New ChatGPT Models Seem to Leave Watermarks on Text

What’s Going On

Our team at Rumi has discovered that the newer and more advanced GPT-o3 and o4-mini models appear to be embedding special character watermarks in generated text. In our testing, the watermarks were only added to longer responses for example when asking GPT-o3 to “Write a full essay on the Department of Education”. These watermarks consist of special Unicode characters (primarily the Narrow No-Break Space) that look identical to regular spaces but have different ASCII-codes.  In our testing we did not observe watermarks in older models such as GPT-4o.

Users can detect these hidden characters by pasting text into online tools like this or text editors like Sublime Text, which reveal these normally invisible markers. The pattern of these characters appears to be systematic rather than random, suggesting an intentional implementation.

‍

This comes on the heels of recent announcements of OpenAI testing with watermarks on images. OpenAI has made no official announcement about this feature, likely because publicizing it would undermine its effectiveness in detecting plagiarism. While potentially useful for identifying AI-generated content, the watermarking is relatively easy to circumvent once users are aware of it – a simple find-and-replace operation can remove these special characters.

Why is this Important

ChatGPT is now free for students until the end of May (chatgpt.com/students). Many, if not most, students will likely use it for their work, particularly the newer and more advanced models. The timing is significant—coming during final projects and papers.

Students who are unaware of these invisible markers and directly copy/paste ChatGPT-generated content as their own work may face consequences once word spreads that instructors can use specialized tools to detect these characters. However students that become aware of this change will have a significant advantage in incorporating entire ChatGPT answers as their own answers, hence further exacerbating the imbalance for penalizing student who use or don’t use AI.

What is Watermarking with Special Characters

This approach embeds special Unicode characters such as Narrow No-Break Space (NNBSP) (Unicode U+202F) into generated text.

These special characters look identical to regular spaces in normal word processors and browsers, making them impossible to distinguish to the naked eye. However, they can be easily revealed using:

• Online tools like SoSciSurvey’s character viewer
• Code editors such as Sublime Text or Visual Studio Code
• Simple text analysis tools that identify non-standard Unicode characters

When revealed, these characters create a distinctive pattern that clearly identifies text copied directly from newer ChatGPT models without any modification.

‍

How to Remove Watermarks

To remove these watermarks, open any text editor that displays special characters and replace them with their standard counterparts. See video below.

‍

Pros and Cons of this Approach

Unlike AI detectors which have proven to be inaccurate, this approach’s main benefit is that special characters can indicate text copied from ChatGPT. The chance of false positives—unfairly accusing someone of cheating—is practically zero since students wouldn’t naturally use Narrow No-Break Space (NNBSP) characters in academic papers.

However, the main drawback is that this is likely a temporary measure, as students will quickly become aware of the method and can easily bypass it by using tools to replace all watermarks with standard characters. Also, this may create false confidence for instructors in their ability to “catch” AI generated content.

What’s Next and The Long term Approach to Identifying Authorship

This feature is likely only in the testing phase. If these issue receives widespread attention, OpenAI might remove this watermarking feature completely similar to when they quietly shutdown their AI detector over inaccuracies. However ChatGPT, to the the best of our knowledge, is the first major LLM to implement this feature.

Rather than relying on easily bypassable watermarks, we at Rumi we advocate for a process-focused approach to student writing that:

• Tracks development of ideas through multiple drafts and checkpoints
• Incorporates customizable AI for assignments
• Emphasizes reflection on research and writing choices
• Enables real-time group collaboration and peer reviews

This method not only addresses academic integrity more effectively but also develops AI literacy skills that serve students beyond the classroom.