OpenAI has changed the AI-detection debate. On 5 October 2026, it announced textGrain, an invisible watermark coming to eligible ChatGPT and Codex text in the European Union. The rollout follows EU AI Act transparency rules requiring machine-readable marking of synthetic content. OpenAI detailed the change in its official text provenance announcement.
That does not mean every ChatGPT paragraph can suddenly be exposed by a public checker. OpenAI is restricting its text-watermark detector to approved researchers and expert organisations at launch.
Key Highlights
- textGrain places a statistical signal inside word and token choices; it does not insert hidden spaces, characters or visible labels.
- At a 1% target false-positive rate, OpenAI detected about 80% of 200-token watermarked psychology passages and about 95% of 400-token passages.
- Replacing 10% of words with synonyms cut detection in one 400-token test from about 92% to 66%; replacing 25% dropped it to 17%.
- EU ChatGPT and Codex text will receive the watermark gradually, while API watermarking remains optional globally.
- A detected watermark suggests OpenAI involvement, but it does not identify the user, prompt, ownership or factual accuracy.
What Is The ChatGPT Text Watermark And How Does It Work?
The new system does not stamp “Made by ChatGPT” onto a document. OpenAI says textGrain subtly changes the statistical pattern used when the model chooses among possible words or word pieces. A compatible detector can later search for that pattern.
Copying and pasting does not automatically remove the watermark because there are no secret characters to delete. OpenAI’s provenance guidance specifically says textGrain does not rely on invisible spaces, extra tokens or unusual punctuation.
That differs from many commercial AI-writing detectors, which examine finished prose and estimate whether its style resembles machine-generated text. textGrain adds a provenance signal during generation.
A Simple Detection Example
Take a 400-token watermarked response. In one OpenAI evaluation, detection was around 92% before synonym replacement. After 25% of the words were replaced, it fell to 17%.
- A positive result: an OpenAI system likely generated or processed some of the passage.
- It does not reveal: who used ChatGPT, how much a human edited, or whether the passage is accurate.
Why OpenAI Is Introducing Watermarking Now
Regulation is the immediate driver. The European Commission’s AI transparency rules began applying from 2 August 2026 and require certain generated or manipulated content to carry machine-readable marks. Its transparency framework covers text alongside images, audio and video.
OpenAI says eligible ChatGPT and Codex users across all plans in the EU will receive text watermarking over the coming weeks. It is not becoming the worldwide ChatGPT default at launch.
In 2023, OpenAI’s public AI classifier correctly identified only 26% of AI-written text in one challenge set and incorrectly labelled human writing as AI-generated 9% of the time. The company withdrew that classifier in July 2023 over low accuracy. Those figures remain available in OpenAI’s original AI classifier report.
- Then: a classifier tried to infer whether completed writing looked AI-generated.
- Now: textGrain plants a statistical pattern while the model writes.
That shift could make provenance more useful than stylistic guessing, but it still cannot deliver certainty.
Can Schools, Publishers Or Employers Reliably Detect ChatGPT Text?
Not in every case, and OpenAI’s text detector is not publicly available today. Initial access is limited to approved researchers and expert organisations.
Length changes the odds. OpenAI reported better detection for 400-token passages than 200-token passages. Mathematics and other tightly constrained writing can be harder because the model has fewer wording choices.
Editing is the larger weakness. Paraphrasing, translation and substantial rewriting can reduce the signal. A university or publisher therefore should not treat a failed watermark check as proof that a person wrote the material.
OpenAI also says a watermark cannot identify the account behind a passage, reveal prompts, establish ownership or show how much of a document came from AI assistance.
- For reviewers: provenance can support an investigation, not replace one.
- For writers: editing does not automatically prove or disprove how the original text was produced.
What Changes For ChatGPT Users And AI Content Detection?
For most users outside the EU, there is no automatic change yet. API customers globally can choose watermarking for supported models, but OpenAI is keeping the consumer rollout regional at launch.
Instead of guessing from phrases or sentence style, provenance systems try to carry a machine-readable signal from creation.
Three issues now stand out: broader detector access, whether signals survive editing, and whether providers make different watermarking systems interoperable.
FAQs
Can ChatGPT text now be detected with certainty?
No. textGrain offers a useful provenance signal, but short passages, rewriting, translation and constrained wording can still reduce detection reliability.
Is textGrain on every ChatGPT response worldwide?
No. OpenAI is rolling it out for eligible ChatGPT and Codex text in the EU; global API customers can opt in separately.
Does the watermark reveal a user or prompt?
No. OpenAI says it does not identify a user, account, organisation, prompt or conversation, and it cannot establish ownership.
Can rewriting ChatGPT text weaken detection?
Yes. In one OpenAI test, replacing 25% of words with synonyms reduced detection from about 92% to 17%, showing editing remains a major limitation.
Is textGrain the same as an AI-writing detector?
No. Typical detectors analyse finished writing patterns. textGrain embeds a statistical signal during generation, which a compatible detector later checks.
What’s New Around ChatGPT and Finance?
Why Is ChatGPT Login Not Working?
Check out common sign-in issues and simple ways to fix authentication errors.
Can Employees Face Action Over AI Data?
Find out why officials are warned against sharing taxpayer information with public AI tools.
How Can ChatGPT Enable Fake Documents?
Uncover how generative AI can potentially be misused for impersonation and fraud.
Will ChatGPT Ads Affect Your Answers?
Explore who may see sponsored content and what it means for AI responses.
How Much Can PPF Earn Now?
See the latest PPF interest rate, expected returns, and maturity details.
