OpenAI is introducing an invisible watermark for eligible ChatGPT and Codex text generated in the European Union. But this is not a magic “AI detector” that can tell who wrote something, whether the information is true, or how much AI was involved.
OpenAI says it will begin adding an invisible statistical watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks. The company is also allowing API customers worldwide to opt in to text watermarking for selected models.
The move is part of OpenAI's response to the European Union's AI transparency requirements, which require certain AI-generated content to be identifiable in a machine-readable way. The EU's Article 50 transparency obligations became applicable on August 2, 2026. ([OpenAI](https://openai.com/index/eu-text-provenance/?utm_source=chatgpt.com))
But there is an important detail that could easily get lost in the headline: an invisible watermark is not the same thing as proof that a person used AI to write an entire article.
Understanding that difference matters for students, writers, publishers, businesses and anyone who uses AI-generated text.
What Is OpenAI's Invisible Text Watermark?
OpenAI calls its text watermarking technology textGrain.
Unlike a traditional watermark placed visibly across an image, textGrain does not add a logo, special symbol or visible mark to the text. You can read the resulting text normally, and copying and pasting it does not reveal anything unusual to the reader.
Instead, the system changes the statistical pattern of the model's word choices.
When an AI model generates a sentence, it does not simply select the next word from a fixed list. It considers multiple possible words or pieces of words and assigns different probabilities to them. OpenAI's watermarking system subtly adjusts those choices according to a secret pattern.
Across enough text, those small choices can form a statistical signal that a detector can search for.
In simple terms, the watermark is not something hidden inside the letters. It is a pattern created by how the AI chooses its words. ([OpenAI Help Center])
Why Is OpenAI Doing This?
The immediate reason is AI transparency.
The European Union's AI Act includes transparency requirements for providers of generative AI systems. Under Article 50, certain AI-generated or manipulated content must be marked in a machine-readable format so that it can be identified as artificially generated or manipulated.
The European Commission says these rules are intended to address risks involving deception and manipulation and to improve transparency around AI-generated content.
OpenAI's approach is therefore not simply a new feature designed to help people catch students or writers using ChatGPT. It is part of a broader effort to provide content provenance information that can help determine whether content contains a signal associated with an AI system.
Will You See the Watermark?
No.
There will not be a visible “AI-generated” stamp appearing beside every paragraph.
OpenAI says the textGrain watermark is embedded through the statistical pattern of word choices. It does not insert hidden characters, invisible spaces or unusual punctuation into the text.
That means a person reading the text normally would not know whether the watermark is present simply by looking at the words.
The signal is intended to be detected by software designed to look for the watermark.
Who Will Be Affected First?
OpenAI says eligible ChatGPT and Codex text output in the European Union will receive the watermark as the feature rolls out over the coming weeks.
For API customers, the situation is different. OpenAI says customers around the world can opt in to text watermarking for selected models, and it will remain off by default in the API.
This distinction matters because many websites, applications and business tools use AI through APIs rather than through the ChatGPT consumer interface. ([OpenAI](https://openai.com/index/eu-text-provenance/?utm_source=chatgpt.com))
Does the Watermark Prove That ChatGPT Wrote the Entire Article?
No.
This is probably the most important point to understand about the technology.
According to OpenAI, detecting a watermark provides evidence that an OpenAI model likely generated or processed the text. It does not establish who authored the content, who owns it, who is legally responsible for it, or how much of the final text came from AI.
For example, imagine someone writes an article themselves and then asks ChatGPT to improve several paragraphs.
If the resulting text contains an OpenAI watermark, the signal could indicate that an OpenAI model generated or processed some of that material. It would not, by itself, prove that the entire article was written by AI.
OpenAI specifically says its watermark cannot determine the extent of an AI model's involvement. ([OpenAI Help Center])
It Also Cannot Tell You Whether the Information Is True
A watermark is a provenance signal, not a fact-checking system.
If ChatGPT produces a paragraph containing an incorrect claim and that paragraph carries a watermark, the watermark does not make the claim more or less accurate.
Likewise, the presence of an OpenAI watermark does not tell a reader whether the content has been properly researched, edited or fact-checked.
OpenAI's own guidance makes this distinction clear: provenance signals can provide information about origin, but they do not guarantee that content is accurate, unedited, legally owned or presented in the correct context. ([OpenAI Help Center])
What Happens If Someone Edits the AI-Generated Text?
This is where things become more complicated.
The watermark is statistical, so significant changes to the wording can make detection more difficult.
OpenAI says substantial rewriting, paraphrasing or translation can reduce the reliability of detection. Short passages can also be difficult to identify because there may not be enough text for the statistical signal to become strong enough.
OpenAI's documentation also explains that detection performance varies by language. Its testing found different detection rates across the European Union's official languages, showing that the technology does not behave identically in every linguistic environment.
