How to make ChatGPT sound less like AI.

To make ChatGPT sound less like AI, give it samples of your writing and the facts it can use, ban the sentence patterns that give AI away, and edit the draft by hand. The five real examples here are lines AI wrote for this site, each cut before the page went live.

  1. Write a short voice brief with two samples of your own writing.
  2. Keep a banned list of the words and sentence patterns you’d never use.
  3. Ask for one draft from your outline and a list of everything it added.
  4. Read the draft out loud and rewrite every line that isn’t you.
  5. Add each line you cut to the banned list, with the reason.

Why does ChatGPT sound so robotic?

ChatGPT writes by predicting likely next words, so anywhere your instructions leave a gap, it fills the gap with common wording. A draft can follow every instruction you gave it and still contain lines you’d never say.

Each line in this table came from an AI draft of a page on this site, written with Claude Code or Codex.

#What the AI draft saidWhy I cut itWhat I published
1“You can’t train ChatGPT on your business in the way the word suggests.”It hedges a claim that should be plain“You can’t train ChatGPT on your business with one prompt you saved from Instagram or Threads.”
2“That’s the whole job, and the first two levels only take about fifteen minutes.”It tells the reader the task is easy“Writing that description and saving it in ChatGPT or Claude takes about fifteen minutes.”
3“Read every line and fix anything that isn’t true yet... Then save it somewhere you can find it in ten seconds.”A coaching voice, plus a timing nobody measured“Check the draft line by line and correct anything that’s wrong or out of date, then save it somewhere you can grab it fast, like a pinned note.”
4“I run a one-person marketing agency, and I check everything AI helps me make before it goes out.”It isn’t how I’d describe my own work“I run a one-person marketing agency, and I’ve integrated AI into my business to help me be more productive.”
5“...the mistakes AI made in my own marketing work...”It made AI’s mistakes sound like mine“Pretty Smart Prompts shares the checks I run on every AI draft before it goes out, with real examples of what they catch.”

After cutting lines 1 to 3, I added a check for each one to AI slop checks, so the tool flags all three today. Claude Code wrote lines 4 and 5 with my banned list already loaded, and they still pass, because what they get wrong is a fact about my business. A pattern check can’t know what your business does or how you’d describe it, so the final pass is always up to you.

Terminal output of AI slop checks on the five table lines, with lines 1 to 3 flagged as rejected and lines 4 and 5 showing 0 flags
AI slop checks run on lines 1 to 5 of the table on November 15, 2026. Lines 1 to 3 trip checks I added after cutting them, and lines 4 and 5 pass because the checker can’t see a wrong fact about my business.

Can ChatGPT match my writing style?

ChatGPT can get close to your writing style when you give it a short voice brief. Adjectives alone, like “warm and direct,” describe almost every brand, so pair them with the specific patterns you’d never use.

  1. 01
    Who you help.

    “I help [specific person] with [specific job].”

  2. 02
    Your point of view.

    “My take on this is [opinion you can defend].”

  3. 03
    How you sound.

    Paste two short paragraphs you’ve written, on the same subject if you have them.

  4. 04
    Words you avoid.

    Name the phrases and promises that don’t belong to you, and keep each line you’ve cut next to what you wrote instead.

  5. 05
    Current facts.

    Add your offer, your policy, and any details the draft is allowed to mention.

The voice line from the context block I wrote for this site in October pairs three adjectives with four bans.

Voice: plain, warm, skeptical, first person “I,” with contractions. No em dashes, no hype, no three short sentences in a row, no “it’s not X, it’s Y.”

I tested the difference on November 15, 2026, in a ChatGPT temporary chat, which ignores memory and custom instructions. Asked for an About page intro with no brief, it opened with “Welcome to Pretty Smart Prompts—where marketing know-how meets the power of AI” and promised to “streamline workflows.” With my brief in the prompt from the next section, it opened with “If you use AI for client work and want the final draft to sound like you, this site is for you,” with no em dash and no buzzwords.

Ban the sentence patterns that give AI away

Banning words like “delve” and “tapestry” is standard advice, and Google’s AI answer for this search recommends it too (checked November 10, 2026). Swapping one word still leaves a sentence built on the same formula, so my banned list covers sentence patterns as well, including these six.

