OpenAI Introduces Text Watermarking in ChatGPT and Codex
OpenAI is rolling out invisible, machine-readable text watermarking to ChatGPT and Codex, launching first for users in the European Union to comply with regulatory requirements.
Radwan AbdulhadiEditor and technology writer
Regional Rollout in the European Union
OpenAI has announced that an invisible, machine-readable watermark is rolling out to ChatGPT and Codex, but only for users in the European Union at first. According to The Verge, the company is launching its textGrain watermarking system to meet the demands of the EU's AI Act, though it is not yet making the feature a global default.
The regional approach is intended to provide the company with room to learn from real-world use and feedback. Over the coming weeks, the watermarking feature will be introduced to eligible ChatGPT and Codex users across all plans within the EU. Meanwhile, API customers around the world will be able to opt in to watermarked text outputs for select models, as detailed by Engadget.
How the textGrain System Works
The proprietary textGrain watermarking tool works by adding an invisible statistical signal to the model's word choices. OpenAI reports that its approach matched or exceeded other alternative methods, such as Google DeepMind's SynthID for text. Anthropic also adopted a similar watermarking system in August to align with the regulatory mandates.
Article 50 of the EU's AI Act requires providers of generative artificial intelligence systems to ensure that text outputs are identifiable in a machine-readable format. Pre-existing entities must comply with these transparency rules, prompting major tech organizations to implement technical solutions.
Performance and Detection Limitations
While internal benchmark scores showed similar performance between watermarked and unwatermarked text, OpenAI cautions that textGrain does not guarantee reliable detection under all everyday conditions. Shorter texts, mathematical content, and edited text outputs can reduce the success rate of the detection mechanism.
OpenAI explicitly notes that text watermarks do not verify factual accuracy, determine text ownership, measure human contribution, or prove ultimate human authorship. Furthermore, because of the risk of missed watermarks and false positives, the company is not making the detection software publicly available at launch.
Access for Researchers and Developers
To support evaluation and improvement, approved researchers and expert organizations can apply for access to the detection tool. In accordance with the Code of Practice, access is granted initially on a case-by-case basis.
The detection tool is designed solely to report whether an OpenAI watermark is detected, without identifying the user or revealing their underlying prompts or conversations. Additionally, OpenAI is working alongside cloud partners to ensure watermarking becomes accessible for model outputs accessed via their services in the coming weeks.
Sources
- The VergeOpenAI is adding text watermarking in ChatGPT and Codex
- EngadgetOpenAI will add a digital watermark to text and code generated in the EU
Continue chronologically
Related entity coverage
Related stories
OpenAI to Watermark ChatGPT Text in the European Union
OpenAI is rolling out invisible watermarks for ChatGPT and Codex text outputs in the European Union to comply with the new AI Act.
OpenAI Details Text Watermarking Approach for EU AI Act
OpenAI announces its phased text watermarking approach under the EU AI Act, featuring textGrain technology, API opt-ins, and researcher access.
OpenAI Hosts DevDay 2026 With Over 20 Major Announcements
OpenAI highlights more than 20 major announcements at DevDay 2026, introducing new agents, Codex updates, models, and builder tools.
OpenAI Expands Academy with Role-Specific AI Learning Paths
OpenAI has expanded its OpenAI Academy to include specialized learning paths for developers, leaders, educators, and students to build practical AI skills.
