OpenAI is introducing invisible text watermarking for eligible ChatGPT and Codex text generated in the European Union, creating a new way to identify content produced with its AI systems.
Artificial intelligence is becoming part of everyday writing, education, business and digital publishing. As AI-generated content becomes harder to distinguish from human-written material, questions around transparency and content origin are becoming increasingly important.
In its latest announcement, OpenAI has revealed a new approach to text provenance, introducing an invisible watermarking system called textGrain for eligible ChatGPT and Codex text outputs in the European Union.
Related Post
The move is linked to the transparency requirements of the European Union’s AI Act and represents a significant development in how AI-generated text may be identified in the future.
OpenAI says the EU rollout will take place over the coming weeks and will apply to eligible ChatGPT and Codex users across plans. The company has also confirmed that it is not making text watermarking a global default at launch.
What Is ChatGPT Text Watermarking?
Unlike a visible logo or label, OpenAI’s new text watermark is designed to be invisible to people reading the content.
The technology, called textGrain, subtly changes the statistical pattern of the words or word pieces selected by the AI model while generating a response.
A specialised detector can then analyse a passage and look for the statistical signal associated with the watermark.
This means readers should not expect to see a message such as “Written by ChatGPT” appearing inside an article or paragraph. The watermark exists within the pattern of the generated wording itself.
OpenAI says the system does not insert hidden characters, invisible spaces or special punctuation into the text.
Why Is OpenAI Introducing the Watermark?
The announcement is closely connected to the EU AI Act, which includes transparency requirements for providers of generative AI systems.
OpenAI says the EU AI Act requires generated text to be identifiable in a machine-readable way. Its new text watermarking approach is intended to support that requirement while recognising that watermarking technology still has limitations.
The wider goal is to improve content provenance – essentially helping people and organisations understand where digital content came from and whether an AI system was involved in creating it.
This is part of a broader effort by technology companies to make AI-generated content more identifiable as AI becomes increasingly integrated into online communication.
How Does textGrain Work?
When ChatGPT generates a sentence, the model generally has several possible words or word pieces it could choose next.
OpenAI’s textGrain system subtly influences those choices according to a secret pattern. Over a longer passage, those choices can create a statistical signal.
A detector that knows the appropriate signal can then analyse the writing and estimate whether an OpenAI watermark is present.
The important point is that the text itself remains readable and does not contain an obvious visual marker.
OpenAI says the watermark is designed to be part of the model’s word-selection process rather than something added afterwards.
Will ChatGPT Responses Look Different?
According to OpenAI, text watermarking is not expected to meaningfully reduce the quality of model responses.
The company tested watermarked and unwatermarked outputs across several benchmarks and reported that performance differences were generally within the normal variation seen between evaluation runs.
OpenAI also says the effect on model speed is negligible.
For everyday users, this means the change is intended to happen largely behind the scenes.
A person using ChatGPT should not necessarily notice a visible difference simply because the response contains a watermark.
Can the Watermark Prove That ChatGPT Wrote Something?
No.
This is one of the most important points about OpenAI’s new system.
A watermark is intended to provide a provenance signal, but it does not establish exactly who wrote a piece of content or how much AI was involved.
For example, a watermarked passage could have been heavily edited by a human after being generated. The watermark itself cannot measure the amount of human contribution.
OpenAI also says the signal does not identify a user’s account, prompt, conversation or personal information.
Therefore, a watermark should not automatically be interpreted as proof that an entire article was written entirely by AI.
AI Watermarking Is Not the Same as an AI Detector
There is an important distinction between OpenAI’s watermarking system and many existing third-party AI detection services.
Traditional AI detectors generally examine the finished text and use statistical or linguistic patterns to estimate whether it appears AI-generated.
OpenAI’s textGrain approach works differently.
The signal is introduced during generation, allowing a compatible detector to search specifically for that embedded provenance pattern.
OpenAI says this approach is designed to support the EU requirement for a machine-readable signal in generated text.
Watermark Detection Has Limitations
Although the technology is designed to identify OpenAI-generated text, OpenAI acknowledges that detection is not perfect.
Short passages can be difficult to identify because there may not be enough text for the statistical signal to become clear.
Highly constrained writing can also present problems. For example, mathematical answers may offer fewer alternative ways of expressing the same information.
OpenAI’s testing found that detection was stronger for longer passages and more flexible forms of writing.
Editing Can Weaken the Signal
Human editing can also affect detection.
In one OpenAI evaluation involving 400-token passages, replacing around 10% of the words with synonyms reduced detection from approximately 92% to 66%.
When around 25% of the words were replaced, detection fell to approximately 17%.
This demonstrates why text watermarking should not be treated as an infallible AI authorship test.
Translation and substantial rewriting can also make the watermark harder to detect.
What Happens If No Watermark Is Found?
A negative result does not necessarily mean a human wrote the content.
There are several possible explanations for a missing signal.
The text could be:
Too short for reliable detection
Heavily edited or paraphrased
Translated into another language
Generated before watermarking was introduced
Produced using an unsupported model
Generated using a different AI system
OpenAI specifically warns against treating the absence of a watermark as definitive proof of human authorship.
What Does This Mean for ChatGPT Users Outside the EU?
For ordinary ChatGPT users outside the European Union, the announcement does not mean that text watermarking is becoming a global default immediately.
OpenAI says it will initially introduce watermarking for eligible ChatGPT and Codex text output in the EU.
The company is taking a regional approach so that it can learn from real-world use and feedback before deciding how its approach should evolve.
For users in countries such as New Zealand and India, the EU rollout therefore does not automatically mean that their normal ChatGPT conversations will receive the same watermark under this initial policy.
However, the development is still important internationally because it could influence future AI transparency standards.
What About OpenAI API Users?
OpenAI is taking a different approach for API customers.
Starting with the announcement, API customers globally can choose to enable text watermarking for supported models.
The feature remains off by default for the API, allowing organisations to decide whether watermarking fits their own transparency requirements and workflows.
This could become particularly relevant for businesses, education platforms, publishing systems and other services that use OpenAI models to generate content at scale.
Will There Be a Public ChatGPT Watermark Detector?
Not immediately.
OpenAI says access to its text watermark detector will initially be limited to approved researchers and expert organisations.
The company wants researchers to help evaluate the technology, understand its limitations and improve the reliability of text provenance.
This is different from OpenAI’s image and audio verification tools, which remain available for supported content.
OpenAI says the text detector will report whether it detects an OpenAI watermark but will not reveal a user’s prompts or conversations.
What Does This Mean for Writers and Publishers?
For writers, journalists, students, businesses and publishers, the new system raises an important distinction between AI assistance and AI authorship.
Someone might use ChatGPT to brainstorm ideas, create an outline, improve grammar or rewrite a paragraph and then substantially develop the final piece themselves.
A watermark, where applicable, does not measure that human contribution.
OpenAI explicitly says its watermark does not determine ownership, legal responsibility or the extent of human creativity involved in a piece of content.
This means organisations will still need their own policies about when AI use should be disclosed.
Disclaimer: This article is based on information published by OpenAI and is intended for general information only. AI regulations and technology policies can change, so readers should check official sources for the latest requirements applicable to their location or organisation.
Read more Technology Desk insights and AI updates from NZ Indian Insights.
