OpenAI adds textGrain watermarks to ChatGPT and Codex in the EU

OpenAI began rolling out invisible, machine-readable text watermarks to eligible ChatGPT and Codex users across all plans in the European Union only, over the coming weeks, starting on 5 October 2026. The company said it is not making text watermarking a global default at launch. API customers around the world can opt in to watermarked text outputs for select models starting on 5 October 2026.theverge.com

The rollout is designed to meet Article 50 of the EU's AI Act, which mandates that providers of generative AI systems make text outputs identifiable in a machine-readable way. Pre-existing entities including OpenAI, Anthropic, Microsoft, Google and Meta have until 2 December to comply, while the requirement already applies to new companies entering the market. OpenAI's regional approach gives it room to learn from real-world use and feedback.

What the rollout covers

OpenAI is using a proprietary system called textGrain that adds an invisible statistical signal to the model's word choices. Once added, a detector can spot these signals and ascertain if there is a watermark. OpenAI says textGrain matched or exceeded other approaches like Google DeepMind's SynthID for text, which is also the basis for the watermarking Anthropic announced in August.

OpenAI included scores from AI benchmarks showing similar performance from watermarked and unwatermarked text. The company also plans to make the technology available in open source so others can tweak it. Watermarking will not be enabled by default except for eligible ChatGPT and Codex text output in the European Union, and API customers can opt in for select models.

How textGrain marks output

OpenAI notes that textGrain does not guarantee reliable detection, and says strong performance under ideal conditions does not guarantee reliable detection in everyday use. The company states that text watermarks do not verify accuracy, determine who owns the text, measure how much a human contributed, or prove human authorship. Shorter text is harder to detect, with around an 80 percent success rate, and content such as mathematics is also harder to detect. Editing text brings down the success rate. OpenAI is being more cautious with the detection software: access will initially be limited to approved researchers and expert organizations, who can apply starting on 5 October 2026.

Approved researchers and expert organizations can apply for detector access starting on 5 October 2026. In accordance with the Code of Practice, access will be initially granted on a case-by-case basis to support evaluation and improvement of text provenance. The tool will report whether it detects an OpenAI watermark, without identifying the user or revealing their prompts or conversations. Given the risk of missed watermarks and false positives, OpenAI is not making it publicly available at launch.

Detection limits OpenAI disclosed

Anthropic made a similar move earlier in 2026 to meet the requirements of the EU's AI Act, and OpenAI's announcement follows that approach. OpenAI says its watermarking matched or exceeded SynthID for text, the system underlying Anthropic's watermarking announced in August. Not everyone was happy to learn about the addition, and questions remain over how well the technology works with code.

The December compliance deadline

OpenAI is also working with cloud partners to make watermarking available for OpenAI model outputs accessed through their services in the coming weeks. The company says API customers can decide how watermarking fits their transparency obligations and the experiences they provide to users. OpenAI plans to make the textGrain technology available in open source. Pre-existing entities including OpenAI, Anthropic, Microsoft, Google and Meta have until 2 December to comply with Article 50 of the EU's AI Act, which requires machine-readable identification of generative AI text output. OpenAI's EU rollout and global API opt-in both begin on 5 October 2026, ahead of that deadline.