OpenAI Rolls Out Hidden Watermarks in ChatGPT, Starting with EU Users

OpenAI has begun introducing an imperceptible, machine-readable watermarking system called textGrain to its ChatGPT and Codex products. The feature embeds hidden markers into AI-generated text, allowing detection of content produced by OpenAI's models—a move driven largely by regulatory pressure from the European Union.

EU-First Rollout Strategy#

The initial rollout targets users in the European Union only, where the deployment will cover all ChatGPT and Codex users across free and paid plans over the coming weeks. OpenAI deliberately chose not to make watermarking a global standard at launch, framing the EU as a testing ground where the company can evaluate real-world performance and gather user feedback before wider expansion.

How Watermarking Will Work#

For API customers worldwide, OpenAI is offering an opt-in watermarking capability for select models, giving developers control over whether and when to include watermarks in their outputs. Cloud partners will soon gain access to the watermarking feature as well. The company emphasizes that textGrain performs comparably to competing systems like Google DeepMind's SynthID and the approach Anthropic introduced in August—both similarly motivated by EU AI Act compliance.

What Watermarks Cannot Do#

OpenAI is transparent about the technology's limitations. The watermark "does not guarantee reliable detection," and the company clarifies that watermarks do not verify text accuracy, establish ownership, measure human contribution, or prove human authorship. Benchmark testing shows that watermarked and unwatermarked text perform similarly.

Restricted Detector Access#

Rather than releasing detection tools to the general public, OpenAI is granting access only to approved researchers and expert organizations on a case-by-case basis. The detector will confirm whether OpenAI's watermark is present without revealing the user, their prompts, or conversation history. This cautious approach reflects concerns about false positives and missed detections.