Diploma can't be generated — ChatGPT leaves "marks" in texts

The Rumi platform team discovered that the latest OpenAI models — GPT-o3 and GPT-o4 mini — invisibly insert special "marks" into the generated text. They resemble ordinary spaces, but have different Unicode codes and thus can indicate the machine origin of the material. It's easy to remove such markers, so this is probably just a short-term experiment, not a permanent tool for verifying authorship.

Rumi writes about it.

What marks does ChatGPT leave in the texts it generates?

Rumi notes that hidden characters appear only in long responses, for example, when GPT-o3 is asked to write a "full essay". This is mainly the Narrow No-Break Space (NNBSP, code U+202F), which looks like a regular space but has a different ASCII sequence. In older versions, such as GPT-4o, this watermark is not present.

You can check for markers by copying the text to an online character viewer or opening it in an editor like Sublime Text, which displays invisible characters. The pattern of NNBSP appearance seems to be systemic rather than random, so experts assume that it was deliberately implemented.

At the same time, the marks are preserved after copying to Google Docs and similar editors, so teachers or employers could potentially find an "invisible" trail if a student or author copies answers from ChatGPT in their entirety. Removing the markers is easy: just replace the special characters with regular spaces using the Find and Replace feature in your text editor.

Rumi believes that hidden markers will not be a reliable long-term verification method. Instead, they suggest focusing on a process-oriented approach to writing: tracking the development of ideas in several drafts, using AI tools while working on a task, and recording the stages of the creative process. This format, in their opinion, is more effective in maintaining integrity and developing students' AI literacy.

As a reminder, OpenAI loses money every time you add "Please" or "Thank you" to a ChatGPT request. As it turned out, being polite with the chatbot consumes a little more electricity, which has already cost the company tens of millions of dollars.

We also wrote that ChatGPT has received the significant memory upgrade that will allow it to analyse all previous dialogues with the user. The chatbot is able to take into account the context of previous conversations.