OpenAI To Watermark ChatGPT Text In EU, Offers API Opt-In

OpenAI announced on March 18, 2024, that it will begin implementing an invisible watermark on text generated by its ChatGPT and Codex models for users within the European Union. This measure is intended to help identify AI-generated content, aligning with evolving regulatory landscapes concerning artificial intelligence. The company stated that initial tests indicate that while editing can weaken the watermark, it remains detectable. This initiative comes as the EU continues to develop and implement comprehensive AI regulations, such as the AI Act, which aims to foster responsible AI development and deployment.

In addition to the EU rollout, OpenAI is also offering an opt-in feature for its API users to enable this watermarking technology. This allows developers and businesses integrating OpenAI's models into their applications to leverage the watermarking capabilities. The decision to offer an opt-in rather than a mandatory implementation for API users provides flexibility for developers who may have specific use cases or concerns about the impact of watermarking on their services. The specifics of the API opt-in process and its technical implementation were not detailed in the initial announcement but are expected to be made available to developers.

The introduction of watermarking is a significant step in OpenAI's efforts to address concerns about the misuse of AI-generated content, including the spread of misinformation and the potential for academic dishonesty. By making AI-generated text traceable, OpenAI aims to increase transparency and accountability. The company has been actively engaged in discussions with policymakers and researchers regarding the ethical implications of AI and has committed to developing safeguards to mitigate potential risks. The effectiveness of the watermark against sophisticated editing techniques will likely be a subject of ongoing evaluation and development.

Codex, a model developed by OpenAI, is specifically designed for code generation and understanding, and its inclusion in the watermarking initiative suggests a broad application of the technology across different AI functionalities. The move to watermark text in the EU also reflects a proactive approach by OpenAI to comply with or anticipate regulatory requirements in key global markets. The company has not specified a definitive timeline for the full deployment of the watermark within the EU, but the announcement signals an imminent start to the process. Further details regarding the technical aspects of the watermark and its detection methods are anticipated to be released by OpenAI.