing, but if the voice is not on the page, there is nothing to represent. Jane Friedman's writing community has spent the last two years circling exactly this point - that in an age of assistance, knowing what your own voice does is no longer optional equipment; it is the survival trait.

Here is the uncomfortable part. The writers most at risk are not the lazy ones. They are the diligent ones - the writers who show up daily, who use every tool available, who accept the improved sentence because it is, by any local measure, improved. The UCL and Exeter data says the lift is real. Your individual chapter genuinely reads better this afternoon. The cost is not local, so no local check catches it. It shows up two hundred pages later, when the book holds together beautifully and sounds like nobody in particular.

Sameness is not a quality problem. Your pages can get better and less yours at the same time.

I want to be precise about what I am not saying. I am not saying assistance ruins writers, and I am not saying the solution is to write alone in a cabin with a fountain pen. The same studies that measured the flattening also measured the lift, and the lift is real. The writers who will finish strong books in the next decade are not the ones who refuse help. They are the ones who take help from something that knows what their voice is and refuses to average it. That distinction - between assistance that pulls toward the middle and assistance that holds your line - is the whole game, and it is where this series is headed.

Full disclosure, since this is our blog: that line is the reason Write in a Click exists. We build voice profiles precisely because we think the flattening is the defining craft problem of this writing generation. But you do not need our workspace to take part one seriously. You need twenty minutes and your own manuscript.

Naming It Is the First Move

Every structural problem in a manuscript gets easier the moment it has a name. A sagging middle. An orphan scene. Now this one: the beige drift - the slow, suggestion-by-suggestion slide from your voice toward the average of every voice the model has read.

You cannot fix it with vigilance alone, because each individual suggestion passes inspection. You fix it the way you fix any craft problem: by understanding what it costs (part two), and by building a working method that protects the asset while keeping the help (part three).

For now, run the diagnostic above. Find your baseline page. And trust the instinct that made you pause over a paragraph that was polished, competent, and somehow not yours. That pause is your ear working. The rest of this series is about giving it something to do.