OpenAI is introducing a new way to identify text generated or processed by its AI systems. The company plans to add an invisible text watermark to eligible ChatGPT and Codex outputs in the European Union over the coming weeks.
The move is part of OpenAI’s response to the European Union’s AI transparency requirements. The company is also opening its text-watermarking capabilities to API customers worldwide on an opt-in basis.
The technology, called textGrain, is designed to introduce a statistical signal into the way an AI model selects words. A dedicated detector can then analyse a passage to check for the signal associated with OpenAI-generated content.
However, OpenAI acknowledges that this technology is not a perfect AI detector. Editing, translating or shortening text can make the watermark harder to detect, and a detected watermark cannot reveal who used ChatGPT or how much AI contributed to the final content.
Here’s what OpenAI’s new text-watermarking system means for ChatGPT users, businesses, writers and developers.
What Is OpenAI’s New Text Watermark?
OpenAI’s new system is designed to place an invisible statistical signal into eligible AI-generated text.
Unlike a visible watermark placed over an image, users will not see a special symbol, word or mark in their ChatGPT responses.
Instead, textGrain subtly influences statistical patterns in the model’s word choices. A detector can then analyse a sufficiently long passage and look for the signal associated with OpenAI-generated content.
The purpose is to provide a form of content provenance information that can help indicate where content originated and whether it contains a signal associated with an OpenAI system.
Text watermarking is one part of OpenAI’s broader approach to identifying AI-generated content. It is intended to provide an additional source of evidence rather than a definitive answer about who wrote a particular passage.
Why Is OpenAI Watermarking ChatGPT Text in the EU?
The European Union’s AI Act introduces transparency obligations for certain AI-generated content.
OpenAI’s text-watermarking approach is intended to help address applicable requirements by making eligible AI-generated text identifiable through a machine-readable signal.
The company has previously expressed concerns about the reliability of text watermarking. The latest approach reflects an effort to improve content provenance while recognising that the technology still has limitations.
Rather than immediately introducing watermarking as a worldwide default, OpenAI is taking a phased approach, beginning with eligible ChatGPT and Codex outputs in the European Union.
It is important to note that watermarking is one approach to content provenance, not the only possible way to address AI transparency requirements.
How Does textGrain Work?
The technology behind OpenAI’s system is called textGrain.
It introduces an invisible statistical signal through the model’s word choices without adding a visible mark to the text.
A detector can then analyse the passage and search for evidence of that signal.
According to OpenAI’s reported evaluations, detection works better on longer passages than shorter ones. At a target false-positive rate of 1%, the detector identified the watermark in approximately:
- 80% of 200-token passages
- 95% of 400-token passages
These figures apply to the specific content and testing conditions reported by OpenAI. Performance was considerably weaker for highly constrained content, such as mathematics, where the model has less flexibility in its word choices.
This means watermark detection should not be interpreted as an absolute test of whether something was written by AI.
Will Every ChatGPT Message in the EU Be Watermarked?
Not necessarily.
OpenAI says it plans to introduce text watermarking for eligible ChatGPT and Codex text outputs in the European Union over the coming weeks.
However, the company has not said that every possible piece of text will always produce a detectable watermark.
Several factors can affect detection, including:
- Very short passages
- Extensive editing or rewriting
- Translation into another language
- Highly constrained writing formats
- Use of unsupported models or systems
OpenAI is also not making text watermarking a global default for all ChatGPT users at launch.
The initial rollout is therefore regional and limited to eligible outputs, rather than a universal change affecting every ChatGPT response worldwide.
What About ChatGPT Users Outside the EU?
For regular ChatGPT and Codex users outside the European Union, OpenAI is not introducing text watermarking as a global default at launch.
The EU-focused rollout gives the company an opportunity to evaluate real-world usage and feedback before deciding how the technology should develop.
This distinction is important because the announcement does not mean every ChatGPT response worldwide will suddenly become watermarked.
Users outside the EU should also distinguish between regular ChatGPT access and API-based applications, which have a separate opt-in arrangement.
What About OpenAI API Customers?
OpenAI’s approach is different for API customers.
Starting October 5, 2026, API customers worldwide can opt in to text watermarking for select supported models.
The feature is off by default, meaning developers must choose to enable it where it is available.
This gives businesses and developers the option to add a provenance signal to content generated by their applications.
For example, organisations using OpenAI models for content creation, customer communications or other text-generation workflows may be able to incorporate watermarking into their content processes.
OpenAI is also working with cloud partners to make watermarking available for supported OpenAI model outputs accessed through their services.
However, availability depends on the supported models and implementation arrangements.
Can OpenAI’s Watermark Detect AI-Generated Text Every Time?
No.
One of the most important points about the announcement is that OpenAI’s watermarking and detection technologies have significant limitations.
The watermark can become harder to identify when text is edited, rewritten, translated or shortened.
OpenAI’s testing illustrates how even relatively modest changes can weaken the signal.
In one test involving 400-token English passages, replacing approximately 10% of the words with synonyms reduced detection from roughly 92% to 66%. Replacing 25% of the words reduced detection to approximately 17%.
These results demonstrate why a watermark should not be treated as a foolproof AI detector.
A detector may fail to identify a watermark even when AI-generated text originally contained one. Similarly, detection results must be interpreted in the context of the content and the testing conditions.
Does an OpenAI Watermark Prove That ChatGPT Wrote the Text?
No.
