OpenAI textGrain Adds Invisible Watermark to ChatGPT and Codex Text in EU

• OpenAI textGrain adds an invisible watermark to ChatGPT and Codex text for eligible EU users.
• The watermark works through statistical word-choice patterns and is designed to survive copying and light edits.
• Detection performance falls sharply after synonym replacement, while short text, code and heavy edits remain harder.

OpenAI textGrain is a new invisible watermark designed to identify text generated by ChatGPT and Codex. It will roll out over the coming weeks to eligible users in the European Union across all ChatGPT and Codex plans.

The system is not being launched as a global default for users. However, developers using the OpenAI API anywhere in the world can enable textGrain for selected models, although the feature is off by default. MobileTelco has also covered other recent OpenAI and ChatGPT developments as the company expands its AI products.

What Is OpenAI textGrain?

OpenAI textGrain is an invisible watermarking system for AI-generated text. It subtly changes how the model randomly chooses between possible words or word pieces while generating a passage.

These choices create a statistical pattern across the text. A detector can then analyse the passage to check whether the OpenAI watermark pattern is present.

The watermark is not added through hidden characters, invisible spaces, unusual punctuation or special watermark-only tokens. Readers therefore cannot see it in the generated text.

Why OpenAI Is Introducing the Watermark

The rollout is linked to European Union AI transparency requirements. The EU AI Act requires AI providers to mark generated text in a machine-readable way.

OpenAI has also committed to the EU Code of Practice on Transparency of AI-Generated Content. The initial consumer rollout of textGrain will therefore be limited to eligible users in the European Union rather than becoming a global default.

How the ChatGPT Invisible Watermark Works

When generating text, the model can have several possible words or word pieces available at different points. textGrain subtly influences the random selection among those possibilities.

One individual word choice does not reveal the watermark. Instead, the choices accumulate across a passage and create a statistical pattern that a detector can search for.

Copying and pasting the text does not remove the watermark, and the system is designed to survive light edits.

OpenAI textGrain Detection Results

OpenAI tested the system at a 1 percent false-positive target. In tests involving content such as psychology answers, around 80 percent of passages of roughly 200 tokens were detected, while detection reached about 95 percent for passages of roughly 400 tokens.

However, the results change significantly when the text is modified. Replacing 10 percent of the words with synonyms reduced detection in 400-token text from about 92 percent to 66 percent. Replacing 25 percent of the words reduced detection to about 17 percent.

Limits of the AI-Generated Text Watermark

OpenAI textGrain is not equally effective across every type of content or every editing method. Short passages, maths answers, code, heavily edited text, translations and text generated by other AI models are harder to detect.

The EU code does not require watermarks for outputs under 200 tokens, which is about 150 English words, or for code. These limitations mean a detector result cannot be treated as a universal test for whether any piece of text was written by AI.

What the textGrain Detector Can and Cannot Show

The detector is designed to indicate whether an OpenAI watermark was found in the text. It does not identify who wrote the passage, which prompt was used or how much a person edited the content.

A missing watermark also does not prove that a person wrote the text. This is particularly important for short content, heavily edited material, code, translations and content produced by other AI models.

Readers can learn more about how OpenAI approaches content provenance through the OpenAI’s provenance signals help page.

OpenAI textGrain Availability

For ChatGPT and Codex, textGrain will roll out over the coming weeks only to eligible users in the European Union across all plans. It will not be a global default at launch.

Developers using the OpenAI API can turn on textGrain anywhere in the world for selected models. The API feature is off by default.

Access to the detector itself is currently limited to approved researchers and qualified organisations.

Impact on ChatGPT and Codex Performance

OpenAI says watermarking had no meaningful effect on model quality in its benchmarks. It also reported a negligible effect on speed.

This allows the watermark to operate through the model’s word-selection process rather than by adding visible or hidden formatting to the generated passage.

The change comes as ChatGPT continues to receive updates across different features, including recent changes to ChatGPT image generation limits.

Conclusion

OpenAI textGrain introduces an invisible statistical watermark for text generated by ChatGPT and Codex, with the initial user rollout limited to eligible users in the European Union. Developers worldwide can enable it for selected API models, while detection remains limited and can become less reliable after substantial text changes.