ChatGPT Text Watermarking: How OpenAI’s New EU System Works

The use of ChatGPT text watermarking is an important topic in the discussion about how transparent AI is and where content comes from. OpenAI has announced that it will begin adding an invisible watermark to some of the text generated by ChatGPT and Codex in the European Union. The move is designed to help make AI-generated content identifiable in accordance with transparency requirements under the EU AI Act.

Unlike a visible logo or special character, OpenAI’s system works by subtly influencing the model’s word choices. The resulting statistical pattern is not visible to readers, but it can be identified by a specialised detector. OpenAI calls this technology textGrain, and will first launch it in the EU.

Key Takeaway:

• OpenAI is introducing ChatGPT text watermarking for eligible ChatGPT and Codex text in the EU.
• The watermark is invisible and is created through patterns in the model’s word choices.
• OpenAI calls the technology textGrain.
• The system is intended to support transparency requirements under the EU AI Act.
• Watermark detection has limitations, especially with short, edited, translated, or highly constrained text.
• A detected watermark does not identify the user, establish ownership, or measure how much a human contributed.

What Is ChatGPT Text Watermarking?

ChatGPT text watermarking is a way of adding an invisible statistical signal to AI-generated text. Instead of adding a visible mark, OpenAI changes the probability of certain word or token choices while the model generates a response.

If you look at a long enough piece of string, you will see that these small changes make a pattern that you can recognise. A detector can then check the text for the watermark.

This is different from just adding a label like “Generated by ChatGPT.” The signal is part of the generation process and is designed to be machine-readable.

OpenAI says its AI text watermarking technology does not add hidden characters, invisible spaces, or unusual punctuation. The watermark is in the statistical pattern of the text’s wording.

Why Is OpenAI Adding a ChatGPT Watermark in the EU?

The main reason is that the rules are clear and easy to understand. The EU AI Act includes rules about identifying AI-generated content. OpenAI says its text watermarking approach is intended to help meet those requirements while providing a machine-readable signal about the origin of generated text. For a broader explanation of these requirements, see our guide to the EU AI Transparency Code. 

OpenAI says that ChatGPT and Codex text that is made in the European Union will have the watermark added over the next few weeks. The company is not going to make text watermarking a global default for ChatGPT when it is first launched. This is part of a bigger change in the industry where companies are working on ways to show if digital content was made by AI or not.

How Does OpenAI’s textGrain Work?

OpenAI calls its watermarking technology textGrain. When a language model generates text, it looks at lots of different possible next words. The watermarking system quietly changes the statistical preferences among those possible choices according to a secret code.

One word choice doesn’t necessarily tell us much. But when this is done over and over again with hundreds of tokens, the choices added together can create a statistical pattern.

A detector can then check the passage for that pattern. This means the OpenAI text watermark is not like a visible watermark on an image. There is nothing for a reader to see or remove manually. Instead, the signal is in the statistical structure of the generated language.

Does ChatGPT Watermarking Change the Text?

The watermark is designed to be invisible to the reader. OpenAI says textGrain does not insert special characters, hidden spaces, or additional watermark-only tokens. Instead, it changes the model’s choices of possible words or word pieces during generation.

OpenAI also says that its tests showed no big differences in how well the models worked when watermarking was switched on. This suggests the goal is to add a provenance signal without noticeably changing the normal generation experience. But the watermark is not always easy to see in all types of text.

How Reliable Is AI Content Detection?

One of the most important things to know about this system is that it doesn’t always detect watermarks perfectly. OpenAI’s tests showed that it’s harder to detect when text is short or when there are fewer opportunities for flexible word choices. Mathematical responses, for example, can be harder to spot than other kinds of writing.

Editing can also make the signal weaker. In one OpenAI test, replacing about 10% of words with synonyms lowered the detection rate from around 92% to 66%. Replacing 25% reduced detection further, to around 17%. This shows an important difference between AI content detection and certainty about authorship.

If a detector finds a watermark, it can show that some of the text was generated or processed by an OpenAI system. But the fact that there is no watermark does not mean that a person wrote the content.

Can Editing Remove the ChatGPT Watermark?

Editing can make the watermark more difficult to detect. Because textGrain is based on how people use words, changing a lot of words can make the results less reliable. Translating it can also make it harder to detect.

This is important because text often changes after it is created. A writer might edit a draft, combine it with their own writing, translate it into another language, or rewrite sections of it a lot before publishing it.

OpenAI does not say that watermark detection is the only way to prove who owns something. Instead, it describes the watermark as a signal that shows where a piece of text was likely generated or processed by an OpenAI system.

What Does an OpenAI Watermark Tell You?

If it works, it can provide useful information, but it’s not clear what it means.

A signal from OpenAI can show that a part of a text has been created or processed using an OpenAI system. It does not prove:

• Who used ChatGPT
• Who wrote the original prompt
• How much a human contributed
• Who owns the content
• Whether the content is accurate
• Whether the use of AI was legally permitted
• Whether AI generated the entire passage

OpenAI says that its watermark does not identify the user or connect the text to an account, prompt, or conversation. This difference is important for schools, workplaces, publishers and other organisations that need to decide how to evaluate AI-generated content.

