OpenAI textGrain: Why Editing Beats the EU Watermark

OpenAI’s Invisible ChatGPT Watermark in the EU: How textGrain Works and Why Editing Matters

OpenAI says it will add an invisible watermark to eligible ChatGPT and Codex text generated in the European Union. The technology, called textGrain, is designed to help identify text from OpenAI systems in response to EU AI Act transparency requirements. It does not insert a visible label or hidden characters. Instead, it subtly adjusts word choices so a detector can look for a statistical pattern. That signal has limits, especially when text is short or edited.

What Is OpenAI’s textGrain Watermark?

textGrain is OpenAI’s text-watermarking technology. As a model generates text, it makes small adjustments to its word or word-piece choices. Across a passage, those choices can form a statistical signal that a detector can test for.

The watermark is not visible to readers, and it is not an invisible character embedded between words. It is a pattern in the wording itself. That means ordinary copying and pasting does not necessarily remove it, although rewriting or replacing words can weaken the signal.

Why Is OpenAI Adding Watermarks in the EU?

OpenAI says the rollout is intended to support the EU AI Act’s transparency obligations for identifying AI-generated content. The company plans to add watermarking to eligible ChatGPT and Codex text output in the EU over time. At launch, it is not a global default for those products.

OpenAI has also said API customers worldwide can opt in to text watermarking for select models. The API option is separate from the planned EU rollout in ChatGPT and Codex.

How Well Can the Watermark Be Detected?

Detection is not guaranteed. OpenAI’s published evaluations show that performance depends on passage length and subject matter. At a target false-positive rate of 1%, the detector identified watermarks in about 80% of 200-token psychology passages and about 95% of 400-token passages. Results were substantially weaker for math, where there is less freedom to vary the wording.

• Longer passages generally give the detector more wording patterns to evaluate.
• Short passages are harder to classify reliably.
• Constrained writing, including some technical or mathematical content, can make detection less effective.
• Edited text may retain some signal, but detection can decline as more wording changes.

How Editing Affects textGrain Detection

Editing can substantially weaken the watermark. In OpenAI’s evaluation of 400-token passages, replacing 10% of the words with synonyms reduced detection from about 92% to 66%. Replacing 25% of the words lowered it to 17%.

These figures describe results in a specific evaluation, not a universal detection rate for every passage. They show why a detector’s result should be treated as an indicator rather than definitive proof of a text’s origin.

What a Watermark Can and Cannot Establish

A detected signal may indicate that an OpenAI system generated or processed part of a passage. It does not reveal exactly how much a person contributed, identify the user, establish ownership or responsibility, or verify whether the text is accurate.

Likewise, failure to detect a watermark does not prove that a person wrote the text. The passage may be too short, edited, or generated by another system. A detector result should not be treated as a standalone authorship judgment.

What This Means for ChatGPT Users

For EU users, the planned watermark is intended to work in the background, without changing how text appears to readers. For organizations assessing AI-generated content, the rollout offers another provenance signal, but not a complete verification method.

• Do not assume that every ChatGPT response will be detectable under all conditions.
• Do not use a missing watermark as proof of human authorship.
• Interpret positive results in context, especially for short or heavily edited passages.
• Use other evidence and review processes when provenance or accountability matters.

Conclusion

OpenAI’s textGrain watermark is an invisible statistical pattern in model word choices, intended to help identify eligible ChatGPT and Codex text in the EU. It can support transparency, but its reliability varies with length, subject matter, and editing. A detected watermark is a useful clue, not proof of authorship or accuracy, and an undetected watermark does not establish that a person wrote the text.