[ Patois · in testing ]

Undetectable AI writing.Reads like a person wrote it.

Paste a draft that reads as AI. Patois finds the passages a detector flags and rewrites them a sentence at a time, keeping what you said. Measured in the open, misses included.

See how it scored Try the phrase checker

Not open to the public yet.

Rewrite

A LinkedIn post, 132 words

7.9 s

Before

99%

reads as AI

After

reads as a person

2 passages flagged and rewritten · no number added

[ Before and after ]

A real run, unedited.

An AI draft of a 2021 LinkedIn post, and what Patois gave back. Every figure in the draft is still there, and no number was added.

The draft

99%AI likelihood · reads as AI

Antarctica holds 90% of the world's ice and 70% of its fresh water, yet these vital resources face unprecedented threats from warming air and waters. Protecting the Southern Ocean is a defining challenge of our time, offering a unique opportunity to unite nations for a common cause. By establishing large marine protected areas, we can boost global ocean conservation by an order of magnitude. We now have a once-in-a-generation chance to safeguard over 7 million square kilometers of the Southern Ocean. To create the largest act of ocean protection in history, we need your help to reach one million petition signatures. Please join us immediately at only.one/antarctica. Every signature counts. Antarctica and the incredible creatures that call it home need us more than ever. Let's act decisively to secure our planet's future.

Typical model phrasing

Patois's rewrite

6%AI likelihood · reads as a person

Antarctica contains 90% of the world's ice and 70% of its fresh water, and these are under threat from warming air and waters. Protecting the Southern Ocean is an opportunity for many nations to unite on one common cause. Marine protected areas would boost ocean conservation around the world by an order of magnitude. Now is a once-in-a-generation opportunity to protect more than 7 million square kilometers of the Southern Ocean and if you help us reach one million signatures for the petition then we'll have the biggest act of ocean protection in history at only.one/antarctica Antarctica and the incredible creatures that call it home need us more than ever.

Said differently

Scored by MELD, an open-source detector
Next: Pangram, GPTZero, Turnitin

[ How it rewrites ]

The detector steers the writing, not just the grade.

Three steps, the same for every passage, and you see every change at the end.

2 passages flagged
01Detect
A detector scores every paragraph and marks the ones that read as AI
02Rewrite
Each next sentence four ways; the detector keeps the one that reads most human
- Meaning kept
- Length within range
- No new numbers
03Guard
Meaning and length kept, no new numbers, or it's thrown out and redone

[ What it writes ]

Trained on how people actually write.

A 9-billion-parameter open model, trained on thousands of pairs: an AI draft in, the human original out. It learned to undo what AI drafting does.

Measured on 300

Posts

Trained on thousands of real LinkedIn posts from 2021, and tested on 300 more by people it never saw.

Trained on

Explainers

Well-edited educational writing from before 2022, so explanations read as taught, not generated.

▲42▼

Trained on

Answers

Voted forum answers: plain and direct, by someone who knows the thing. Emails get the same rewrite; they aren't measured yet.

[ The results ]

AI drafts all land in one place. Rewrites land where people are.

Detector scores for 300 LinkedIn posts by 12 people the writer never trained on. Each dot is two posts.

← Reads like a personReads like AI →

of real posts flagged: 0.7%
of AI drafts flagged: 57.3%
flagged after one rewrite: 2.3%
flagged after the best of four: 0.3%

[ Not proven yet ]

What these numbers don't show, so you can weigh them.

One detector so far: MELD, an open-source detector. It also picked the best of four, so the one-try number is the fair one. Pangram, GPTZero and Turnitin haven't been run.
The flag line was set from other posts by the same 12 people, so 0.7% for real posts is the line agreeing with itself, not a fresh test.
The test posts are from 2021. How it does on posts written today is untested.
It flattens bullet lists into paragraphs, and once added a job title the draft didn't have. Read it before you post it.
English only.

[ Why ]

Models learned to write from millions of people. Then they were tuned until they all sound like one. Detectors learned that one voice. Patois takes it back out.

[ Questions ]

The short answers.

Is undetectable AI writing actually possible?

Against one detector, on these posts, nearly: after the best of four rewrites, 1 post in 300 was still flagged. No tool can promise every detector will pass every text, so this page shows how it was measured instead.

Does it get past Turnitin or GPTZero?

Not tested yet. The results here are from MELD. Outside detectors are next, and their numbers will go on this page when they've run.

Will it change what I said?

It's built not to. Each rewrite has to stay close in meaning and length to your draft and can't add a number you didn't write, and you see every change. It can still add a name or a detail now and then.

Can I use it now?

Not yet. Patois is in testing and isn't open to the public.

Can I use it for schoolwork?

Not where your school bans AI help. It's made for your own posts, emails and messages, where the problem is that an AI draft reads like AI, not who wrote it.

[ Keep reading ]

Free AI phrase checkerSee what reads as AI in your own draftRead Do AI detectors flag human writing?27% of Wikipedia openings were flaggedRead Signs of AI writing, countedWhich tells hold up across 1,500 textsRead Can you bypass AI detection?What moved the score, and what didn'tRead