Scroll through LinkedIn for five minutes and try to tell one post from the next. You can't. Not because the ideas are bad, but because every AI-assisted post shares the same rhythm, the same transition words, the same paragraph structure. AI writing has become wallpaper.
This is the central argument in a recent analysis from writing tool maker Noren, and they're not wrong. The piece pins down something most daily AI users have felt but struggle to articulate: the problem isn't quality. It's sameness.
How LLMs Flatten Your Voice
Large language models work by predicting the most likely next word (technically, "token") based on patterns in their training data. The output gravitates toward the statistical average of everything the model has ever read. Your quirks, your sentence rhythm, that weird habit of starting paragraphs with single-word fragments? Those are outliers. The model smooths them out.
Concretely, that means short punchy sentences get lengthened. Blunt openings get softened. Transition words like "however" and "moreover" appear where you'd never use them. The result reads fine. It just doesn't read like you.
Noren draws a useful distinction here: tone versus voice. Tone is the easy part, the surface-level "professional but friendly" instruction you drop into a system prompt. Voice is deeper. It's the words you reach for without thinking, the punctuation habits, the way you build an argument across paragraphs. A system prompt captures maybe 10% of that.
Why the Common Fixes Fall Short
Three approaches get floated constantly, and all have real limits:
- Prompt engineering - Describing your style in a prompt is like describing a song to someone who's never heard music. Too much nuance gets lost.
- Fine-tuning (retraining a model on your own writing) - Expensive, requires thousands of examples, and produces a black box you can't inspect or tweak.
- RAG (retrieval-augmented generation, where the AI pulls from your past writing as reference) - Good at recalling facts and phrases, but bad at extracting the structural patterns underneath your prose.
The practical result: people spend the time they "saved" on generation editing the AI's output back toward their own voice. Eventually most give up and accept the generic version. Multiply that across millions of users on the same models, and you get entire platforms where every post feels interchangeable.
The Real Cost Is Brand Erosion
For anyone whose writing is their product, from newsletter authors to consultants to founders building in public, this matters beyond aesthetics. Readers develop an intuitive sense of who wrote something. Swap that for AI-average prose and you trade years of built-up recognition for a marginal speed boost.
Noren is positioning itself as the fix, a macOS app that analyzes your writing patterns and steers AI output to match. Whether their specific approach works remains to be tested. But the diagnosis is sound: the current generate-then-edit workflow treats voice as an afterthought, and the content landscape is visibly worse for it.
The practical takeaway for anyone using AI to write: if you're not actively editing for voice, you're publishing someone else's. And that someone is a statistical composite of the entire internet.