This is one reason it would be misleading to describe text watermarking as a perfect AI detector.
What About Copying and Pasting the Text?
Copying and pasting the text does not remove the watermark simply because the watermark is not stored as an obvious hidden character.
The signal is connected to the statistical pattern of the generated wording.
However, this does not mean the signal will survive every possible transformation. Substantial rewriting, paraphrasing and translation can weaken detection, and very short passages may not provide enough information for reliable detection. ([OpenAI Help Center)]
How Is This Different From Ordinary AI Detectors?
This distinction is important.
Many third-party AI detection systems work by analyzing a piece of text after it has already been written. They use statistical or machine-learning classifiers to estimate whether the writing resembles AI-generated text.
OpenAI's textGrain approach is different because it attempts to place a detectable signal into the generated text at the time the model produces it.
OpenAI says this embedded-signal approach is specifically intended to address the EU requirement for machine-readable identification of AI-generated content.
Does This Mean Every AI Detector Will Now Know When ChatGPT Was Used?
No.
OpenAI's system is specifically designed to detect its own watermark. That is different from creating a universal system capable of identifying every AI model.
A document produced by another AI system would not automatically contain an OpenAI watermark simply because it was generated by artificial intelligence.
Similarly, a piece of human-written text would not become “AI-generated” merely because it contains wording that sounds similar to something an AI model might produce.
That distinction is particularly important when AI detection is used in education, employment, publishing or other situations where an incorrect accusation could have serious consequences.
What Does This Mean for Writers and Publishers?
The development could make AI provenance increasingly important in digital publishing.
Publishers may eventually have more ways to distinguish between content generated by AI systems and content that does not contain those systems' provenance signals.
But publishers should not treat a watermark as a replacement for editorial review.
A publication still needs to determine:
- Whether the information is accurate.
- Whether important claims have reliable sources.
- Whether the article adds original value.
- Whether the content has been properly edited.
- Whether the article is useful to the intended audience.
- Whether the writer or publication has represented the use of AI honestly.
In other words, provenance can answer part of the question “Where might this text have come from?” but it cannot answer the much larger question “Is this a good and trustworthy article?”
Can the Watermark Identify the Person Who Used ChatGPT?
No.
The watermark does not function as a personal identification system.
Finding an OpenAI watermark in a passage does not automatically tell a detector:
- Who entered the prompt.
- Which person copied the text.
- Who edited it afterward.
- Who published it.
- Who owns the final article.
OpenAI says the watermark itself does not establish authorship, ownership or legal responsibility. ([OpenAI Help Center])
Why This Technology Is Still Important Despite Its Limitations
The limitations do not make the technology useless.
Digital content is increasingly produced with AI, and platforms, publishers and researchers need better ways to understand where some content originated.
A machine-readable provenance signal can provide one additional piece of evidence that would otherwise be unavailable.
The key is to use that evidence appropriately.
It should be treated as one signal among several, rather than as a final verdict about authorship or authenticity.
That distinction is particularly important because OpenAI itself describes text watermarking and detection as early technologies with significant limitations. ([OpenAI])
Frequently Asked Questions
Will ChatGPT text be watermarked everywhere?
Not initially. OpenAI says eligible ChatGPT and Codex text output in the European Union will receive the watermark as the rollout begins. API customers globally can separately opt in for selected models.
Can I see the watermark in ChatGPT text?
No. The watermark is designed to be invisible to ordinary readers and is created through statistical changes in the model's word choices.
Does an OpenAI watermark prove that ChatGPT wrote an entire article?
No. It can indicate that an OpenAI model likely generated or processed text, but it does not establish the amount of AI involvement or the identity of the human author.
Can editing remove the watermark?
Significant rewriting, paraphrasing or translation can make detection less reliable. OpenAI also notes that short passages can be difficult to detect reliably.
Does an AI watermark prove that information is false?
No. A watermark says something about content provenance, not factual accuracy. AI-generated information still needs to be checked against reliable sources.
Does the watermark identify the person who used ChatGPT?
No. OpenAI says the watermark does not establish authorship, ownership or legal responsibility.
The Bigger Picture
OpenAI's move is part of a much larger shift in how the internet handles AI-generated content.
As generative AI becomes part of everyday writing, the challenge is no longer simply creating AI-generated text. It is also becoming important to understand where content came from, how it was produced and how much confidence we should place in it.
Invisible watermarking could become one useful part of that system.
But it should not be confused with a perfect AI detector.
For readers, the practical lesson is simple: don't judge an article's reliability solely by whether it appears human-written or AI-generated. Check the quality of the information, the evidence behind important claims, the sources being used and whether the publisher demonstrates genuine editorial responsibility.
And for publishers, the arrival of AI provenance technology makes a strong editorial process even more valuable not less.
AI can help create content. But trustworthy technology journalism still depends on research, verification, explanation and human judgment.