You probably noticed it. The em-dashes everywhere. The way every essay opens with "Let's dive into..." That exhausting symmetry of "It's not just X, it's Y." Lists of three, lists of three, lists of three. The closing paragraph that pivots gently to a forward-looking platitude. Once you see the texture, you can't unsee it. Within a year anyone reading a lot of writing online will pick out ChatGPT prose at fifty paces.
So why? The model trained on the entire literary corpus of the internet. Hemingway, Didion, Baldwin, Lebowitz, every great essayist who ever lived. It has access to every voice. So why does it always reach for the same one?
BECAUSE GENERIC IS THE SAFEST BET
Large language models work by predicting the most statistically likely next word given the words before it. That's the whole mechanism. It's not generating writing the way writers generate writing. It's not reaching for the right word, the surprising image, the specific verb. It's averaging the response patterns that most reliably appeared after similar inputs in the training data.
"Most likely" is by definition the middle of the distribution. It's the safest, most predictable, most consensus-validated next word. When the model is asked to write about marketing, it produces writing that statistically most resembles writing about marketing. Which means it produces writing that most resembles the average of every marketing blog post ever written.
This isn't a bug. It's the design. The model is doing exactly what it was built to do. The thing it was built to do, producing the most plausible response, is structurally the opposite of what good writing requires.
THEN HUMAN FEEDBACK MAKES IT WORSE
After the base model is trained, it gets fine-tuned through a process called reinforcement learning from human feedback. Workers rate outputs as better or worse. The model learns to produce outputs that rate well. The intent is to make the AI helpful, harmless, honest.
The unintended consequence: the model learns to produce outputs that please a wide pool of human raters. Pleasing a wide pool means inoffensive. Balanced. Hedged. Structured. Polite. Avoiding strong claims that some raters might disagree with. Defaulting to that cautious both-sides mildly-helpful tone nobody loves but nobody can object to.
That tone is the smell. It's why even a beautifully written ChatGPT response feels weirdly empty. The writing has been optimized to never offend, and writing that never offends rarely says anything either.
THE WRITER'S TRICK NEVER FIRES
Real writers do something the model can't. They make a choice. They prefer one word to another not because it's statistically most likely but because it's right. They cut the safe sentence and replace it with the strange one. They kill the symmetry. They let a paragraph end on the unexpected beat.
Those moves require a point of view. An internal sense of what's true and what's interesting that overrides the average. The model has no such thing. It has no point of view. It cannot prefer. It can only generate the most likely, then sand it smooth.
The reason ChatGPT sounds the same to everyone is that it's giving everyone the average answer. The average answer is, by definition, no one's actual answer.
HOW TO STOP IT
Inside general-purpose tools like ChatGPT, you have three real moves. None are perfect. Stacked together they help.
Constrain the output, not the topic. Don't say "write a blog post about loyalty programs." Say "write a blog post about loyalty programs that takes a position the average marketer would disagree with, uses no bullet points, no em-dashes, and ends mid-thought." Constraints force the model away from its default attractor.
Give it a voice to imitate. Paste in 500 words of writing you actually like and tell the model to write in that texture. Imitation isn't original. But it's at least specific. The model can mimic surface texture even when it can't generate original taste.
Use it for thinking, not finishing. Use the model to generate options, pressure-test ideas, find the angle. Then write the final draft yourself. The most generic part of any AI-assisted piece is usually the part you let the AI finish. Your voice goes in the last 20%, by hand.
OR USE A DIFFERENT KIND OF TOOL
The deeper fix is to stop using a tool optimized for the average when you need work that isn't average. There's a small but growing category of AI built on a curated philosophy instead of the open internet. Tools that don't try to give you the most likely answer because they were never trained to.
That's the problem Dante Peppermint was built to solve. Not to be a faster ChatGPT. To be the opposite kind of thing. An AI that thinks from a specific point of view rather than averaging every possible one. For most operational work, the average is fine. For creative and strategic work, the average is the enemy.
The em-dash isn't going anywhere. The hedging tone isn't going anywhere. They're features of how the dominant tool works, not bugs to be patched. If you want writing that sounds like a person, you need a tool that thinks like one. Or you need to do enough of the work yourself that the model never gets to finish the sentence.