OpenAI's textGrain Watermark for the EU: What It Does and Its Limits
OpenAI will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks. OpenAI The system is proprietary and called textGrain. Engadget It covers text and code.
The rollout is tied to the EU's AI transparency rules. Article 50 of the EU AI Act requires providers of generative AI systems to make text outputs identifiable in a machine-readable way, in other words a signal software can read even though readers cannot see it. The obligation applies from 2 August 2026. European Commission Pre-existing providers have until 2 December to comply. OpenAI states it supports the EU Code of Practice on AI content transparency, which supports compliance with marking and labelling obligations. Its newsroom lists the update as Our approach to EU text provenance rules, dated 5 October 2026.
textGrain adds a hidden statistical signal to the model's word choices. You can think of it as a faint pattern in which words get picked, not a visible stamp. A detector can test for that pattern to check if the watermark is present. The detector does not identify the user or reveal prompts or conversations. Access to the detection software will initially be limited to approved researchers and expert organizations.
OpenAI says textGrain matched or exceeded the performance of other approaches such as SynthID for text. The company cautions that strong performance under ideal conditions does not guarantee reliable detection in everyday use. Detection has around an 80 percent success rate for shorter text. Content such as mathematics is harder for the detector to detect. Editing watermarked text reduces the detection success rate.
The watermark will not be enabled by default except for eligible ChatGPT and Codex text output in the European Union. Customers will be able to opt in to watermarking for select models. OpenAI plans to make the watermarking technology available in open source, meaning the code can be inspected and reused.
Anthropic made a similar move earlier this year watermarking text generated by Claude. OpenAI already uses SynthID watermarks and Content Credentials (C2PA) as provenance signals to help people understand whether content was generated with OpenAI. Seven companies including OpenAI and Google pledged in 2023 to develop a system to watermark AI-generated text, images, audio and video. Reuters The European Commission provides a set of icons that deployers of generative AI systems can use to label certain AI-generated content. From 2 August 2026, the EU AI Act will require clear labelling of AI-generated content in key cases. Under the transparency rules, deployers cannot simply rely on the machine-readable marking embedded in the content.
For teams building or operating on top of these models, the limits are as important as the mechanism. Detection is probabilistic. Short outputs, formal notation, and routine edits all degrade it. That caveat matters for any workflow that would treat a detector score as ground truth for provenance, academic integrity, or incident response. An opt-in model for select models also means coverage will be uneven at first, and gated detector access keeps independent testing narrow while the system matures.
In my view, the structure of the rollout reflects that reality. Machine-readable marking at the provider layer, human-visible labelling at the deployer layer, and gated detectors for researchers create separate controls instead of a single point of failure. Open sourcing textGrain could help independent evaluation and interoperability across detectors. The long arc here is constructive. Persistent, testable provenance signals give platforms, enterprises, and regulators something concrete to build on, even if no watermark survives determined rewriting.
