OpenAI Introduces Text Watermarking in ChatGPT and Codex, Starting with EU Users
OpenAI has begun rolling out text watermarking technology called 'textGrain' to users of its AI models ChatGPT and Codex within the European Union. The feature is not yet a global default but will be available regionally first as part of OpenAI's compliance efforts with the EU’s AI Act. The company claims that this method performs at least as well as similar approaches from competitors like Google DeepMind and Anthropic. OpenAI also announced that API customers worldwide can now opt in to watermarked outputs for select models, allowing them flexibility regarding transparency requirements. Additionally, approved researchers and expert organizations will have access to the watermark detection tool starting immediately, though it is not being made publicly available due to risks of false positives or missed detections. The textGrain system does not verify accuracy, determine ownership, measure human contribution levels, nor confirm authorship by a person. OpenAI emphasized that while watermarks are part of its transparency strategy, they do not guarantee reliable detection and should be used with caution.
Topics: ChatGPT Codex EU AI Act OpenAI text watermarking