OpenAI Is Bringing Invisible Watermarks to ChatGPT Text in the EU

OpenAI is taking a significant new step toward making AI-generated writing identifiable. The company has announced that it will begin adding an invisible watermark to eligible ChatGPT and Codex text generated in the European Union, bringing text provenance into the same broader conversation that has already surrounded AI-generated images and audio.

The change is not a visible label placed at the top or bottom of a ChatGPT response. Readers will not see a special symbol, watermark, hidden message or unusual formatting when they read the text. Instead, OpenAI's system subtly influences the model's choice between words or word pieces, creating a statistical pattern that a specialized detector can later analyze.

OpenAI says the technology, called textGrain, is being introduced as part of its response to the transparency requirements of the European Union's AI Act. The company says eligible ChatGPT and Codex users across all plans in the EU will receive the watermark over the coming weeks. The initial rollout will remain limited to the EU rather than becoming a worldwide default.

What Exactly Is the ChatGPT Watermark?

The word "watermark" can make it sound as though OpenAI is going to stamp every AI-written response with a visible notice. That is not what is happening.

Instead, textGrain works inside the generation process itself. When a model has several words or word pieces that could naturally fit a sentence, the system can subtly influence which option it selects. Across a sufficiently long passage, those choices form a statistical signature that a detector can look for.

For a normal reader, the result should simply look like ordinary ChatGPT writing.

This also means the watermark travels with the wording when the text is copied and pasted. It is fundamentally different from metadata that can disappear when a document is exported or reformatted. OpenAI says textGrain does not insert hidden characters, invisible spaces or strange punctuation into the text.

Why the European Union Is Driving the Change

The decision is closely tied to the EU AI Act, which introduces transparency obligations for providers of AI systems that generate synthetic content.

Article 50 requires providers of general-purpose AI systems that generate synthetic text and other forms of content to ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, taking into account technical feasibility and the limitations of the technology. The rules also contain specific provisions concerning AI-generated text published to inform the public about matters of public interest.

For OpenAI, that creates a difficult technical problem. Making AI-generated text identifiable sounds simple in principle, but text is fundamentally different from an image or an audio recording.

An image can carry an embedded signal without changing what a person sees. Text is much more sensitive to tiny changes in word selection because language has enormous flexibility. A system that is too aggressive could make AI responses sound unnatural, while a system that is too weak could become difficult to detect.

OpenAI's approach is therefore designed to operate underneath the visible writing rather than adding a conspicuous label to it.

OpenAI Admits AI Text Detection Still Has Serious Limits

Perhaps the most interesting part of OpenAI's announcement is not the watermark itself, but the company's acknowledgement that text watermarking is still an imperfect technology.

OpenAI says its internal evaluations found that textGrain performed well against the approaches it tested, including Google's SynthID for text. But the company also warns that strong performance in controlled testing does not mean that detection will always be reliable in real-world situations.

Short passages are particularly difficult.

According to OpenAI's testing, at a target false-positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages in one set of tests, compared with roughly 95% for 400-token passages. Results were considerably weaker for areas such as mathematics, where there is less flexibility in choosing between equivalent expressions.

That distinction matters. A detector saying that a passage contains an OpenAI watermark should not automatically be interpreted as absolute proof that every word was produced by ChatGPT.

Editing Can Weaken the Watermark

The biggest challenge may come after the text has already been generated.

OpenAI's own evaluation found that relatively modest rewriting could substantially weaken the signal. In one test involving 400-token passages, replacing 10% of the words with synonyms reduced detection from about 92% to 66%. Replacing 25% of the words reduced detection to about 17%.

That makes the technology more useful as a provenance signal than as a perfect AI-content lie detector.

Someone could take AI-generated writing, edit it, paraphrase portions of it or translate it, and the detector may become less certain that the original watermark remains detectable.

