ChatGPT Text Is Getting an Invisible Watermark in the EU: How textGrain Works

OpenAI is bringing invisible watermarking to text generated by ChatGPT and Codex in the European Union.

Announced on October 5, 2026, the system is called textGrain. Instead of adding hidden characters, unusual punctuation, or visible labels, it changes the statistical pattern of the model's word choices so specialized software can look for a watermark later.

The rollout is tied to the European Union's AI transparency requirements, but OpenAI is also making watermarking available as an optional feature for API customers around the world.

Read OpenAI's official text provenance announcement

How textGrain works

TextGrain does not place a normal watermark on the page.

OpenAI says the system subtly changes how the model selects between possible words or word pieces while generating a response. Across a long enough passage, those choices create a statistical pattern that a detector can search for.

That means the watermark is effectively carried by the wording itself.

Copying and pasting the text does not add a hidden file or secret character. If the wording is preserved, the statistical signal may remain detectable.

Where it will appear

OpenAI says it will introduce text watermarking to eligible ChatGPT and Codex users across all plans in the European Union over the coming weeks.

It is not becoming a worldwide default at launch.

API customers globally can opt in to watermarked text outputs for select models, with watermarking remaining off by default in the API.

OpenAI is also working with cloud and distribution partners to make the feature available through additional services.

Editing can weaken the watermark

This is where the technology becomes more complicated.

OpenAI says text watermark detection is not guaranteed, especially when text is short, constrained, translated, or substantially rewritten.

In one test cited by OpenAI and reported by TechCrunch, replacing 10% of the words with synonyms reduced detection from about 92% to 66%.

Replacing 25% of the words reduced detection to roughly 17%.

That does not mean every edited passage will behave exactly the same way. It demonstrates why watermarking should not be treated as a perfect AI detector.

A missing watermark does not prove a human wrote it

One of OpenAI's most important warnings is that detection results can be misunderstood.

If the detector finds a watermark, it can provide evidence that an OpenAI system generated or processed at least part of the passage.

But OpenAI says a watermark does not tell you:

• who used the model;
• who owns the text;
• how much a human edited or contributed;
• whether the information is accurate; or
• whether the text was used in the correct context.

Likewise, failing to detect a watermark does not prove that a person wrote the passage.

Short answers, unsupported models, older outputs, edited text, translated text, and material generated by another AI provider may all lack a detectable OpenAI watermark.

Why the detector is not public yet

OpenAI is not launching the text detector as an unrestricted public tool.

Instead, access will initially be limited to approved researchers and expert organizations that can help evaluate reliability and responsible uses.

OpenAI says this cautious rollout is partly because text watermark detectors can produce both false positives and false negatives.

A false positive could incorrectly suggest that human-written material contains an OpenAI watermark. A false negative could miss a watermark that is actually present.

The company says approved detector users will only receive a result about whether an OpenAI watermark is detected. The tool is not designed to reveal the person who generated the text, their prompts, or their private conversations.

Does watermarking affect ChatGPT's quality?

OpenAI says its evaluations have not found a meaningful reduction in model quality from text watermarking.

According to the company, benchmark differences with and without the watermark fell within normal evaluation variation. OpenAI also says the technology has a negligible effect on generation speed.

The challenge is therefore less about making the model function and more about making the watermark survive real-world editing while avoiding incorrect detections.

Why OpenAI is doing this

The rollout is part of OpenAI's response to the EU AI Act, which includes transparency requirements for AI-generated content.

OpenAI already uses other provenance technologies for different media.

Supported generated images use Content Credentials and invisible watermarking, while supported audio also carries provenance signals. Text is more difficult because words can be copied, paraphrased, shortened, translated, or completely rewritten with very little technical friction.

TextGrain is OpenAI's attempt to add a machine-readable signal directly to generated wording while acknowledging those limitations.

What this means outside the EU

For users in the Philippines, there is no announced default textGrain rollout for normal ChatGPT use as of October 6, 2026.

The mandatory ChatGPT and Codex rollout described by OpenAI is EU-specific.

Developers and organizations using the OpenAI API outside Europe can opt in where supported, so watermarked OpenAI-generated text could still appear outside the EU through products that enable the feature.

For ordinary ChatGPT users in the Philippines, however, OpenAI has not announced that all text responses will receive textGrain by default.

The bottom line

OpenAI's text watermark is not a hidden label tucked inside a document.

textGrain changes the statistical pattern of the words the model chooses, creating a signal that specialized detection software can search for.

It will roll out to eligible ChatGPT and Codex text in the EU, while API customers around the world can opt in on supported models.

The technology could make AI-generated text easier to identify in some situations, but OpenAI is explicitly warning that it is not definitive proof of authorship. Editing, translation, short passages, false positives, and missed detections all remain significant limitations.

That makes textGrain useful as a provenance signal, but not a universal test for deciding whether a person or an AI wrote something.