OpenAI introduces text watermarking for ChatGPT and Codex in Europe

OpenAI is launching a new feature called textGrain watermarking for its AI models, ChatGPT and Codex. Initially, this feature will be available only to users within the European Union. The watermark is invisible and machine-readable, embedded within the text output of these models. According to OpenAI, their watermark matches or exceeds the effectiveness of similar technologies, such as Google DeepMind's SynthID, which serves as a base for other companies' watermarking solutions.

This move aligns with the requirements set by the EU’s AI Act, although OpenAI clarifies that this will not become a global default immediately. The purpose of this phased regional rollout is to gather feedback and observe real-world usage before considering broader implementation.

Limited Availability and Functionality

Despite its capabilities, textGrain does not ensure reliable detection of all watermarks. It also does not verify the accuracy of the text, determine ownership, assess human contribution, or authenticate authorship. For now, OpenAI is offering opt-in watermarking for API customers worldwide on select models, allowing users to meet transparency obligations if they choose.

For product developers and small software teams in the EU, it means evaluating how this watermark might affect user experience and data handling practices. Access to the detector is limited to approved researchers and expert organizations, meaning practical use cases are still emerging and need careful assessment before integration into existing workflows.

Source: AI | The Verge