Watermarks will now be applied to ChatGPT's AI-generated text; OpenAI has launched 'textGrain'..

If you copy and use text generated by ChatGPT for note-taking or other tasks, it may soon become easy to identify. OpenAI has developed a new invisible watermark system called "textGrain" for AI-generated text. This watermark will not be visible to the human eye, but specialized detection tools will be able to detect its presence. Currently, OpenAI is implementing this on eligible text outputs from ChatGPT and Codex within the European Union (EU). How exactly does this watermark work? Can it be removed? Will it affect the quality of ChatGPT's responses? Find out all the details here...

**Invisible Watermark to be Added to ChatGPT Text**
In the coming weeks, OpenAI plans to add an invisible watermark to eligible text outputs generated by ChatGPT and Codex in the EU. This watermark will not be visible to the reader; the text will appear just like standard AI-generated content, without any distinct marks, extra spaces, or odd characters.

The objective is to enable machines to detect whether or not OpenAI's watermark is present in a given text. For now, the rollout is limited to the EU. OpenAI has not announced plans to make this the default for ChatGPT users worldwide at launch; the company first wants to understand its real-world usage and gather feedback within the EU.

OpenAI has named this watermarking technology "textGrain." Instead of placing a visible mark on the text, it introduces subtle statistical changes to the way the AI ​​model selects words. When the AI ​​writes a sentence, it has multiple potential options for the next word or token. TextGraph generates a specific type of statistical signal within the patterns of word choices. While humans cannot typically detect this, OpenAI's detector can identify this pattern within the text.

Notably, the system does not insert hidden characters, invisible spaces, or distinct watermark tokens into the text. Instead, the watermark resides within the statistical patterns of word selection.

**Detectability even after copy-pasting**
A key feature of TextGraph is that the watermark is embedded in the statistical patterns of the text's wording. Consequently, simply copying and pasting the text from one location to another does not automatically erase this signal.

However, this does not mean that every piece of AI-generated text will always be detected. OpenAI acknowledges that factors such as text length, subject matter, language, and subsequent editing can influence detection.

No... according to the company, TextGraph's detector is designed solely to determine whether or not an OpenAI watermark is present in the text. It does not reveal which user generated the text, what prompt was used, or the specific conversation from which the text originated. Nor does the system indicate the extent of human contribution to the final text or identify its owner. In other words, while the watermark serves as an indicator of AI-generated content, it does not, in itself, prove the author's identity or establish ownership of the text.

**How ​​effective is it for texts of 200 and 400 tokens?**
OpenAI has conducted various tests to understand the limitations of TextGraph. According to the company, targeting a 1% false-positive rate, the watermark was successfully detected in approximately 80% of cases for texts containing around 200 tokens. However, for texts containing around 400 tokens, this rate stood at approximately 95 percent. These figures are derived from evaluations conducted on specific types of content, such as responses related to psychology. This makes it clear that the shorter the text—or the more limited its vocabulary—the more difficult it may be to detect the watermark.

TextGen is not equally effective across all subjects. In OpenAI's tests, detection performance was notably weaker in subjects like mathematics. One reason for this is that math-related responses offer less flexibility regarding word choice or phrasing compared to general topics. Consequently, the AI ​​has less scope to embed a watermark signal within its choice of words.

**Watermark effectiveness diminishes with minor editing**
OpenAI's tests also revealed that watermark detection can rapidly weaken if the text is edited. For a text of about 400 tokens, replacing just 10 percent of the words with synonyms caused the detection rate to drop from approximately 92 percent to 66 percent.

If 25 percent of the words were changed, the detection rate fell to around 17 percent. This implies that while TextGen can assist in identifying AI-generated text, it cannot be considered a system capable of catching AI content in every scenario.

**Detection is difficult after translation**
Identifying the watermark can also prove challenging if the AI-generated text is translated into another language. Similarly, very short texts, heavily edited content, material generated by other AI models, or text created before the watermark was implemented may evade detection. Therefore, the company clarifies that the absence of a watermark does not prove the text was written by a human.

Disclaimer: This content has been sourced and edited from Amar Ujala. While we modified it for clarity and presentation, the original content belongs to its respective authors and website. We do not claim ownership of the content.