OpenAI Introduces Text Watermarking for ChatGPT and Codex in EU

Innovative Watermarking Technology

OpenAI has announced the rollout of its text watermarking feature, known as textGrain, specifically tailored for users in the European Union. This invisible, machine-readable watermark will initially be available only to a select group of customers using ChatGPT and Codex.

Comparative Performance

According to OpenAI, the textGrain watermarking technology performs comparably or even superior to existing solutions, such as Google DeepMind’s SynthID. Both initiatives aim to comply with the regulatory requirements set forth in the EU’s AI Act.

While some users welcomed the news, others expressed dissatisfaction regarding the implementation of this feature. To address concerns, OpenAI provided benchmark scores demonstrating that the performance of watermarked text is on par with unwatermarked counterparts.

Key Features and Limitations

Despite its advantages, OpenAI cautioned that textGrain does not guarantee absolute detection reliability. Users should note that watermarked text may not reflect the accuracy of the content, indicate ownership, measure human input effectively, or serve as proof of authorship.

OpenAI states, “We are introducing text watermarking for eligible users across all plans in the EU only. This localized rollout allows us to gather insights from user experiences and feedback before expanding globally.”

Opt-In Options for API Customers

In addition to the EU rollout, OpenAI is providing global API customers the choice to opt-in for watermarked text outputs for specific models. This flexibility enables businesses to align the use of watermarks with their transparency commitments and user experience considerations.

Collaboration with Researchers

OpenAI is also looking to collaborate with researchers and expert organizations by granting them access to a detection tool. Interested parties can apply for access, which will be granted on a case-by-case basis to facilitate the evaluation and improvement of text provenance.

This detection tool aims to inform whether an OpenAI watermark is present, ensuring that user identities or their corresponding prompts and conversations remain confidential. Due to concerns about potential inaccuracies, including missed watermarks or false positives, the tool will not be publicly available at the feature's launch.

Overall, OpenAI's initiative is a significant step forward in promoting transparency within AI-generated content and aligning with regulatory requirements in Europe.