GPTZero Detects AI Paraphrasers

Our core mission at GPTZero is to “preserve what's human,” and improving content creation transparency is one of our strategies that requires us to keep adapting — especially as more powerful LLMs and AI prompting techniques evolve.

Starting today, we are now introducing more robust detection that includes detecting AI bypassing tools and AI “humanizers,” which paraphrase AI content for the purposes of avoiding AI detectors.

What are AI paraphrasers?

AI paraphrasers are tools that are designed to take AI-generated text and “bypass” AI detection, without requiring any human originality or input. In this case, we are not looking at tools that paraphrase human text for the purposes of improving writing quality (e.g. Grammarly and other AI-assisted writing aids) or avoiding plagiarism (e.g. Quillbot).

Instead, we detect tools that take AI-written text and automatically tweak the phrases, sentence structures, and word choices. With more aggressive settings, some tools claim to “humanize” AI-written text in a way that avoids AI detection. 

What are the problems with AI humanizers?

Individuals seeking to portray AI-written content as their own often turn to humanizing paraphraser tools to alter their text “just enough” to avoid detection. Some people might not see the ethical or qualitative distinction between an AI bypasser and another AI writing aid.

However, unlike Grammarly and other writing aids, these AI paraphrasing tools are not intended to “improve” the quality of human writing. Their entire goal is merely to bypass potential detectors. In many cases, AI paraphrasing tools actually degrade the quality of the text by introducing grammar and spelling errors or creating awkward synonyms and phrases. Sometimes, this degradation is not noticeable to the reader, but other times, it’s obvious.

Here is an example of text run through paraphrasing where degradation occurs:

Original Text (Generated from GPT-4o): “Income inequality is a persistent global challenge, characterized by disparities in wealth distribution among different segments of society. This socio-economic issue not only affects the quality of life for individuals but also hampers economic growth and social cohesion.”

Paraphrased Text (From a popular bypasser tool): “Income inequality remains a issue marked by variations, in how wealth is spread across various societal groups.This economic and social problem impacts individuals well being. Also hinders both progress and community unity.”

The new text has several grammatical errors, from comma splices to punctuation formatting errors and sentence fragments. In short, the writing is worse, not better, because all the tool cares about is avoiding "AI" labels from detectors like ours.

Maintaining integrity in writing is essential to ensure a level playing field, especially in academics where authentic work is instrumental in judging students' progress and understanding. AI bypassers undermine this learning experience and degrade the value of a student’s education. Furthermore, most of the bypassing tools are paid-only, giving an unfair advantage to individuals who are able to afford these tools and exacerbating equity issues in education.

What is GPTZero’s solution to AI bypassers?

GPTZero has developed an innovative solution that can not only distinguish between text that is human or AI-generated but also flag AI-paraphrased content. Our approach works by first:

• detecting if the modified text is originally AI-written, and only then
• analyzing if the content has been altered using other writing tools to appear human.

Here is an example of how we detect the same AI-generated text that was humanized using popular bypassers compared to other detectors currently on the market.

How accurate is GPTZero at detecting AI paraphrasing?

In our beta phase, our solution successfully detects AI-paraphrased text generated using certain popular AI bypassing tools with ~95% accuracy. Our model is industry-leading in capturing bypassing attempts, but like all current detection, it is not perfect; in less than 1 out of every 1000 (<0.1%) cases, we may mislabel human text as coming from a paraphraser.

Like all current AI detection, our results are probabilistic and should not be used to academically punish or be considered as 100% absolute proof of intent. We encourage writers to adopt our AI authorship tools to transparently verify their own human writing. 

We aim to continuously improve our model to enhance our coverage to all the prevalent AI bypass tools in the market. We're committed to keeping up with detection to prevent bypassers from introducing poor quality writing and more inequity into education.