OpenAI Introduces Text Watermarking in ChatGPT and Codex
The company is rolling out textGrain watermarking in the EU—but stops short of making it a global default.
OpenAI has begun embedding invisible watermarks into text generated by its ChatGPT and Codex models, marking a significant step in the company's approach to AI content transparency. The feature, called textGrain, is initially launching only within the European Union, where regulatory pressure has intensified following the full enforcement of the EU AI Act's transparency provisions in 2025.
According to OpenAI, the watermarking system will not yet be enabled as a "global default," leaving users in other regions—including the United States and Asia—without the same provenance signals for the time being. The company has not provided a timeline for broader rollout.
What Is textGrain?
textGrain is OpenAI's proprietary watermarking technique for text. Unlike visible disclaimers or metadata tags, textGrain subtly alters word choices and sentence structures in a statistically detectable pattern. The watermark is imperceptible to human readers but can be identified by compatible detection tools.
This approach mirrors earlier watermarking efforts for AI-generated images and audio, but applying it to text has proven far more challenging. Text offers a narrower information space for encoding hidden signals without degrading readability or model performance.
Why the EU First?
The European Union's AI Act, which came into full force in August 2025, requires providers of general-purpose AI systems to mark and disclose AI-generated content. Article 50 of the regulation specifically mandates that outputs be "marked in a machine-readable format and detectable as artificially generated or manipulated."
By launching textGrain in the EU first, OpenAI is signaling compliance with these rules while carefully managing the technical and commercial risks of a worldwide deployment.
Not a Global Default—Yet
OpenAI's decision to limit textGrain to the EU reflects ongoing concerns:
- Detection accuracy: Watermarks can be removed through paraphrasing, translation, or adversarial editing.
- User experience: Writers and developers may resist subtle alterations to generated text.
- Competitive pressure: Rival models like Anthropic's Claude and Google's Gemini have not yet implemented mandatory text watermarking at scale.
An OpenAI spokesperson said the company is "evaluating feedback and performance data" before deciding whether to expand textGrain globally.
What This Means for 2026
As AI-generated content floods the web, provenance and authenticity have become central policy battlegrounds. The EU's lead on watermarking mandates is pushing major AI labs to act, but a fragmented global landscape is emerging:
- EU: Strict transparency requirements, active enforcement
- US: Patchwork of state laws, no federal watermarking mandate
- China: Deep synthesis regulations requiring labels on AI content
- Global South: Largely unregulated, raising concerns about misinformation
OpenAI's EU-first strategy may become a template for other AI companies—comply where required, wait and see elsewhere.
The Bottom Line
textGrain represents a meaningful technical achievement, but its limited rollout underscores a broader tension: the AI industry still lacks a unified, global standard for marking machine-generated text. For now, if you're using ChatGPT or Codex in the EU, your output may carry a hidden signature. Everywhere else, it remains business as usual.
via The Verge AI