Edward Tian, a 22-year-old senior at Princeton University, built an app called GPTZero that helps detect whether a text is written using AI tools like ChatGPT.
Though ChatGPT is still far from being a completely accurate and reliable source of information, it has already got teachers worried about students turning in essays written using it, and that too within seconds!
Schools in New York City recently impeded ChatGPT from their devices and internet networks concerning issues of AI plagiarism by students. This might be an essential step taken to prevent students from cheating. But it is not a very effective measure as students can still use ChatGPT and other similar tools outside the school premises.
Educators around the world who are worried about the unethical use of AI-writing tools, now have a tool of their own to detect machine-generated text. Hence, institutes and universities can now easily start using GPTZero to detect whether a text is human-written or not.
The app, GPTZero was diligently built to identify whether the text is written by a human or artificial intelligence app like ChatGPT—the chatbot that has ignited fears for its potential unethical uses in academia.
“The app can quickly and efficiently detect whether an essay is ChatGPT and other AI or human-written,” Tian said in a tweet.
Though the app is still in the development phase, it is already getting overwhelmed with traffic because many users want to test whether the text they’re entering is human-written or not. GPT Zero can be used on the developer’s website and Streamlit.
More than 30,000 people used the app within the first week of its launch which caused it to crash due to high traffic. Streamlit, the free forum that hosts GPT Zero, has since trekked in to help Tian with additional resources to regulate the web traffic.
How does GPTZero work?
GPTZero works by analyzing the text for two indicators, “perplexity” and “burstiness”, to decide the complexity of the text and if it is human-generated or not.
Burstiness: The diffusion of sentences—based on the notion that humans tend to blend long and brief sentences with rhythmic patterns while AI sentences are more likely to be uniform.
Perplexity: A measure of how nicely a language model predicts a text piece. It involves the randomness of the text to determine if it is combined by AI tools.
Lower perplexity implies that the text is easy to foretell; hence, it sounds more like a machine. Humans are rather wise and theatrical, they tend to compose text with higher perplexity.
The founder of the app, Tian, has publicized multiple cases in which his algorithm was accurate. However, he conceded that his app isn’t foolproof and that if the piece of text is very random, the app might toggle; as some users have reported when testing it out.
Tian wants to improve the app with time, and he’s still working on increasing the accuracy of the beta version. Developing software that can distinguish human-written text from the machine-generated text can bring transparency to using tools like ChatGPT for multiple purposes.
To catch things ostensibly, we tried the app ourselves to test the results. Here is the text we entered (human-written):
‘The Apple MacBook Air M2 presently is number one. Mainly, because it is a substantial option for many. It has got everything you’ll ask for—reliable enactment, long battery life, and a configuration that can be used by many people. The latest MacBook Air costs around $1,199 but you can always opt for the 2020 MacBook Air M1 as a cheaper alternative. However, if you’re seeking a valuable option then HP’s Pavilion Aero 13 is an outstanding and lightweight laptop costing around $550, pretty reasonable.’
The analysis rightly indicated that it is human-written content with high perplexity score. The tool was super fast in detecting that the text is likely to be human-written!
We even tested ChatGPT text and it successfully detected that. However, AI detection was somewhat inaccurate with shorter texts.
Is GPT Zero good enough for Educators to detect AI-written text?
Well, it depends on how OpenAI tools like ChatGPT evolve over the coming months.
The app created by Edward Tian will be helpful to assist teachers and educators in identifying artificial intelligence generated text, but the results might be variable and depend on the length of the text as well as other factors.
ChatGPT was released on November 30, 2022, and crossed 1 million users within 5 days. AI-writing tools might become more potent and accurate with time. ChatGPT might be able to write more human-like and naturally with future updates and development.
For scenarios like these, models like GPT Zero would have to emerge with more innovative ways to detect artificial intelligence written texts.
Tian isn’t alone against the misuse of ChatGPT. OpenAI has also signaled an obligation to preclude AI plagiarism. Last month, Scott Aaronson, a researcher presently concentrating on Artificial Intelligence security at OpenAI, disclosed that the company has been performing on a way to “watermark” GPT-generated text with an “unnoticeable secret signal” to determine its basis.
Furthermore, the open-source AI community Hugging Face has created a tool to detect whether the text was produced by GPT-2 (an older version of the OpenAI model) after a philosophy professor in South Carolina caught a student submitting AI-written work.
Tian is not opposed to the use of AI tools like ChatGPT.
He said that GPT Zero is “not meant to be a tool to stop these technologies from being used”.
“But with any new technologies, we need to be able to adopt it responsibly and we need to have safeguards”