PatternExampleWrite this instead
A colon that sets up a reveal“The result: twice the calls.”Join the two halves into one sentence
A filler transition“Here’s the thing”Cut it and make the point
The same sentence frame repeated in a row“No contracts. No setup fees. No surprises.”Make each line add something new
A false contrast“It’s not X, it’s Y”State the point directly
Three short sentences of the same length in a row“We listen. We plan. We deliver.”Join two of them so the rhythm varies
A section that opens with a rhetorical question“But what does that actually mean?”Open with the answer

The cut lines in the first table added three more patterns to my list, covering hedged claims, easy-task reassurance, and a coaching voice.

The full list is free on GitHub as AI slop checks. It has a checklist with a prompt for ChatGPT or Claude chat, plus a skill for Claude Code and Codex that runs 21 checks on a draft and rewrites each flagged sentence. The checker started with 16 checks on November 9, 2026, and had 21 by November 14.

Ask for one draft and a list of what it added

This prompt asks for a draft plus a change log, a short list of the facts it used and what it added on its own. The log won’t catch every addition, so compare the draft with your sources yourself.

Copyable prompt

I run [type of one-person business] for [specific audience]. I need a [email / page section / caption] that helps the reader [specific task].

My voice examples: [paste two short things you wrote].
My point of view: [one specific belief].
My outline: [paste your points in the order you want them].
Facts this draft may use: [paste verified offer, process, price, timing, or policy].
Words or claims to avoid: [your list].
The reader's question or situation: [paste it].

Write one draft in plain language. Keep it under [word count]. Answer the question before describing my business. Do not invent proof, testimonials, urgency, prices, results, or policies. Mark any missing fact as [CHECK].

Then give me a short change log: which facts you used, which voice examples shaped the draft, and what still needs my review.

In the test above, the change log flagged one addition, “my daily work” where my facts said “my business.” It missed that my point of view had turned into a general claim, “AI works better with context going in and a human edit coming out,” which I only caught by comparing the draft with what I gave it.

If the draft sounds polished but vague, ask it to name the customer’s exact next step. When it can’t, add source material before you touch the tone.

Should I edit ChatGPT's first draft?

Yes, edit ChatGPT’s first draft line by line after you’ve given it your outline. Custom instructions don’t fix your voice. A humanizer or a prompt that strips out AI tells is fine as a first pass, but neither one knows the facts that made lines 4 and 5 wrong. AI can’t speak for you, because you’re the one with the experience, so keep the final say and plan on a few rounds of edits.

  1. 01
    Read the draft out loud

    and mark every line you wouldn’t say to a customer.

  2. 02
    Find the plain claim

    under each marked line.

  3. 03
    Rewrite it

    in your own words.

  4. 04
    Save the edited version

    as your next writing sample.

Take line 3 from the table. Out loud, “somewhere you can find it in ten seconds” sounds like a coach, and nobody timed the ten seconds. The plain claim under it is “keep it somewhere easy to reach,” so the published version says “save it somewhere you can grab it fast, like a pinned note.”

Before anything goes out, check its promises against your real price, timing, result, and policy, using the full review in How to Edit AI-Generated Content.

Each line you cut goes on the banned list with the reason, the same way lines 1 to 3 in the table became checks. The edit takes minutes to hours. For example, reviewing this site’s homepage, About page, and guides line by line took about four hours.

Will ChatGPT remember my voice in every chat?

ChatGPT remembers your voice only where you’ve saved the voice brief. Custom instructions cover every chat, while a project or a custom GPT covers the chats inside it, and how much each one holds depends on your plan. How I Trained ChatGPT and Claude on My Business walks through each option. The voice brief and banned list are plain text, so the same files work in Claude.

Check your next draft for AI writing patterns.

  • The full banned list and a paste-in prompt for ChatGPT or Claude, no install needed
  • A Claude Code and Codex skill that runs 21 checks on your draft and rewrites each flagged sentence
  • Free on GitHub, with no email signup
Get AI slop checks on GitHub

Sources

  • Google search results and AI answer for “how to make chatgpt sound less like ai,” United States, checked November 10, 2026
  • AI slop checks, the public banned list and checker. The 21-check count and the flags on lines 1 to 3 were rechecked November 15, 2026.
  • How I Trained ChatGPT and Claude on My Business, the page lines 1 to 3 were cut from, and the guide to custom instructions, projects, and custom GPTs
  • How to Edit AI-Generated Content, the full review for a finished draft
  • About and the homepage, the pages lines 4 and 5 were cut from. Line 5’s replacement was the homepage subhead from November 10 to 15, 2026.