You've noticed it. Everyone has. AI-generated text has a sameness to it. Different people, different prompts, different topics, and yet the output reads like it was all written by the same mildly enthusiastic middle manager. There's a mathematical reason for this, and understanding it is the first step to fixing it.

Token prediction, simply

Language models work by predicting the next word (technically the next token, but "word" is close enough). Given everything that came before, the model assigns a probability to every possible next word and picks from the top candidates.

This is the entire game. Every word in every AI-generated sentence is the result of a probability calculation. The model isn't thinking about meaning or style or voice. It's running statistics on sequences.

When you ask ChatGPT or Claude to "write a blog post about productivity," the model starts predicting words. And it predicts the most likely words. The words that appeared most often in similar contexts in its training data. The average words. The consensus words.

Why this creates convergence

Think about what "most likely next word" means across millions of users. If a thousand people ask for a blog post opening, the model gravitates toward the same high-probability sequences every time. "In today's fast-paced world" isn't a creative choice. It's a statistical attractor. The model produces it because those words, in that order, appeared constantly in its training data.

This is voice flattening. The model has seen millions of writers and learned a composite of all of them. When it generates text without specific style constraints, it outputs that composite. The average of a million voices isn't a voice. It's background noise shaped like prose.

The measurable differences

Linguists have known for decades how to distinguish individual writing styles from each other. The same metrics reveal the gap between AI writing and human writing. (We dig into the underlying research on our research page.)

Burstiness. Human writers alternate between short punchy sentences and longer complex ones. The pattern is irregular and personal. AI writing has remarkably consistent sentence lengths. A human might write sentences of 4, 22, 8, 31, 6 words. AI tends to hover in the 15-20 word range with small variations. Measured burstiness in human writing is typically 2-3x higher than in AI output.

Type-token ratio. This measures vocabulary diversity: how many unique words appear relative to total words. Human writers, especially experienced ones, use more varied vocabulary. AI text recycles the same words and phrases. Measured across thousands of samples, human writing scores 15-25% higher on type-token ratio.

Sentence length variance. Related to burstiness but specifically about the standard deviation of sentence length. Human writers average a standard deviation of 8-12 words. AI writing sits around 3-5. That's a massive difference. It's the difference between prose that has rhythm and prose that drones.

Opening patterns. Human writers start paragraphs in unpredictable ways. AI writing overwhelmingly opens with a subject-verb construction or a dependent clause. Measure the first three words of every paragraph in an AI-generated piece and you'll find maybe 4-5 patterns. Do the same with a human writer and you'll find 15-20.

What "voice flattening" actually means

Voice flattening isn't a metaphor. It's a literal statistical phenomenon. Your writing voice is a distribution across many measurable dimensions: word choice, sentence structure, paragraph length, opening patterns, vocabulary range. Plotted on a graph, your voice has peaks and valleys. It's spiky. It's distinctive.

AI output has the same peaks and valleys as every other AI output. Plotted on the same graph, it's flat. Not flat as in boring (though it often is), but flat as in lacking the statistical distinctiveness that makes one writer sound different from another.

When you use AI without style constraints, your output loses those personal spikes. Your 6-word sentences disappear. Your unusual word choices get replaced by common ones. The AI keeps your ideas but strips your fingerprint.

How voice profiles counteract this

A voice profile is a set of constraints that push the model's probability distribution away from the average and toward your personal patterns.

When your profile says "average sentence length: 11 words, standard deviation: 9 words," you're telling the model to produce burstier text. When it says "never use the word 'comprehensive,'" you're blocking a statistical attractor and forcing the model toward less common choices. When it says "open 30% of paragraphs with a single-word sentence or fragment," you're injecting a structural pattern the model would never choose on its own.

Each constraint is small. Together they shift the model's output from "generic average" toward "sounds like a specific person." The more constraints, the more distinctive the output.

This isn't magic. It's statistics pushing back against statistics. The model wants to converge on the mean. The profile forces it to stay near your personal distribution instead.

The gap between knowing and fixing

Understanding why AI sounds generic is easy. Fixing it requires knowing your own statistical patterns well enough to encode them as constraints. And as I've written about in how to make AI write like you, people are bad at self-reporting their writing style.

The reliable approach is measurement. Analyze what you actually write. Extract the patterns. Convert them into model constraints. That's what we built MyDamnVoice to do.

Your voice has a shape. It's worth preserving.

Why AI writing sounds the same (and the math behind it)