A new app can find out whether an essay has been written by a human or a chatbot
In today’s tech-oriented world, plagiarism has left everyone worried, not to forget technologies like artificial intelligence and machine learning which have made it even more difficult to figure out whether certain things are the work of a machine or a human.
A similar problem was presented during a computer science lecture last year at the Princeton University where Edward Tian, a senior grad student and his peers were asked to differentiate between AI-generated text and a human written script. Guess what? Many of them got it wrong. This made the 22-year-old Tian wonder how powerful AI has become over the years and how potent, smarter and more accessible it will become in a few more years.
Realising that AI is the future and it is here to stay, Tian decided to find a way to responsibly tackle it; find a tech-based solution to a tech-based problem. During the winter break, he sat in his hometown Toronto and he came up with GPTZero. It is an app that can determine if a text has been written by a person or a chatbot with high precision.
He has deemed his creation as a response to ChatGPT, an AI content generator that emerged in November last year and has ever since been a rage. It is a free online chatbot launched by American AI research laboratory OpenAI as a supervised and reinforced learning technique. It is designed to effectively write almost any content, starting from essays to news articles to marketing plans to computer codes, all from a simple prompt.
Apart from quality concerms, there are two other fears that have arose following the launch of ChatGPT. It may snatch away jobs from writers and editors and other creative personnel and it may be be used by hackers for unethical purposes such as developing toxic malware. This is where Tian’s app comes in to play.
Okay, so here’s how it works. At first, it looks at two variables in a text, namely, perplexity and bustiness and then it assigns a score each to the variables. Considering it is a pre-trained software, the app first measures how familiar it is with the given text. If the familiarity is more, it is more likely to be written by ChatGPT and vice versa. Why? Because human written content is more perplex or random in nature. Next comes the burstiness part. Here, the app checks whether the text is variable or not. Wondering what it means? It simply means that if a text is a combination of short and long sentences, variant words and other fluctuating language nuances, it is more likely to be human-written. In contrast, if it is more levelled and uniform, chances are it is bot-generated.
At present, Tian is working on improving the app. However, he has launched a beta version for public use. In fact, he demonstrated his app via Twitter by using two samples: A hand-written essay of New Yorker magazine and a letter composed by ChatGPT. As you can imagine, it gave out perfect results with less than 2 per cent false positives only.
Since its release, GPTZero has been widely used by academics to combat plagiarism in the field with more than 30000 downloads within a week. Tian however hopes that his app is used for more intensive purposes such as to quickly and efficiently disarm online disinformation practices.