OpenAI Begins Rolling Out Text Watermarking for ChatGPT and Codex in the EU
OpenAI has announced the launch of a new text watermarking system for ChatGPT and Codex, starting with users in the European Union. The feature adds an invisible, machine-readable marker to AI-generated text, making it easier to identify content created by OpenAI's models.
The company said its watermarking technology, called textGrain, performed as well as or better than similar systems developed by other AI companies, including Google DeepMind’s SynthID. Anthropic recently introduced a comparable feature as part of efforts to comply with the European Union’s AI regulations.
Despite the rollout, OpenAI emphasized that watermarking is not being introduced worldwide for now. The company plans to launch the feature only in the EU initially, allowing it to gather feedback and evaluate how the system performs in real-world situations before considering a broader release.
OpenAI also noted that the watermark does not guarantee perfect detection. It cannot confirm whether information is accurate, prove who owns a piece of content, determine how much a human contributed to it, or verify that a person wrote the text without AI assistance.
In addition to the EU rollout, OpenAI is giving API customers around the world the option to enable watermarking for selected AI models. This allows businesses and developers to decide whether the feature fits their transparency and compliance needs.
The company is also working with cloud service providers to make watermarking available through platforms that offer access to OpenAI models.
To support research and improve the technology, OpenAI will provide a watermark detection tool to approved researchers and expert organizations. The tool will only indicate whether an OpenAI watermark is present in a piece of text and will not reveal user identities, prompts, or conversation details.
However, OpenAI said the detector will not be publicly available at launch because watermarking systems can sometimes miss watermarks or mistakenly flag content, which could lead to inaccurate result.
0 Comments