Why AI writes
like that.

I use these models every day. This post exists because I got tired of deleting em dashes.

You know the voice. Confident, warm, slightly breathless, allergic to committing to anything. It shows up in cover letters, LinkedIn posts, product pages, and lately in the abstracts of published research papers. The tells:

Here’s the part that makes it interesting: none of these are bugs, and none of them are habits the model “picked up”. Each one falls straight out of how the thing is built. Trace the mechanics and the style stops being a mystery. It becomes a receipt.

A language model never sees the sentence it is about to write. It sees the text so far, and it produces one thing: a probability for every possible next token. A token is a chunk of text, roughly three quarters of a word. Pick one, append it, repeat. That loop is the entire act of writing.

The bug was caused by ___

  " a"        21.4%   safe
  " the"      18.9%   safe
  " an"        8.2%   safe
  " missing"   4.1%
  " improper"  2.7%
  " race"      0.9%   <- the interesting one
  " cosmic"    0.002%
  ...and ~100,000 more, most near zero

Look at where the mass sits. The safe continuations soak up most of the probability. The interesting one is down at 0.9%. A human writer reaches for the 0.9% option on purpose, because surprise is the whole job. The model, sampling proportionally, takes the safe word almost every time. Now stack that choice thousands of times per page. Every individual pick is reasonable. The sum is prose where nothing unexpected ever happens.

To be fair to the machine: it is not purely myopic. Anthropic’s interpretability team found that when a model writes a rhyming couplet, it activates the rhyme word before starting the line, then writes toward it. So something like planning happens inside. But the output interface is still one token after another, and at each step the expected token usually wins. The plan is real; the accent it gets executed in belongs to the distribution.

# Voice lives in the low-probability choices. The model is built to avoid them.

There is literally a knob for this. Temperature rescales the distribution before sampling. At zero, the top token always wins and the model becomes a very expensive autocomplete. Turn it up and the tail gets a chance.

prompt: "The deadline was"

t = 0.0   the deadline was approaching fast.
t = 0.8   the deadline was two days away and nobody
          had opened the ticket.
t = 1.6   the deadline was a lighthouse nobody was
          swimming toward.

The lighthouse sentence is the kind of thing a good writer might keep. It is also two bad rolls away from word salad, and at that setting the model can’t tell the difference. Chat products ship somewhere around 0.7 to 1.0 because that’s the range where the model is reliable. Reliable is the operative word. Nobody ever described a writer they loved as reliable.

Pretraining is a compression problem. Take a huge slice of the written internet, and tune hundreds of billions of parameters so the model predicts every next token in it as well as possible. Hemingway and terms-of-service pages, physics forums and recipe blogs, all through the same weights.

The model that scores best on that objective is a kind of weighted average of every writer in the pile. And averages have a known property: they have no accent. Your strange sentence rhythm, the punctuation you abuse, the words you refuse to use, all of that is noise to the loss function. The flatness is the signal.

Here’s the twist though: the base model that comes out of pretraining is actually a decent mimic. It has seen every voice, so with the right prompt it can imitate almost anyone. Raw base models are weird, varied, occasionally brilliant. The sameness you and I are complaining about mostly arrives in the next step.

After pretraining, models are tuned on human feedback. The recipe: show people two candidate answers, record which one they prefer, train a reward model to predict those preferences, then push the language model toward whatever the reward model scores highly. This is what turns an autocomplete engine into a helpful assistant, and it works.

It also has a side effect with a name: mode collapse. Raters, working fast and paid per task, reliably prefer answers that look confident, structured, thorough, and inoffensive. The reward model learns that preference. The language model then abandons the base model’s whole range of registers and converges on the one that wins ratings. Every major lab runs a version of this pipeline, which is why every major model sounds like the same diligent intern.

My favorite piece of evidence is a single word. After ChatGPT launched, the word “delve” exploded in published medical paper abstracts, an order of magnitude above its old baseline. Why that word? The most-reported hypothesis: a lot of the human feedback work was done by annotators in Nigeria and Kenya, and in Nigerian business English “delve” is an ordinary, unremarkable word. The raters wrote and rated in their own dialect, the reward model soaked it up, and the model learned to delve.

Sit with that chain for a second. A word traveled from office English in Lagos, through a spreadsheet of preference clicks, into the weights of a model, and out into the abstracts of oncology papers written by people who never chose it. That is how style moves through this system: invisibly, statistically, and at scale.

# Nobody decided AI should write this way. A million small preference clicks did.

The obvious fix is to prompt the style away. “No em dashes. No rule of three. Write plainly.” It half works, and the half that fails is instructive.

• Prompts steer, weights pull. An instruction is a nudge applied at inference. The register is baked into the parameters. Ban a word and the model routes around it with a synonym carrying the exact same energy. You banned “delve”; enjoy “dig into”.
• Long outputs regress to the mode. The first paragraph obeys your style rules. By the fourth, the distribution has drifted home. Instructions decay; gradients don’t.
• The model can’t hear it. You detect the AI voice instantly because you’ve read a thousand samples of it. The model has no outside ear. From inside the distribution, the distribution is invisible.

This is why detection-by-vibes still works better than any classifier. The tells are statistical fingerprints of the training process itself, and the training process is the one thing the model cannot route around.

I still write with these models daily. The trick is assigning them the role they’re actually good at. A model is a mediocre author and a genuinely useful critic. So flip the seating.

Here's my draft. Do not rewrite it. Do not add anything.
Tell me: which sentences carry no information, where I
repeat myself, and where I buried the point. Quote the
exact lines.

Cut this by 30%. Removal only, no new phrasing,
no smoothing. If cutting changes the meaning
somewhere, flag it instead of fixing it.

Here are three things I wrote: <paste>.
Describe how I write, concretely. Sentence length,
rhythm, what I never do. Then check this new draft
against that description and list where it drifts.

The model writes like the average of everyone because that is, quite literally, what it is. Which settles the question of what your writing should be. The average is the one thing it already has.

# Your voice is a probability distribution the model has never sampled. Keep it that way.