OpenAI Will Introduce Watermarks for ChatGPT Text in the EU

• OpenAI has begun rolling out invisible watermarking technology for its models’ text responses.
• Over the next few weeks, the feature will be added to supported versions of ChatGPT and Codex in the EU.
• API clients worldwide can already opt in to watermarks for specific models.

OpenAI introduced an approach for applying invisible watermarks to text, which the company developed in response to the requirements of the EU AI Act. The technology is intended to help determine whether a text was created or processed by OpenAI systems.

Under EU requirements, providers of generative AI must ensure machine-readable identification of generated content.

OpenAI said that over the next few weeks it will add invisible watermarks to text responses in supported versions of ChatGPT and Codex for users in the European Union.

Globally, the feature will also become available to API customers. They will be able to voluntarily enable watermarking for specific models. By default, the feature will remain disabled.

The system, called textGrain, adds an invisible statistical signal to the model’s word choices. A dedicated detector analyzes the text and determines whether it contains this signal. According to OpenAI, textGrain delivered results on par with or better than other tested approaches, including SynthID for text. At the same time, the company emphasizes that lab test results do not guarantee the same effectiveness in real-world conditions.

In particular, at a 1% false positive rate, the detector found the watermark in about 80% of 200-token excerpts and around 95% of 400-token excerpts in topics such as psychology. For math, the figures were lower due to less freedom in word choice.

Editing also significantly affects the detector’s performance. Replacing 10% of words with synonyms reduced the watermark detection rate from about 92% to 66%. If 25% of words were replaced, the figure dropped to 17%.

The Watermark Does Not Identify the Author of the Text

OpenAI stressed that the technology provides only a limited signal about a text’s provenance. In particular, the verification result does not show:

• how much of the text was contributed by a human
• who owns the text and who is responsible for it
• who exactly used the OpenAI system
• which prompt or conversation was the source of the text
• whether the information in the text is true or false.

At the same time, the absence of a watermark will not mean that a human wrote the text. The reason could be an insufficient text length, editing, translation, the use of a different model, or the material being created before the technology was introduced.

The Detector Will Not Be Publicly Available Yet

OpenAI has opened applications for access to the detector for approved researchers and expert organizations.

The company explained this by the risk of false positive and false negative results. At the initial stage, the tool will be provided on an individual basis to assess its reliability and ways to use it responsibly.

OpenAI also plans to open-source the textGrain technology so other developers can build their own solutions on top of it. According to the company, during testing of the new advanced Astra model, they found no significant difference in benchmark results with and without watermarks.

The company noted that no single method can provide full verification of content provenance. Therefore, OpenAI plans to continue improving watermark detection, including evaluating its robustness against editing and translation.

In the future, the company also wants to better distinguish between AI assistance in writing and full AI authorship.

As a reminder, we previously reported that OpenAI is seeking to raise $30 billion at a $1.4 trillion valuation.

Сообщение OpenAI Will Introduce Watermarks for ChatGPT Text in the EU появились сначала на INCRYPTED.