Text Watermarking

OpenAI is bringing text watermarking to ChatGPT and Codex, describing the move as part of its strategy for meeting transparency expectations in the European Union. The EU AI Act includes requirements around identifying certain AI-generated content, making it increasingly important for providers to show how synthetic material can be recognized. Watermarking is one possible way to help distinguish model-generated text from human-written content, complementing other approaches such as disclosure labels and provenance information.

Unlike a visible label, a text watermark may be designed to be detected through analysis rather than noticed by readers. That makes the approach potentially useful across different kinds of generated content, including conversational responses and coding assistance, but it also raises practical questions. Detection needs to work reliably across varied writing styles and languages, while edits, copying, or paraphrasing may affect whether a signal remains recognizable. The articles focus on OpenAI’s rollout and the reasoning behind its EU-facing strategy, rather than establishing how well watermarking performs in every use case. For users and developers, the change signals a shift toward making AI-generated text easier to identify; for the broader industry, it reflects the challenge of translating regulatory transparency goals into technical safeguards that can function at scale.