A detected watermark can indicate that a passage contains a signal associated with an OpenAI system, but it cannot answer every question about how the content was produced.
According to OpenAI, a watermark does not:
- Measure how much a human contributed to the content
- Identify the person who used the AI system
- Reveal the user’s account, prompt or conversation
- Establish ownership of the text
- Determine whether the use of AI was lawful
- Verify whether the information in the text is accurate
- Prove that a person did not write the text
Similarly, the absence of a detected watermark does not prove that a human wrote the content entirely without AI assistance.
This distinction is particularly important for schools, employers, publishers and organisations considering the use of AI-detection technologies.
A watermark should be treated as one piece of evidence, not a final judgement about authorship.
Who Will Be Able to Use OpenAI’s Watermark Detector?
OpenAI is not making its text detector publicly available at launch.
Instead, approved researchers and expert organisations can apply for access.
The initial approach is intended to support the evaluation and improvement of the technology. OpenAI plans to provide detector access on a case-by-case basis to organisations that can contribute to this work.
The detector can report whether it detects an OpenAI watermark, but it does not identify the person who used ChatGPT or reveal their prompts and conversations.
Restricting access also reflects the need to evaluate false positives and missed detections before relying more widely on the technology.
What Happens When Humans Edit ChatGPT Text?
Human editing is one of the biggest challenges for AI text provenance.
A person might use ChatGPT to generate an initial draft and then substantially rewrite it. They may change the structure, replace words, translate the content or combine AI-generated material with their own writing.
These changes can weaken the statistical signal and make the watermark more difficult to detect.
OpenAI’s reported tests demonstrate this limitation: synonym replacement reduced detection rates, and more extensive changes weakened the signal further.
For businesses, writers and students, this means a watermark alone cannot reliably establish exactly how much AI assistance went into a final article, essay or document.
It also reinforces the importance of using appropriate context when evaluating how a piece of content was created.
Will the Watermark Affect the Quality of ChatGPT Responses?
OpenAI reports that its evaluations found minimal to no difference between outputs with watermarking enabled and outputs without it in the benchmarks it tested.
The company is therefore aiming to introduce the provenance signal without significantly changing the quality or usefulness of generated text.
However, this does not guarantee identical results in every situation. Performance and output quality may vary depending on the model, the type of content and the specific use case.
Further evaluation will help establish how the system performs across different writing tasks and real-world conditions.
How Does OpenAI’s Approach Compare With Other AI Tools?
AI companies are exploring different ways to identify AI-generated content and provide information about its origin.
Text watermarking is one approach because it attempts to embed a detectable signal directly into generated text. Other provenance methods may rely on metadata, content credentials or information about how digital content was created.
For example, Anthropic has also explored watermarking approaches for AI-generated text.
However, the effectiveness of any method depends on its implementation, the content being analysed and how much the content changes after generation.
No single technique should automatically be treated as conclusive proof of authorship. Text watermarking is best understood as one developing approach within the wider field of AI content provenance.
What Does This Mean for Students and Writers?
The technology could eventually give schools, publishers and other organisations another signal when evaluating AI-generated content.
However, treating a watermark as definitive evidence that someone used AI improperly would be risky.
For example, a detected watermark cannot determine whether a student wrote an original assignment and used ChatGPT only to improve grammar, or whether the system generated most of the work.
Likewise, the absence of a watermark cannot prove that an assignment was written entirely by a human.
Students and writers should continue to follow the relevant rules and disclosure requirements of their schools, publishers or employers. Organisations should also consider the limitations of detection technology before making decisions based on its results.
What Does This Mean for Businesses?
Businesses using ChatGPT for marketing, customer support, research and content creation may eventually have another way to establish the provenance of AI-generated material.
For developers, the API opt-in could provide an additional option when building applications that generate text.
However, organisations operating across different regions should understand that the rollout arrangements may differ between eligible EU ChatGPT and Codex outputs and supported API models available worldwide.
Businesses should avoid building policies that assume a watermark is a definitive measure of AI authorship.
Instead, watermarking could be treated as one signal within a broader content-verification process that considers editorial records, source material, human review and the intended use of the content.
OpenAI’s Broader Approach to AI Content Provenance
OpenAI is taking a layered approach to content provenance rather than relying on a single technology.
For supported images and audio, the company uses or supports approaches that include Content Credentials, C2PA standards and invisible watermarks in applicable cases.
Text presents different challenges because it can easily be rewritten, translated or edited without leaving an obvious visual trace.
OpenAI plans to continue researching how text watermarking performs under these conditions. The company may also expand detector access as the technology and supporting evidence develop.
The broader goal is to improve the ability to understand where digital content comes from while recognising that no individual method answers every question about its creation.
Final Thoughts
OpenAI’s decision to introduce invisible watermarks for eligible ChatGPT and Codex text in the EU represents an important development in AI content provenance.
Its textGrain system aims to help identify passages containing a signal associated with an OpenAI system, while the API opt-in gives developers worldwide an option to use the technology with supported models.
However, it is not a foolproof solution for identifying AI-written content.
Editing, translation, short passages and highly constrained writing can make detection more difficult. More importantly, a watermark cannot identify the person who used ChatGPT or determine how much AI contributed to the final content.
The significance of OpenAI’s announcement lies less in creating a perfect AI detector and more in establishing another layer of provenance for AI-generated text.
As AI-generated content becomes increasingly common, technologies such as textGrain could become an important part of how platforms, businesses and researchers understand where digital content comes from.
