OpenAI is preparing to introduce invisible watermarks for text generated by ChatGPT and Codex in the European Union. The feature is designed to help identify AI-generated content without changing how the text looks to users.

The company says the watermark will not be visible when users read, copy or paste the generated text. Instead, its new textGrain technology makes small changes to word choices. These changes create a statistical pattern that can later be identified by a dedicated detection system.

OpenAI to Add Invisible Watermarks to ChatGPT and Codex Text in Europe

OpenAI says the watermark will be added to eligible ChatGPT and Codex outputs in the EU over the coming weeks. The company has not announced plans to make the feature a default worldwide.

API Developers Can Opt In

While the EU rollout will apply to eligible ChatGPT and Codex outputs, OpenAI is also giving API developers around the world an option to use watermarking on supported models.

The feature will remain disabled by default for API users, meaning developers will need to choose to enable it.

OpenAI is also accepting applications for access to its watermark detection system. Initially, the detector will only be available to selected researchers and expert organisations.

Editing Can Reduce Watermark Detection

OpenAI acknowledges that text watermarking has limitations. Its testing shows that relatively small changes to AI-generated text can make the watermark much harder to detect.

In one evaluation involving 400-token passages, replacing just 10% of the words with synonyms reduced detection accuracy from around 92% to 66%. When 25% of the words were replaced, detection dropped to only 17%.

The length of the text also affects detection. At a 1% false-positive rate, OpenAI detected watermarks in about 80% of 200-token psychology responses. The rate increased to around 95% for responses containing 400 tokens.

OpenAI also found that detection can vary depending on the subject. Mathematics, for example, can be more difficult because there is often less flexibility in choosing alternative words.

A Detected Watermark Will Not Reveal the User

OpenAI says the watermark is intended to identify whether text was generated by its AI systems, but it does not provide information about the person who created it.

A successful detection cannot reveal the user’s account, prompt or conversation. It also cannot determine how much of the final text was produced by AI or how much was written or edited by a person.

The company warns that failing to detect a watermark does not prove that a person wrote the text. Short content, edited material, and translated text may not retain enough of the watermark pattern for reliable detection.

Little Impact on Model Quality

OpenAI says the watermarking system has been designed to preserve the quality of generated text. According to the company’s testing, enabling textGrain does not meaningfully affect the performance of its GPT-6 Astra model, with benchmark results remaining broadly similar when the technology is active.

The EU rollout could give researchers, educators and organisations a new way to identify AI-generated content. However, OpenAI’s own testing highlights that the technology is not foolproof, particularly when users make changes to generated text.