ChatGPT Watermark vs. Traditional AI Detectors

OpenAI’s watermarking approach is different from many other AI detection tools.

Normal AI detectors look at finished text and try to work out if it seems like it was written by a computer. OpenAI’s text watermark is added during the generation process.

In other words:

• Traditional AI detector: Analyzes text after it has been written.
• ChatGPT text watermark: Creates a statistical signal while the AI generates the text.

OpenAI says the EU AI Act asks for a signal to be put into the text that’s been made, which is one reason why it developed textGrain instead of relying only on classifiers made by other companies.

Who Will Be Affected by ChatGPT Text Watermarking?

The initial rollout is particularly relevant to people using ChatGPT and Codex in the European Union.

Content Creators: Writers and creators using ChatGPT should know that some of the text generated by the AI can contain an invisible signal showing where it came from.

Businesses: Companies using generative AI for marketing, documentation, or customer communications might need to think about how to include AI provenance in their internal policies and compliance processes.

Developers: People who use the OpenAI API can choose to add a watermark to their text for certain models. Businesses and developers using AI APIs are increasingly incorporating AI capabilities into their creative and content workflows. 

Educators and Researchers: Schools and researchers might want to use provenance signals to study AI-generated writing. But a watermark doesn’t automatically prove who made something or how much work went into it.

Does the Watermark Prove That Someone Used ChatGPT?

No.

This is one of the most important things to understand. If a detector identifies an OpenAI watermark, it suggests that an OpenAI system generated or processed some of the text. It doesn’t identify the person who used the system or determine how much of the final work was produced by AI.

Also, if a detector doesn’t find a watermark, that doesn’t prove that the text was written by a person. The passage may be too short, heavily edited, translated, or produced by another AI system. It may also have been generated before watermarking was available. So, we should think of AI-generated content detection as one piece of evidence, not as a perfect way to determine authorship.

What Does This Mean for AI Transparency?

The new feature of adding a watermark to ChatGPT text shows a bigger change toward making AI technology more clear and understandable. As generative AI becomes a part of everyday writing, businesses and regulators need better ways to tell the difference between content created by traditional software, humans, and AI systems. This is closely connected to the growing focus on AI-generated content identification under Europe’s AI transparency framework. 

Watermarking is one way to solve this problem. OpenAI already uses different provenance technologies for other types of content. Its current framework includes Content Credentials and SynthID for supported images and SynthID for audio, while text uses textGrain in the EU.

However, OpenAI admits that no provenance technology is perfect. Editing, transformation, language, content length, and other factors can affect whether a signal remains detectable.

What Happens Next?

OpenAI says that they will start adding text watermarks to ChatGPT and Codex in the EU over the next few weeks. The company is also making text watermarking available as an opt-in feature to API customers worldwide for supported models. OpenAI is also letting approved researchers and expert organizations use a text watermark detector.

OpenAI also says it plans to make textGrain available as open-source technology, which could allow researchers and developers to study and improve the approach.

As standards and detection technology improve, AI watermarks might be used for more than just making sure people follow the rules. They could be used in different ways to show where content comes from.

Conclusion

ChatGPT text watermarking is a big step forward in making AI-generated content easier to identify. OpenAI’s textGrain system creates an invisible statistical signal through the model’s word choices. Specialized detection tools can look for evidence that an OpenAI system generated or processed text.

For users in the EU, the change is closely connected to the EU AI Act’s transparency requirements. At the same time, OpenAI’s own testing shows that watermark detection has its limits. Short passages, tight writing, translation, and editing can all make detection less reliable.

In the end, the OpenAI watermark is best understood as a provenance signal, not a definitive authorship test. As AI-generated content becomes more common, technologies such as textGrain could become an increasingly important part of how platforms, businesses, researchers, and regulators approach making AI transparent.

FAQs

1. What is ChatGPT text watermarking?

ChatGPT text watermarking is OpenAI’s method of adding an invisible statistical signal to AI-generated text through patterns in the model’s word choices.

2. What is OpenAI textGrain?

textGrain is OpenAI’s text watermarking technology. It subtly influences the model’s token choices so that a detector can look for a statistical watermark in the resulting text.

3. Does ChatGPT watermark all AI-generated text?

No. OpenAI’s initial rollout applies to eligible ChatGPT and Codex text generated in the European Union. API customers worldwide can separately opt into watermarking for supported models.

4. Can an AI watermark prove that a person used ChatGPT?

No. A detected watermark can indicate that an OpenAI system generated or processed some text, but it does not identify the user or establish how much human contribution was involved.

5. Can editing remove an AI text watermark?

Editing can weaken the watermark and make detection less reliable. OpenAI’s testing found that replacing words with synonyms can substantially reduce detection rates.