One of the latest issues that writers, professors, and other professionals are facing is the rise of AI-written content. Getting an entire article or essay with a simple prompt could easily cause cheating issues and loss of job opportunities. Hence, the creation of tools that can identify AI-written text like GPT-2 Output Detector and Originality.ai. One platform that adds to this list of tools is GPTZero.
Edward Tian made GPTZero and launched it on January 2, 2023. The platform aims to detect AI-written essays, articles, and other works. Ideally, educators could use this to prevent students from resting on their laurels and letting the AI do the work for them.
How it works
The main factors the GPTZero uses to differentiate human and AI-written content are the Total and Average Perplexity. They define perplexity as “the randomness of the text.” The higher the perplexity, the lower the chance that an AI will generate it.
In the context of GPTZero, this refers to the overall perplexity of a certain work. The higher it is, the more likely it is that a human wrote it. However, the longer a work is, the lower its total perplexity gets. Of course, you can’t say an essay or article is AI-written just because it was long. This is where average perplexity comes in.
Accompanying the total perplexity is the average perplexity. This is based on each sentence rather than the entire work. In GPTZero, getting the average of all sentences will give a clearer value of just how perplexing a work is, no matter its length.
Of course, humans can also write sentences with low perplexity. However, GPTZero’s research has shown that humans are naturally bound to have some randomness in their writing. Meanwhile, current AIs write with a constantly low perplexity.
Spotting AI-written content
I tried inputting two groups of texts to test GPTZero’s accuracy. One group of texts is made up of previous works I wrote. The other group was written by ChatGPT, a popular AI writing tool. I gave the AI the same topics/prompts from the works in the first group to compare them better. To my surprise, the platform got it right every single time. It correctly determined which texts were written by an AI. This is great news for educators out there who are worried about their students relying on AI writing their papers.
Can it mistake human-written content for AI-written content?
It might be possible, especially with more basic written works. To test this out, I searched online for a 4th-grade-level essay, copied a part of it, and fed it to GPTZero. Unsurprisingly, it had a relatively lower perplexity score, but still enough to be judged as human-written. Not bad for something that’s still pretty new. However, even basic written works might trip up the app.
Of course, AI writing isn’t all bad. Certain companies or professionals might prefer using an AI tool as a guideline to ease the writing process. AI can certainly help humanity out if we learn to use them correctly. You can try out GPTZero for yourself here.
Photo credits: The feature image is symbolic and has been taken by Marten Newhall. The screenshots in the body of the article were taken by the author for TechAcute.