Skip to content

textGrain · EU rollout

ChatGPT Watermark Remover for Text

OpenAI's textGrain (text grain) watermark is an invisible statistical signal in the model's word choices, now being added to eligible ChatGPT and Codex text in the EU. Weakening it means changing those words. In OpenAI's own test on 400-token passages, replacing a quarter of the words with synonyms cut detection from about 92% to 17%.

Privacy mode

OpenAI ChatGPT: current status

· Documented by the provider

OpenAI documents textGrain, a word-choice watermark, for eligible ChatGPT and Codex text in the EU.

Status
Rolling out
Mechanism
textGrain adds an invisible statistical signal to the model's word choices. API customers worldwide can opt in, and watermarking stays off by default. Eligible ChatGPT and Codex output in the EU receives it over the coming weeks. It is not a global default.
Hidden characters involved?
No. textGrain is a statistical signal in word choices, not hidden characters. Invisible characters in copied ChatGPT text still usually come from the copy path.
Public detector
No public detector. OpenAI is opening access first to approved researchers and expert organizations. Image and audio verification stays public and separate.

What the provider says weakens the signal

  • OpenAI's evaluation: replacing 10% of words with synonyms cut detection from about 92% to 66% on 400-token passages
  • Replacing 25% of words with synonyms cut that detection to about 17%
  • Shorter text, and text with little freedom in word choice such as mathematics, is harder to detect
  • OpenAI says text provenance has to account for how easily text can be rewritten, translated, or edited

How this works

Deterministic marks can be removed exactly. Probabilistic patterns can only be reduced. Simple Unmark keeps the two separate and does not promise detector outcomes.

  • On 5 October 2026 OpenAI published its approach to EU text provenance. From that day, API customers worldwide can opt in to text watermarking for select models, and it stays off by default. Over the coming weeks, eligible ChatGPT and Codex text in the European Union, on every plan, receives the watermark. OpenAI is not making it a global default.
  • The method is called textGrain. It adds an invisible statistical signal to word choices, and the detector needs the text plus a secret key. It is not hidden Unicode. Zero-width characters in a paste still usually come from the page, the editor, or the chat client, and deleting them does not touch textGrain.
  • OpenAI says editing can weaken the watermark, and that text provenance has to account for how easily text can be rewritten, translated, or edited. In their evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%, and replacing 25% reduced it to 17%. Shorter text, and text with little freedom in word choice such as mathematics, is harder to detect to begin with.
  • A rewrite replaces the word choices the detector scores. We do not have OpenAI's detector, so we do not claim a score. A detection would not say who used ChatGPT or whether the passage is accurate, and a miss would not prove a person wrote it. Image and audio provenance (C2PA and SynthID) is a separate system.

Three steps

  1. 1

    Paste the ChatGPT text. The local scan lists any invisible characters, which are a separate layer from textGrain.

  2. 2

    Run the clean. Hidden characters are removed, then the wording is rewritten so the original word choices are no longer the ones textGrain scored.

  3. 3

    Compare meaning, names and numbers, then copy. The rewrite changes the words; it does not give you a detector result.

Worked example

Before

The team decided to move the launch to 14 November. The main reason is that the payment integration still needs more testing before it can go live.

After

Launch is now set for 14 November. Payments are the holdup: the integration has to be tested further before customers can use it.

Illustrative editing example written by us. OpenAI's synonym test changed 10% to 25% of the words; this rewrite keeps the date and the reason and changes nearly everything else. It was not run through textGrain detection, which is not public, and real output varies per request.

Scope

What this page can and cannot do

Can

  • Remove and count the invisible code points it covers, exactly
  • Rewrite the passage so its word choices are no longer the model's original ones
  • Point you to the free invisible character remover when you only want hidden characters gone and your wording kept
  • Show you the before and after so you can judge the result yourself

Cannot

  • Guarantee that OpenAI's textGrain detector will miss the result
  • Prove or disprove that ChatGPT produced a passage
  • Remove textGrain by deleting invisible characters
  • Remove C2PA metadata or SynthID marks from images or audio. Simple Unmark handles text only

We don’t store submitted content or cleaned results in our application database, and we don’t send content to analytics. For rewrites, Private mode routes plaintext through the app, its load balancer and the processors. Confidential mode encrypts text from the browser to the workload, so the web backend only receives metadata. DeepInfra still rewrites the text. The deterministic scan in Invisible character remover runs entirely in your browser. See Privacy Policy and Methodology.

Questions

About ChatGPT Watermark Remover for Text

Sources

What is textGrain?
textGrain is OpenAI's text watermark, also written text grain. It adds an invisible statistical signal to the model's word choices. The detector looks for that signal. It is not a hidden character, and OpenAI plans to open-source the method later.
Does ChatGPT watermark text?
Eligible ChatGPT and Codex text in the European Union is getting textGrain over the coming weeks, on every plan. It is not a worldwide default. API customers anywhere can opt in for select models, and API watermarking stays off unless they turn it on, so text from an app built on OpenAI's API may or may not carry it.
Will changing words remove a ChatGPT watermark?
OpenAI's own evaluation says editing weakens it. Replacing 10% of words with synonyms cut detection from about 92% to 66% on 400-token passages, and replacing 25% cut it to 17%. A rewrite changes more of the wording than that synonym test. We still cannot promise a detector verdict, because the detector is not public.
Why do I keep finding zero-width characters in ChatGPT text?
Rendered pages, editors and chat clients insert them for layout. OpenAI does not describe any character, including special spaces or em dashes, as part of textGrain. The free invisible character remover lists each hidden character by code point. Deleting them is useful formatting cleanup, but it does not weaken the watermark.
Does OpenAI have a public text detector?
Not at launch. Access is opening first to approved researchers and expert organizations. Image and audio verification stays public. OpenAI retired its separate AI text classifier in 2023; that classifier was a guess, not a watermark check.