OpenAI's own help documentation now makes a similar point: substantial rewriting, paraphrasing or translation can make the watermark harder to detect, while short text may also be difficult to identify reliably.

OpenAI Is Keeping the Detector Restricted

OpenAI is not immediately giving everyone a public tool that can scan any paragraph and declare whether ChatGPT wrote it.

The company has opened applications for access to its text watermark detector, but initial access will be limited to approved researchers and expert organizations. The goal is to allow specialists to evaluate how reliable the technology is and how its results should be interpreted.

That cautious approach makes sense given the possibility of false positives and false negatives.

A publicly available detector could easily be treated as an unquestionable authority, even though OpenAI's own research shows that detection becomes more difficult with shorter passages and significant editing.

For schools, publishers, researchers and organizations dealing with AI-generated content, that distinction could become increasingly important.

ChatGPT Users Outside Europe Are Not Getting the Same Default Yet

For users outside the European Union, the announcement does not mean that ChatGPT text will suddenly receive the same watermark by default.

OpenAI says its initial rollout is specifically focused on eligible ChatGPT and Codex text generated in the EU. The company is deliberately avoiding a global default at launch, giving it time to observe how the system performs in real-world use.

There is a separate development for API customers.

Starting October 5, 2026, OpenAI says API customers worldwide can opt in to text watermarking for supported models. The feature remains off by default for the API. OpenAI also says it plans to expand model coverage over time.

That creates two different tracks: a regional default rollout for eligible ChatGPT and Codex output in the EU, and an optional watermarking capability for API customers around the world.

This Is About Provenance, Not Accuracy

There is another important distinction that users should understand.

An OpenAI watermark is designed to provide information about where text came from, not whether the information contained in that text is true.

OpenAI's own documentation says provenance signals do not guarantee that content is accurate, legally owned, unedited or presented in the correct context.

In other words, a detected watermark can potentially indicate that a passage contains an OpenAI provenance signal, but it cannot tell a reader whether the passage is factually correct.

That could become especially important as AI-generated material becomes more common in journalism, education, business documents and online publishing.

OpenAI's Larger Push Toward AI Content Provenance

The textGrain rollout is part of a wider effort by OpenAI to make its generated content easier to identify.

For images, OpenAI says it uses Content Credentials and SynthID as provenance signals. For audio, it uses SynthID. Text now has its own dedicated system through textGrain in the EU.

The broader idea is straightforward: as AI-generated material becomes harder to distinguish from human-created work, provenance signals can give software another way to determine how content was produced.

But OpenAI is also acknowledging an uncomfortable reality: no watermarking system is perfectly permanent.

The more a piece of content is transformed, the harder its original provenance signal may become to detect. That means watermarking is likely to become one part of a larger ecosystem of AI disclosure, provenance standards and editorial practices rather than a single universal solution.

What This Means for ChatGPT Users

For most EU ChatGPT users, the practical experience should remain largely unchanged. They will still see normal-looking ChatGPT responses, but eligible text generated by the service will contain a statistical provenance signal that software can potentially detect.

There should be no visible watermark to remove, no special button to press and no formatting change that ordinary readers need to understand.

The bigger change will happen behind the scenes.

As governments increasingly demand transparency around synthetic content, companies such as OpenAI are being pushed to build technical systems capable of identifying AI-generated material without fundamentally changing the user experience. The EU's approach could also influence how other AI companies handle provenance in the future.

OpenAI's decision to begin with the European market is therefore more than a regional product change. It is an early test of whether invisible text provenance can work at internet scale—and whether it can provide meaningful transparency without becoming another unreliable AI detector.

Sources

• OpenAI — Our approach to EU text provenance: OpenAI official announcement
• OpenAI Help Center — Provenance signals in OpenAI-generated content: OpenAI Help Center
• European Commission — AI Act Article 50 transparency obligations: EU AI Act Service Desk
• TechCrunch — OpenAI will start watermarking ChatGPT's text in the EU: TechCrunch report