How to make ChatGPT output sound like you wrote it

ChatGPT text sounds alike because the same habits recur from answer to answer. Wikipedia’s editors keep a field guide to them, with entries such as superficial analyses, promotional language and a high density of “AI vocabulary” words. Readers do notice: in a 2025 ACL paper, Jenna Russell, Marzena Karpinska and Mohit Iyyer had annotators label 300 non-fiction articles as human or AI, and the majority vote of five who frequently use LLMs for writing misclassified only one.

Why does ChatGPT text sound like everyone else’s?

The tells are mostly about content, not only word choice. The paper’s annotators leaned on AI vocabulary, and they also judged formality, originality and clarity. A list of banned words therefore ages fast, because Wikipedia’s guide notes that the words models overuse have changed over time. What lasts is a draft that says something specific in sentences you would write.

What should you give ChatGPT so it matches your voice?

Your own sentences carry more voice than any adjective you can use to describe it. OpenAI’s prompt-engineering guide recommends few-shot prompting, where the model learns a pattern from examples, and advises showing a diverse range of inputs with the desired outputs. Anthropic’s guide for Claude calls examples one of the most reliable ways to steer format, tone and structure, and suggests three to five of them. The same steps work in any chat window.

A workable instruction is short and concrete. It says who reads the text, how long the sentences run, whether the writer uses “I”, and which samples to follow, then asks for a draft that uses only the facts supplied. Give ChatGPT the samples first and the request last, so that the examples frame the task and are not read as an afterthought.

• Paste three short samples of your writing on different subjects, such as an email, a report paragraph and a message to a colleague.
• Describe the target in positive terms. Anthropic’s example replaces a ban on markdown with a request for “smoothly flowing prose paragraphs”, and OpenAI’s guide likewise asks for the assistant’s communication style to be described.
• Hand over the material only you hold: the claim you stand behind, the figure you checked, the reader you are writing for.
• Ask for a first draft, and treat it as raw material to edit rather than as the final text.

Which habits should you cut from the draft by hand?

Cutting habits by hand works better than forbidding them in the prompt, because a banned habit often returns in another form. Wikipedia’s guide calls its patterns “only potential signs of a problem, not the problem itself”, and the table below follows that reading: each row is a habit from the guide, with the edit that replaces it with something specific.

Punctuation belongs in the same pass, as a habit to check and not as a verdict. Keep the em dashes you would write yourself and cut the ones that stand in for a decision about sentence structure. Reading the draft aloud finds most of the rest, because stock phrasing is easy to hear and hard to see.

How do you check the draft against your own writing?

To check a draft against your own writing, put it beside the samples you pasted and look at three things. Start with sentence length, because a draft whose sentences all run the same length reads flat. Then check person: if you write “I” and the draft says “one” or “users”, change it. Last, scan for the phrases you never use. Keep that list in a note, add to it each time you cut something that was not yours, and paste it into the next prompt as a short instruction.

A second pass with a specific request works better than a vague one. Asking ChatGPT to “make it sound more human” invites more adjectives. Asking it to shorten the longest sentence in each paragraph, to replace a general claim with the figure you gave it, or to remove the closing summary gives the model something it can do, and gives you something you can verify in the result.

What does matching your voice not do?

Matching your voice makes a draft read like you. It does not make authorship verifiable, and it is not a way past a detector: a detector score is weak evidence in either direction.

A voice edit is also separate from watermarking. Anthropic writes that “Light editing probably won’t remove the watermark completely”, and that statement concerns Claude’s mark specifically. Other providers document their own, and the related articles cover what each provider says.

For a long draft, our humanizer rewrites stock phrasing into plainer language and shows every change so you keep only the edits that sound like you, and our AI detector reads a draft for the same habits and explains what it found, as a reading of style and not evidence of who wrote the text.