How to keep your writing from sounding like AI

Many people have tried getting help from ChatGPT or Claude to write a text and found that the result sounds strikingly uniform and a bit artificial. A language model is a computer program trained to write and understand text by having read enormous amounts of text beforehand. The tech figure Thomas Ptacek believes he has found a solution: use the language model as a proofreader, never as the one who finds the words for you.

His first and most important rule is simple: you may not use a single word the model suggests itself. According to Ptacek, language models are unusually good at landing on elegant and easy-to-read phrasings. But those are precisely the phrasings that reveal a text has been polished by a machine rather than a human.

The second rule is about praise. Ptacek warns against letting the model encourage you along the way. If it tells you your draft is already good, you risk keeping weaknesses you would otherwise have fixed. You need to be able to look critically at your own text, and that becomes harder if a chatbot keeps patting you on the back.

Four steps in practice

Here's how to do it in practice, according to Ptacek:

First, you write the entire draft yourself, without help from a language model. The ideas, structure and phrasing should be your own before the model comes anywhere near the text.

Next, you ask the model only to point out problems, not to solve them. That could be, for example, repetitions, heavy sentences or filler words like "very" and "actually."

Then you rewrite the places the model has pointed out yourself. This is where the discipline lies: you have to find your own words instead of copying the model's suggestions.

Finally, you can do a kind of blind test. You show the model two versions of a paragraph, without saying which one is new, and ask it to judge which is stronger. That forces a more honest assessment than if you just ask whether the text has gotten better.

Other uses of language models

Ptacek also uses language models for more limited tasks such as checking facts, fixing spelling errors and finding a precise word when he's missing one. What these tasks have in common is that the model has to confirm or correct something concrete, not keep spinning out your sentences.

Simon Willison, who develops software and shared Ptacek's text on his blog, makes the same point. The rule of never using a suggestion from the model is uncomfortable to follow in practice, but that very hassle is part of the purpose. It forces the writer to stay in their own style, while the model is used for what it's actually good at: finding weaknesses you're blind to yourself.