OpenAI Introduces Text Watermarking
An invisible, machine-readable watermark in text output is rolling out to ChatGPT and Codex, but only for users in the European Union at first.
OpenAI is adding text watermarking in ChatGPT and Codex
OpenAI is launching textGrain watermarking in the EU, but states it will not be the 'global default' at this initial stage.
Industry Context and Regulatory Compliance
OpenAI claims its textGrain watermarking technology 'matched or exceeded' other approaches, including Google DeepMind’s SynthID for text, which also underpins Anthropic's recently announced watermarking. Like Anthropic, OpenAI's move is intended to meet the requirements of the EU’s AI Act.
While OpenAI provided AI benchmark scores indicating similar performance for both watermarked and unwatermarked text, the company notes that textGrain 'does not guarantee reliable detection.' It also clarifies that text watermarks do not verify accuracy, determine text ownership, measure human contribution, or definitively prove human authorship.
Phased Rollout and Access
Over the coming weeks, text watermarking will be introduced to eligible ChatGPT and Codex users across all plans exclusively in the EU. This regional approach allows OpenAI to gather real-world use data and feedback before considering a global default. Additionally, API customers worldwide can now opt-in to watermarked text outputs for select models, enabling them to align with their own transparency obligations. OpenAI is also collaborating with cloud partners to extend watermarking availability through their services. Approved researchers and expert organizations can apply for detector access, which will be granted on a case-by-case basis to support evaluation and improvement of text provenance. The detection tool will indicate the presence of an OpenAI watermark without identifying the user or revealing prompts, and will not be publicly available at launch due to the risk of missed watermarks and false positives.
