OpenAI Brings textGrain Watermarking to ChatGPT and Codex

OpenAI has begun rolling out textGrain watermarking across ChatGPT and Codex, embedding a hidden signal in AI-generated text so it can later be identified as machine-written. The move marks one of the most significant attempts yet by a major AI lab to make its outputs traceable at scale.

According to OpenAI, the watermark is designed to be imperceptible to readers while remaining detectable by dedicated tooling. It applies to text produced by both consumer-facing ChatGPT and Codex, the company's coding assistant, extending coverage beyond chat responses to developer workflows.

Key details of the rollout include:

• Scope: ChatGPT and Codex outputs are both covered.
• Detection: The watermark is intended to be invisible to human readers but recoverable with the right tools.
• Availability: The feature is not launching globally at once.

The company has not positioned textGrain as a universal solution. Its initial reach is limited, and broader availability will depend on how the first phase performs in practice.

The EU Is the First Launch Market

The rollout of textGrain watermarking starts in the European Union, making the bloc the first market to receive the feature. That sequencing is not accidental: the EU has moved fastest on rules that require AI-generated content to be identifiable, and launching there first lets the company test the system under the closest regulatory scrutiny it is likely to face.

For users in EU member states, the practical effect is that watermarked output may appear before it reaches anyone else. Everyone outside that group — including users in the United States, the United Kingdom, and most of Asia — is, for now, on the outside looking in.

Two details about this first phase are worth keeping in mind:

• Launching in the EU does not mean the feature is permanent there first; it means the EU is where the rollout begins.
• Availability is regional, so two users on the same plan in different countries may see different behaviour from the same model.

Whether the EU-first approach becomes a template for later launches, or simply reflects a one-off compliance decision, will become clearer as the rollout widens.

Not the Global Default Yet

Despite the broad rollout across ChatGPT and Codex, OpenAI is careful to frame the feature as a regional experiment rather than a universal standard. The company states that the watermarking will not be the "global default" yet, meaning users outside the European Union will not necessarily see the same behaviour immediately.

This distinction matters for several reasons:

• Regional variation: The EU is the first launch market, so availability and defaults may differ by country.
• Evolving policy: OpenAI's wording signals that the approach could change as the technology and its reception develop.
• User experience: Outside the EU, outputs may remain unmarked for now, creating an uneven landscape across regions.

The cautious phrasing suggests OpenAI is treating textGrain as a test case, gathering feedback before committing to a worldwide default. For users, the practical takeaway is that watermarking is not yet a guaranteed feature of every ChatGPT or Codex response. Instead, it is a market-specific deployment that OpenAI can expand, adjust, or refine over time.

What This Means for Users

If you use ChatGPT or Codex, the arrival of textGrain watermarking is something you will experience through the text these tools produce, not through any new button or setting. The rollout touches both products, so the change applies across the assistants you already rely on for writing, coding and everyday tasks.

Where you are located shapes when this reaches you. Users in the EU are first in line, making the region the initial launch market for the feature. Everyone else will have to wait, because watermarking is not yet the global default.

For now, that leaves two groups with different expectations:

• EU users: they encounter watermarked output first, ahead of the rest of the world.
• Users elsewhere: they continue without the feature until a global default arrives.

The practical effect is that the same prompt may carry different underlying characteristics depending on where it is run, even though the visible answer looks unchanged.