Summary
OpenAI is preparing to add invisible watermarks to text generated by ChatGPT and Codex for users in the European Union. The move is aimed at helping the company comply with transparency requirements under the EU AI Act.
Unlike a visible label or symbol, the watermark will be embedded directly into the generated text. OpenAI says specialized detection systems will be able to identify the pattern, while the text itself will look normal to readers.
What Happened?
OpenAI announced that it will begin rolling out text watermarking for eligible ChatGPT and Codex users in the EU over the coming weeks.
The company says the feature will initially be limited to the European Union rather than becoming a global default.
Developers using OpenAI’s API outside the EU will also have the option to enable the technology for selected models. However, the feature will be turned off by default.
The change comes as AI companies face increasing pressure from regulators to make it easier to identify content generated by artificial intelligence.
Why Is OpenAI Adding Watermarks?
The move is connected to the transparency requirements of the EU AI Act.
The rules require providers of certain AI systems to ensure that AI-generated content can be identified by systems capable of detecting it.
OpenAI says its new watermarking system is designed to meet this requirement without adding visible markings to the text.
The company has not said that the technology will initially become a worldwide default. Instead, the first major rollout will focus on users in the European Union.
How Does the Invisible Watermark Work?
OpenAI’s watermark does not work like a traditional image watermark.
Instead, it subtly influences the words selected by the AI while generating a response. Those small changes create a statistical pattern inside the text.
A person reading the content would not normally notice anything unusual. However, a specialized detector with the appropriate key could analyze the text and determine whether the pattern associated with OpenAI’s watermark is present.
Because the watermark is embedded within the words themselves, OpenAI says it can remain with the content when someone copies and pastes the text.
The company also says the watermark does not contain information that identifies the individual user.
OpenAI Publishes Technical Details
Alongside the announcement, OpenAI published a technical report describing its watermarking technology, called textGrain.
The research was conducted with researchers from the University of Pennsylvania and Yale.
At a high level, the system influences the model’s word-selection process. When repeated across a sufficiently long piece of generated text, these small changes create a detectable pattern.
OpenAI says its testing showed no meaningful impact on model performance when the watermark was enabled.
However, the company is taking a cautious approach to access to the detection technology.
Initially, detector access will be provided only to approved researchers and expert organizations. OpenAI says this will allow it to evaluate the system’s reliability and potential uses before making it more broadly available.
A Watermark Does Not Prove Human or AI Authorship
OpenAI is also warning users not to treat the watermark detector as a perfect test for determining who wrote a piece of content.
If a watermark cannot be detected, that does not necessarily mean the text was written entirely by a human.
Short pieces of text may not contain enough information for reliable detection. Heavy editing can also weaken the watermark, while translated content and certain types of responses, such as mathematical answers, can be more difficult to identify.
OpenAI’s own testing showed how easily editing can affect detection.
In one test, replacing around 10% of the words with synonyms reduced detection from approximately 92% to 66%.
That means watermarking could be useful as an indicator, but it is unlikely to provide absolute proof of AI authorship.
What About Human Editing?
The technology also raises questions about how AI-assisted writing should be classified.
Someone may use ChatGPT to generate an initial draft and then significantly rewrite, edit or expand it themselves. A watermark could indicate that OpenAI technology was involved in producing part of the text, but it cannot determine how much human work went into the final version.
OpenAI acknowledges this limitation and says its watermark should not be interpreted as a measurement of human creativity or judgment.
OpenAI Is Not the Only AI Company Using Watermarks
OpenAI’s decision follows similar efforts from other major AI companies.
Anthropic recently announced that it would watermark text generated by Claude, although its approach is being applied more broadly.
The issue has already generated debate among some AI users, particularly people who use AI as an assistive writing tool rather than simply copying AI-generated responses.
The broader AI industry is also moving toward common standards for identifying AI-generated content. OpenAI, Anthropic, Google, Meta and Microsoft are among the companies that have committed to following the EU’s code of practice covering AI-generated content.
OpenAI Had Previously Held Back Its Watermark
This is not the first time OpenAI has developed technology for watermarking AI-generated text.
The company had previously created a similar system but decided not to release it publicly. Concerns reportedly included the possibility that users could move to competing AI services that did not use watermarks.
The new EU regulations appear to have changed the situation, at least for users in the region.
What Comes Next?
OpenAI plans to introduce the watermark to eligible ChatGPT and Codex users in the EU over the coming weeks.
The company will also allow developers using its API to enable the feature for selected models, although it will initially remain optional.
The effectiveness of text watermarking will likely depend on how reliably detectors can identify the embedded patterns and how well those patterns survive editing, translation and other changes.
For now, OpenAI is treating the technology as an additional signal rather than a definitive way to determine whether a person or an AI system created a piece of text.
Conclusion
OpenAI is introducing invisible watermarks for ChatGPT and Codex-generated text in the European Union as AI transparency rules take effect.
The technology works by subtly changing word-selection patterns rather than adding visible labels. While this could make AI-generated text easier to identify, OpenAI’s own testing shows that editing and other transformations can reduce detection accuracy.
The move marks another step toward a future where AI-generated content can be identified more easily, but it also highlights the difficulty of drawing a clear line between human writing and AI-assisted work.