Make ChatGPT text sound human
ChatGPT's drafts have a voice people now recognise on sight. The odd part: other models sound so similar that detectors label their text ChatGPT too.
Updated October 1, 2026
How often unedited ChatGPT text is caught
- GPT-3.5 Turbo (PAN 2026)
- Texts200
- Caught at 1 human in 100 flagged100%
- GPT-4o (PAN 2026)
- Texts209
- Caught at 1 human in 100 flagged100%
- GPT-4o mini (PAN 2026)
- Texts214
- Caught at 1 human in 100 flagged99.5%
- GPT-4 Turbo (PAN 2026)
- Texts63
- Caught at 1 human in 100 flagged100%
- GPT-4.5 preview (PAN 2026)
- Texts45
- Caught at 1 human in 100 flagged95.6%
- GPT-6 Sol (rasbt human-vs-ai-50k)
- Texts157
- Caught at 1 human in 100 flagged85.4%
| OpenAI model (public test sets) | Texts | Caught at 1 human in 100 flagged |
|---|---|---|
| GPT-3.5 Turbo (PAN 2026) | 200 | 100% |
| GPT-4o (PAN 2026) | 209 | 100% |
| GPT-4o mini (PAN 2026) | 214 | 99.5% |
| GPT-4 Turbo (PAN 2026) | 63 | 100% |
| GPT-4.5 preview (PAN 2026) | 45 | 95.6% |
| GPT-6 Sol (rasbt human-vs-ai-50k) | 157 | 85.4% |
MELD on public sets, public_report.txt. Unedited ChatGPT text is caught almost every time. Newer models are slightly harder, which is why no detector's numbers stay current for long.
Every model sounds like ChatGPT
We asked a detector to guess which model wrote 600 drafts from Gemma and Qwen. It said GPT for 428 of them and Claude for 148. On a public set it also called Llama 3.1 8B's text GPT 79% of the time. Models are tuned toward the same polite, even, helpful voice, so their writing converges, and detectors learn that shared voice rather than any one model.
For you that's useful: the fixes are the same whichever model wrote the draft.
Edits that work on a ChatGPT draft
- 01Delete the first sentence if it restates the question or says what's coming.
- 02Delete the last paragraph if it starts with In conclusion, Overall or Ultimately.
- 03Cut most em dashes and every Moreover, Furthermore and Additionally at the start of a sentence.
- 04Turn "it's not just X, it's Y" into a plain claim.
- 05Make one sentence much shorter than the rest. Even rhythm is a tell.
- 06Add something ChatGPT couldn't know: what happened, to whom, when, and what it cost.
Before and after (illustration)
Written for this page to show the edits above, not taken from a test set.
- Great question! Choosing a CRM is a crucial decision that can significantly impact your business.
- Edited(cut: answer the question instead)
- Additionally, it's worth noting that pricing can vary widely — so it's important to consider your budget.
- EditedPrices vary a lot, so set a budget first.
- Overall, the right CRM will empower your team to unlock new growth opportunities.
- Edited(cut, or replace with what you'd actually pick and why)
| ChatGPT-style draft | Edited |
|---|---|
| Great question! Choosing a CRM is a crucial decision that can significantly impact your business. | (cut: answer the question instead) |
| Additionally, it's worth noting that pricing can vary widely — so it's important to consider your budget. | Prices vary a lot, so set a budget first. |
| Overall, the right CRM will empower your team to unlock new growth opportunities. | (cut, or replace with what you'd actually pick and why) |
Does asking ChatGPT to sound human work?
Partly, and it depends on the writing. In a public set where models were asked to explain things simply and avoid AI habits, MELD caught only 7.5% of GPT-5.6 Luna Pro's answers at its shipped line. The same approach for advice-forum posts was caught 94.9% of the time for that model. A prompt changes the words more reliably than the shape of the text.
What we haven't measured
Patois's rewrite numbers so far come from Gemma and Qwen drafts, not ChatGPT's. Given how alike their drafts read, we expect similar results, but we'll publish ChatGPT drafts as their own row when they've been run.
Questions
Can AI detectors catch ChatGPT?+
Unedited, almost always. On public test sets MELD caught 95.6–100% of GPT-3.5, GPT-4 and GPT-4o texts while flagging 1 human text in 100. A newer model, GPT-6 Sol, was caught 85.4% of the time.
Can teachers tell if I used ChatGPT?+
Often, by the voice, and detectors catch most unedited drafts. If AI help isn't allowed for the work, rewriting it doesn't change that.
Does asking ChatGPT to "sound more human" work?+
Sometimes. Simple explanations prompted to avoid AI habits were rarely caught, but advice-forum posts from the same prompt mostly were. In our pilot, models rewriting with a couple of examples got to 26% flagged after picking the best of four, and our trained model to 19.3% on one try.
Why do detectors say my Claude or Gemini text is ChatGPT?+
Models are tuned toward the same helpful, even voice, so detectors learn the shared voice. Ours called Gemma and Qwen drafts GPT or Claude 96% of the time.
What's the quickest fix for a ChatGPT draft?+
Cut the opening restatement and the closing summary, delete stacked transitions and most em dashes, and add one specific detail of your own.
Has Patois been tested on ChatGPT drafts?+
Not yet. Its numbers so far are on Gemma and Qwen drafts. ChatGPT drafts will get their own row when they've run.
Sources and dates
- MELD, the open-source detector we test with (Hugging Face)huggingface.co
- PAN 2026 and RAID texts (Hugging Face)huggingface.co
- rasbt/human-vs-ai-50k dataset (Hugging Face)huggingface.co
- mild-rgb ELI5 and AITA human-vs-AI sets (Hugging Face)huggingface.co
Facts about other products come only from these pages, on the dates shown; prices are as published that day. Macaron's own details are as shipped on October 1, 2026. Products change, so check their sites for the latest.