How to make ChatGPT sound more human: prompts that work, and where they stop

To make ChatGPT sound more human, put the words and writing habits AI overuses (like "genuinely", "honestly", em dashes and unasked-for headers and bold) into your custom instructions as things to avoid. Also, for each chat, include samples of your own writing and a description of your audience. These go a long way, but you can't fully change the cadence of ChatGPT's writing. And when you ask ChatGPT to rephrase its own writing, it usually keeps many of the same habits.

What makes ChatGPT sound like AI in the first place?

There are certain words and rhythms which are key to ChatGPT's voice. To see what they are, we ran two studies. In the first, we asked OpenAI's latest models (GPT-5.5, GPT-6.1 and chat-latest), as well as seven other 2026 models, to rewrite a dataset of 7,554 passages from humans' blogs. There are many words these models used much more often than the original authors. When rewriting, across all of these models, they used "genuinely" 12.5 times as often as the original authors, "roughly" 7.7 times, "plenty of" 7.2 times and "essentially" 7.1 times. The humans in our dataset very rarely used the words "for instance", but the AI models used it 0.25 times per thousand words.

ChatGPT also has a very specific cadence. In the rewriting study, the coefficient of variation of sentence lengths in the bloggers' passages was 0.65. In OpenAI's rewrites, it was 0.59. Additionally, 12.0% of the bloggers' sentences started with "And", "But" or "So", compared to 9.7% of OpenAI's. This makes its sentences flow smoothly, but in a somewhat flat way.

In the writing from scratch study, we had 4 different models (all from 2026) write 1,177 documents from scratch based on notes from a real human document. That study used OpenAI's GPT-5.4 mini model, which is different from the models in the rewriting study, so the numbers aren't directly comparable. GPT-5.4 mini produced more uniform rhythm than any other model (coefficient of variation of 0.43 vs. 0.50 for the humans), used "not X but Y" sentences more than any other model (0.29 per 1,000 words, ~2.4x more than humans), and used 1.99 em dashes per 1,000 words vs. 0.72 for the humans. All the models also used unrequested headers, bold, and bullet points.

Some words AI used to be known for no longer stand out. Neither group very often uses the word "delve", and AI uses the word "leverage" about as often as humans do. Thus, adding 2023's list of banned words to a prompt wastes valuable prompt space on words the model already avoids. For the full list, see our guide on words and phrases that AI uses too often.

Should I use custom instructions or a prompt in each chat?

Use both. Things like what you don't want the AI to write (certain words, punctuation, how you want it to end a piece) can go in custom instructions, since they apply to most writing. For each chat, write out a prompt that includes details about the writing: who it's for and what you want the reader to do. If you have some samples of your own writing, include those.

To add custom instructions in ChatGPT on the web or desktop, go to Settings, then Personalization, make sure Enable customization is turned on, and enter your instructions in the Custom Instructions field. In the iOS or Android app, go to Settings, then Customize ChatGPT. Per OpenAI's help center, Free and Go users can write up to 1,500 characters of custom instructions, and Plus, Pro, Business, Enterprise and Education users up to 5,000. The block below is under 900 characters, so it fits every plan.

While you test, make sure the Base style and tone menu on the same page is set to Default. Otherwise, if it's set to Professional, Friendly or one of the other presets, it will mix that style with your instructions and can pull against them. If you're using Projects in ChatGPT, you can add instructions to the project that will override your global custom instructions for any chats in that project. This allows you to create a project for client emails, for example, with its own rules.

What instructions should I give ChatGPT?

Here's what to paste into your custom instructions. It's based on our data about what ChatGPT overuses, not on a generic list.

Write like a person who knows the subject and respects my reader's time.
Start with the point. No warm-up line, no recap at the end.
Don't end with a question or an offer to help further.
Avoid these words: genuinely, truly, honestly, simply, essentially, roughly, crucial, "rather than", "plenty of", "for instance".
No em dashes. Use commas, periods or parentheses.
Vary sentence length. Some sentences should be very short. Starting a sentence with And, But or So is fine.
Don't group things in threes out of habit. List as many items as there really are.
No "it's not X, it's Y" or "not just X, but Y" sentences.
No headers, bold or bullet points unless I ask for them.
When I give you samples of my writing, copy their sentence length, word choice and punctuation instead of your own.
If you need a fact you don't have, write [CHECK] instead of inventing one.

Why each line is there

"Write like a person who knows the subject..." sets the tone by telling ChatGPT to write like a specific person, instead of giving it a list of things to avoid. Negative rules alone tend to produce stiff text. It's better to tell it what you do want, as it results in better writing.

"Start with the point" is there because AI often rephrases the question in the first sentence and then recaps the answer at the end. This makes the writing feel bloated and AI-like.

"Don't end with a question or an offer" is there because ChatGPT often asks if you'd like it to do something else at the end of its response ("Want me to turn this into a checklist?"), and this often gets mixed into drafts you'll send to other people.

The word list is the biggest outliers in our data. When rewriting, across the models, "genuinely" was used 12.5 times as often as by the humans, "roughly" 7.7 times, "plenty of" 7.2 times, "essentially" 7.1 times, "truly" 5.4 times, "honestly" 5.3 times, "simply" 4.7 times and "crucial" 3.3 times. We also included "for instance," as it was almost never used by the bloggers. When writing from scratch, the words that overlap are "honestly" (8.3x the human rate), "rather than" (4.3x), "genuinely" (4.2x) and "roughly" (4.7x), which is why "rather than" is now on the list. The two phrases that were characteristic of rewrites ("for instance", "essentially") showed up less in the drafts than in the people's writing. Words like "delve" and "leverage" are not overused by 2026 models, so we don't include them here. Also, keep this list small. Every word you add is one more thing the model has to track.

