OpenAI is taking a major step toward making AI-generated text easier to identify, announcing that it will begin adding an invisible watermark to content produced by ChatGPT and Codex in the European Union. The move is designed to help the company comply with the transparency requirements of the European Union’s AI Act while creating a technical way to distinguish AI-generated text from human-written content.
The watermarking system, which OpenAI plans to roll out over the coming weeks, will initially apply to eligible ChatGPT and Codex users across all plans in the EU. Unlike a visible label or symbol, the technology works directly within the generated language, creating a subtle statistical pattern that can be detected by a specialized system.
The announcement highlights a growing shift in the AI industry toward content provenance as governments, technology companies, educators, businesses, and researchers grapple with the challenge of determining where digital content comes from.
A New Layer of AI Transparency
Under the EU AI Act’s transparency framework, providers of certain AI systems are required to ensure that AI-generated or manipulated content can be identified by machines in appropriate circumstances. The relevant transparency rules took effect on August 2, putting greater pressure on AI companies to develop practical mechanisms for identifying synthetic content.
OpenAI’s approach is notably different from traditional watermarks that appear as logos, labels, or visible markings.
Instead, the company’s technology subtly influences the model’s selection of words during generation. Individually, these changes are virtually impossible for a reader to notice. When repeated across a longer passage, however, they create a detectable pattern that can be recognized by a compatible detector.
Because the signal is embedded within the text itself, it can remain attached to the content when users copy and paste the material.
OpenAI said the watermark does not contain information that identifies the person who generated the text. The company also reported that enabling the system did not produce a meaningful change in model performance in its testing.
How OpenAI’s TextGrain Technology Works
OpenAI has also released technical research detailing the technology behind its approach, called textGrain.
Developed with researchers from the University of Pennsylvania and Yale, the system is based on manipulating the model’s word-selection process in a way that remains invisible to readers but statistically recognizable to a detector.
At a simplified level, the technique uses a secret key to organize possible next-word predictions. The model then makes selections according to subtle constraints created by the watermarking system. A single decision would not provide enough evidence to identify the source.
But as hundreds of these small statistical nudges accumulate throughout a passage, they can form a distinctive pattern.
The detector can then analyze the text and determine whether the pattern associated with OpenAI’s watermark is present.
This approach is particularly significant because it does not require a visible tag to remain attached to the content. A user can copy the generated text into another document, email, website, or platform without necessarily removing the underlying signal.
Watermarking Will Initially Be Limited to Europe
OpenAI is taking a deliberately regional approach to the rollout.
For ChatGPT and Codex, the watermark will initially be introduced only for eligible users in the European Union. The company said the feature is expected to reach users across all plans during the coming weeks.
For developers using OpenAI’s API, the situation is different. Developers around the world can begin enabling the technology for selected models, although it will remain turned off by default.
OpenAI has made clear that it is not introducing text watermarking as a global default at launch.
That decision reflects the technical and practical challenges surrounding AI detection, particularly because no watermarking technology can provide a perfect answer about who wrote a piece of content or how much AI was involved in producing it.
Editing Can Weaken the Watermark
One of the biggest challenges facing AI watermarking is what happens after generated content is edited.
OpenAI’s own testing indicates that the watermark can be weakened when people substantially modify the original text.
In one test, replacing around 10% of the words with synonyms reduced detection performance from approximately 92% to 66%. This demonstrates that although the watermark can survive ordinary copy-and-paste behavior, it is not necessarily permanent under deliberate editing.
OpenAI also identified several situations where detection becomes more difficult. Very short passages, mathematical answers, and translated text can all pose challenges for the system.
These limitations are important because they prevent the technology from being treated as a definitive AI-or-human verdict.
A detector finding no watermark does not necessarily mean that a human wrote the text.
The content could have been heavily edited, could be too short for reliable analysis, or could have been produced using another AI system whose watermarking technology is different.
OpenAI Warns Against Treating Watermarks as Proof
OpenAI is also emphasizing an important distinction between AI provenance and authorship.
A watermark can potentially indicate that an OpenAI system generated or processed part of a passage. It cannot determine how much human thinking, editing, decision-making, or creativity contributed to the final result.
That distinction could become particularly important in workplaces, schools, publishing, research, and creative industries.
For example, a person may use ChatGPT to brainstorm ideas, reorganize information, improve grammar, or develop an initial draft before extensively rewriting the material themselves. A watermark could indicate that an OpenAI system was involved, but it would not establish how much of the final work came from the AI.
Similarly, the absence of a watermark cannot establish that a piece of writing is entirely human-created.
A Broader Industry Push Toward AI Provenance
OpenAI’s decision comes amid increasing movement across the technology industry toward identifying AI-generated content.
Anthropic announced in August that it would watermark text generated by its Claude models worldwide. The company’s decision generated criticism from some users, particularly those concerned that watermarking could be used to scrutinize employees or students who use AI as an assistive tool.
The debate illustrates the larger question facing the industry: Should AI-generated content be identified because it was produced by an AI system, or should the focus instead be placed on how the technology was used?
OpenAI’s approach appears to acknowledge both sides of that debate. Its watermark can provide an indication that its technology was involved, while explicitly avoiding the claim that the system can measure the level of human contribution.
Several major technology companies, including Anthropic, Google, Meta, Microsoft, and OpenAI, have also committed to following the EU’s voluntary code of practice covering transparency around AI-generated content.
Why This Matters for ChatGPT Users
The introduction of invisible watermarking could have significant implications for organizations and individuals using generative AI.
For businesses, provenance technology could become another tool for establishing whether documents, reports, communications, or other content were generated using AI.
For educators and researchers, it could potentially provide an additional signal when evaluating AI-assisted submissions. However, OpenAI’s own warnings suggest that such systems should not be treated as definitive evidence of misconduct.
For publishers and media organizations, provenance could eventually become part of a broader system for understanding how digital content was created and modified.
At the same time, the technology raises questions about interoperability. If every AI company develops a different watermarking method, detecting content across multiple AI platforms could remain difficult.
The Bigger Battle Over AI-Generated Content
OpenAI’s move represents more than a technical upgrade to ChatGPT. It is part of a much larger effort to build trust around generative AI at a time when synthetic content is becoming increasingly difficult to distinguish from human work.
The EU is emerging as one of the strongest regulatory forces shaping that transition. By requiring greater transparency, the AI Act is pushing technology companies to move from voluntary disclosure toward systems that can make AI involvement more technically identifiable.
Yet watermarking alone will not solve the broader provenance problem.
Text can be edited. Content can be translated. Multiple AI systems can be used together. Human writers can also substantially transform AI-generated drafts. As a result, the future of AI transparency will likely depend on a combination of watermarking, metadata, detection technologies, disclosure policies, and clearer standards for human-AI collaboration.
For now, OpenAI’s European rollout marks a significant milestone. ChatGPT-generated words may soon carry a signal that readers cannot see, but machines can detect—creating a new, largely invisible layer of provenance around AI-generated text.
Read more:Â Volvo Auto India Names Karan Arora as Marketing & PR Director to Drive Brand Growth
