OpenAI Introduces Watermarking System for EU ChatGPT and Codex Services

OpenAI has announced plans to introduce a watermarking system for its ChatGPT and Codex services in the European Union. This move is aimed at complying with the EU AI Act's transparency rules, which require AI companies to clearly identify AI-generated content.

The EU AI Act came into effect on August 2, mandating that AI companies mark their generated content in a way that can be easily detected by other systems. To meet this requirement, OpenAI will introduce an invisible watermark for text produced by its ChatGPT and Codex services within the EU region.

OpenAI has stated that the watermarking system will roll out over the coming weeks to eligible users on all plans within the EU. However, developers using OpenAI's API worldwide can opt-in to use the feature starting today, with a default setting of off. The company has explicitly mentioned that it does not plan to make text watermarking the global default at launch.

The watermark itself is not an actual visual marker but rather a subtle pattern created by shaping word choices within the generated text. This means the watermark is embedded in the words and remains present even when the text is copied or pasted. OpenAI has assured that the watermark does not compromise user anonymity, as it doesn't identify individuals.

A technical report detailing OpenAI's method, called textGrain, was also published alongside the announcement. Co-authored with researchers from Yale and the University of Pennsylvania, the document explains how a secret key can be used to subtly influence word predictions, allowing for the detection of AI-generated content using only the provided key and the text itself.

OpenAI has found that its watermarking system can be compromised by editing text generated by ChatGPT. In one test, replacing 10% of words with synonyms dropped detection rates from 92% to 66%. This suggests that the watermark can be removed or obscured if sufficient effort is made to manipulate the text.

The company acknowledges these limitations and has decided to provide access to its detector technology only to approved researchers and expert organizations. These groups will help evaluate the reliability and responsible use of the detector, which is intended to identify AI-generated content in the EU.

OpenAI also warns that a missing watermark does not necessarily prove human authorship. Text may be too short or heavily edited, or it could have originated from another company's AI system. The watermark only indicates that an OpenAI system was involved in generating or processing part of the passage.

The announcement comes two months after rival Anthropic began watermarking text generated by its Claude system worldwide. This decision sparked backlash from some users who felt they had contributed significantly to the creation of the content, while Claude merely provided a tool for their work.

OpenAI's delay in releasing its own watermarking technology was partly due to concerns that users might switch to unwatermarked rivals if they knew their text could be detected as AI-generated.

OpenAI has announced that it will begin watermarking the text generated by its ChatGPT chatbot in the European Union.

The move is likely aimed at addressing concerns about AI-generated content being used for malicious purposes, such as spreading misinformation or propaganda.

OpenAI's decision to start watermarking ChatGPT's text in the EU comes after a federal judge in the US described another AI-powered tool, Flock, as "indiscriminate mass surveillance".

The development is part of OpenAI's broader efforts to ensure that its technology is used responsibly and for beneficial purposes.