"No em dashes" is there because OpenAI's models added lots of em dashes when rewriting (3.88 per 1,000 words compared to 2.56 for the bloggers). The bloggers were especially likely to use em dashes, though: across 22 million words of all types of human writing, people used far fewer (0.73 per 1,000 words). GPT-5.4 mini also used more em dashes than the humans when writing from scratch (1.99 vs. 0.72 per 1,000 words). Sam Altman said in November 2025 that ChatGPT now follows this instruction when it's in custom instructions, so it's easier to add it there than to include it in every chat. Learn more about why ChatGPT uses so many em dashes.

"Vary sentence length" and the line about And, But and So help create a more natural rhythm, which word bans can't reach. In our rewriting study, the bloggers started 12.0% of their sentences with And, But or So. In our writing from scratch study, humans started fewer sentences this way (3.8% vs. 3.4% for the AI), so this is partly just a feature of bloggers. Sentence length matters more: in the writing from scratch study, GPT-5.4 mini was more uniform than the other models we tested.

"Don't group things in threes" is there because AI tends to group things in threes much more often than humans do. Pangram, which builds AI detection tools, finds about four times as many triads in AI-generated text as in human writing. In our from-scratch drafts the gap was smaller, 1.4 times (5.60 vs 4.00 per 1,000 words), but it was there.

The "not X, but Y" line is mostly just for GPT. Neither of our studies showed that it was generally a feature of AI writing (the rewrites used "not X, but Y" slightly less than the bloggers, and the drafts used it the same as people). But GPT-5.4 mini used it way more than people when writing from scratch (~2.4x more), and it showed up in places like LinkedIn posts and academic abstracts where the original human writing didn't have it. Pangram finds that AI-generated text has about three times as many instances of "not just X, but Y" as human writing. Read more about why AI loves "it's not X, it's Y".

"No headers, bold or bullet points" is there because, in the writing from scratch study, formatting was by far the most telling feature of AI writing. Despite the prompt not mentioning it at all, the AI writing included lots of unrequested formatting: 10.1 bold phrases, 4.2 headers, and 8.9 bullet lines per 1,000 words (compared to 2.4 bullet lines per 1,000 words in the original human writing). Even emails had lots of bold (5.8 per 1,000 words). This line keeps the writing right for emails, not slide decks.

The samples line tells the model what to do with samples of your writing you paste in later, so you don't have to explain it every time.

"[CHECK]" is a safety line. AI often makes specific things up when it's pushed to sound specific, so it's good to have a visible placeholder, which is easier to catch than a confident invented number.

Does it help to give ChatGPT samples of my own writing?

Yes, it's the most helpful thing you can do. Give it two or three examples of previous writing of your own of the same type (e.g. if it's a sales email, show three past sales emails, not a blog post) and tell it what you want it to copy from them (e.g. sentence length, how you start emails, how you end emails), but not to copy the content.

Make sure the examples you give are ones you wrote yourself, without AI. If not, you're just training it to write like AI again. And make them as short as possible. A few hundred words of your real voice should be enough.

Why should I tell ChatGPT who the writing is for?

One of the reasons for generic-sounding writing when using AI is that it's written to a generic audience. So if you ask the model for a follow-up email, you will get a generic follow-up email. But if you ask for a follow-up email to a CFO who saw your demo last Tuesday, liked the reporting but thinks the price is high, and will probably read it on her phone between meetings, then your email will be much shorter and more targeted, because the model now has a reason to leave things out.

Spending three sentences or so describing your reader, what they already know and the one thing you want them to do will help more than another paragraph of style rules.

Can I have ChatGPT edit its own writing to sound more human?

Yes, but tell it specifically how to change the text. If you just ask it to make something sound more human, you'll get another set of clichés. Specific edits work better: "Cut the first and last sentence if they only restate the point." "Break up any sentence over 25 words that holds two ideas." "List every word from my avoid list that's still in the draft." Then read it out loud yourself. Your ear catches the rhythm problems the model can't see in its own work.

Then read the draft out loud yourself to check for rhythm issues. Your ear catches what the model can't see in its own work.

Where does prompting stop working?

At the model's default voice. You can fix specific things like words and formats with prompts. But prompts only partially fix things like rhythm and sentence structure, and even that doesn't stick for long.

There are three reasons for this. First, if you say "don't use the word genuinely", it'll just use a similar word, like "really", instead. Second, as you continue the chat, your instructions will get further and further away from what the model is looking at when it generates text, so it'll gradually revert to how it normally writes. (This is why we suggest you start a new chat for every piece of writing.) Third, you can't use ChatGPT to fully fix ChatGPT. It can fix specific things like the words you name, but it won't recognise its own writing preferences (learn more in why your AI can't fix its own writing).

That's why we suggest using a different model entirely for the rewrite. It won't have learnt ChatGPT's preferences, so it won't have the same blind spots. Below is an example of typical ChatGPT-style writing and the same email after it's been rewritten using Zavus.

For most writing, the custom instructions we suggest above, plus a sample and telling the model who you're writing for, will be enough. But if you're writing something that will go out under your name (whether that's to clients, customers or the press) and it still sounds like ChatGPT even after doing all of the above, then it's worth using a second model. To understand the full picture, from why models converge on one voice to what each fix can and can't do, check out our guide to making AI writing sound human.

Zavus is a separate model you run on your finished writing so that it sounds human. It connects to ChatGPT and other agents via MCP; see how to add a humanizer to ChatGPT, Claude and Cursor. If you're looking to make your team's outreach sound like a person wrote it, check out Zavus for sales teams.