2 Ways Students Can Use ChatGPT to Write Urgent Essays without Detection by GPTZero

In this article, we bust a developing myth most people think will help them use ChatGPT and avoid detection from the viral GPTZero and we show you an evidence-based method that might help you 

Edward Tian, a Princeton University student of computer science, has developed GPTZero, which most people now believe can accurately analyze characteristics embedded in AI-generated texts to determine if it was AI-generated.

Admittedly, Tian’s GPTZero has astounding capability when compared to other AI-detectors such as Copyleaks, Hugging Face, Originality.ai, among others.

Some argue that GPTZero‘s results are ‘mediocre at best’, but after a detailed investigation, we conclude that they did not consider the intended use.

Importantly, if someone were to put your AI-generated content with an overall length of over ten lines (as recommended), the sophisticated algorithm will definitely get you busted!

How does GPTZero work?

The fundamental means of measure for GPTZero is perplexity and burstiness.

Perplexity measures how much the computer model likes the text. The lower the perplexity index, the more likely the model is “in love’ with your text and this means trouble for you.

Burstiness measures the non-common items that usually appears in random clusters that will certainly appear over time in case the text was generated by a human. If the graph appears quite level or linear, as shown below, it might imply trouble for you.

In the end, after approximately a minute of analysis, a combined analysis of the two indicators gives results, which could be

1. “Your text was likely AI generated”

2. “More data may be needed to determine if your text is human or AI generated. Try inputting more text”

3. “Your text was likely human generated”

We tried and tested this by feeding GPTZero with over 100 ChatGPT generated content such as essays and conclusions and the algorithm consistently concluded that “the text was likely generated by AI.”

Comparatively, each and every human generated content was labelled “Your text was likely human generated”

Has ZeroGPT changed the game?

Some people argue that spinning the content using tools such as Quillbot will generate undetectable text because most words will have been paraphrased.

However, unlike other plagiarism detection tools like Turnitin, this algorithm does not rely on the originality of the words per se, rather on a pattern.

Simply put, if you were to use non-AI based paraphrasing tools to spin/rephrase your words, you will be more likely to first loose the meaning/gist of the content before you get to evade GPTZero detection, which is obviously not what you desire.

Additionally, if you were to think a bit out of the box and use AI-based paraphrasing tools such as Quillbot, the purpose will be self-defeating because you will be generating another AI-based content that exhibits these patterns that GPTZero is looking for.

The ‘solution’ 

Paradoxically, a solution lies in “fighting fire with more fire”; using ChatGPT (GPT-2 based) to counter GPTZero (GPT-2).

In one occasion, we instructed ChatGPT to increase the perplexity and burstiness of two paragraphs, which were 100% generated by ChatGPT. Our aim was to evade detection by GPTZero. The results were quite remarkable primarily because ChatGPT understands perplexity and burstiness.

For a text that was initially labelled “Your text was likely AI generated,” the newly generated text was labelled “more data may be needed to determine if your text is human or AI generated. Try inputting more text”.

As little as it may seem, this is a giant leap in the scale of GPTZero detection since it is also possible for a 100% human generated text to be labelled “more data may be needed to determine if your text is human or AI generated. Try inputting more text” under similar circumstances.

A user on YouTube confirmed these results, though he creatively went farther by combining two paragraphs that had applied the same technique; the result was exponential.

He concluded that by combining two individual paragraphs that were labelled “more data may be needed to determine if your text is human or AI generated. Try inputting more text,” GPTZero labelled the combined paragraphs as “Your text was likely human generated.” This is a win!

The findings are both reasonable and plausible because it is likely that there was an increase in burstiness which seems to be measured as deviation from “average perplexity”.

Therefore, by combining the two separately rephrased paragraphs, deviation from the mean value increased; hence avoiding detection by GPTZero.

Is time to celebrate?

Unfortunately, the GPTZero evading methods, though possible, are not rewarding because as a student or ChatGPT user, what you really need is reliability.

Reliability is trusting that every time you click on ChatGPT’s “Submit” button, you will get indisputable results without a need to regenerate.

Another factor is to be sure that there won’t be a need to visit novel AI-detectors such as GPTZero to check ChatGPT generated essays or texts.

Another possible setback to seamlessly creating your content is that ChatGPT creators (OpenAI) are looking for ways to watermark every single line generated by the chatbot in the near future.

Succinctly, the whole truth is that in the present, we must NOT use ChatGPT to generate content since we can’t ascertain for sure what the future of ChatGPT regulation/detection holds for us.

What’s the reliable ChatGPT alternative?

Obviously, as a student who may need academic assistance, you can always rely on traditional methods such as tutoring to get that grade you desire.

As you wait to get back to the future of AI-generated submissions, let Assignmentpapermaster  help you get that grade you desire, in the meantime.

We offer excellent educational consultancy services at affordable prices. We always guarantee that you’ll get a distinction or an A plus in any discipline.

Over 200 satisfied students can’t be wrong!