In early 2023, computer science student, Edward Tian, from Princeton University developed the GPTZero app to analyse the likelihood that a passage of text is written by AI. GPTZero, while still in the development phase, could help universities counter the misuse of ChatGPT in classrooms. ChatGPT has attracted massive attention since its launch on 30 November 2022 and millions of people, including students, have used it.  Over the last few weeks, there has been mounting concern in the education sector that students are using ChatGPT and other AI tools to cheat on assessments.

ChatGPT and GPTZero

A refresher on ChatGPT

ChatGPT is a generative AI program that can produce natural language outputs in response to user prompts. ChatGPT has experienced huge user uptake since its launch in 2022 - more than a million people used the tool within five days post launch. However, since its launch concerns about “plagiarism” have arisen from the university sector. ChatGPT can write basic essays and emulate the style of established writers. This revolutionary technology has forced educators to discourage students from using the software and rethink their methods of assessments.   

Case studies

Australian universities and lecturers have expressed concerns over the use ChatGPT. For example, The Guardian reported that in one class at Deakin University a lecturer detected the use of bots in almost one-fifth of assessments, sparking concerns that the use of artificial technology to cheat in exams is widespread. Sally Brandon, an associate communication lecturer at Deakin University, reported that of 54 summer postgraduate assessments 10 had “significant, detectable bot assistance” and this was the highest rate in the five years she has used software to detect bots.

The NY Times reported that Antony Aumann, a Professor of Philosophy at Northern Michigan University discovered that ChatGPT was used to generate one of the best essays submitted for his class.  Alarmed by his discovery, Professor Aumann has since transformed the essay writing component of his course - he plans to restrict computer use and require students to write the first draft of the assessment in the classroom.

Recently, the ABC reported that New York City school officials have blocked access to ChatGPT to stop students from generating essays, while certain Australian universities are reinstating "pen and paper" exams and beefing up cheating detection measures to address these new issues.

ChatGPT and copyright infringement

Usually, copying your homework from another source would amount to copyright infringement (as well as a breach of school/university code of conduct).  However, according to the terms of use published on OpenAI’s website, users are assigned all rights that subsist in ChatGPT’s output to their prompts, and therefore the user owns any intellectual property rights in both their input and the output created by ChatGPT. It’s also questionable whether copyright subsists in ChatGPT’s responses in the first place – the responses may not qualify as “original” works as they are not created by a human author.  Submitting a ChatGPT generated essay may be a breach of university or school academic policies, but it does not amount to copyright infringement.

Enter GPTZero

22-year-old Princeton student Edward Tian designed GPTZero. It is an application made to quickly and efficiently determine whether a passage, such as an essay, was written by a person or an AI program.

GPTZero uses “perplexity” and “burstiness” to determine whether text has been generated by a bot. Perplexity is how random the text is in a sentence and whether the way a sentence is constructed is unusual or surprising to the app. Burstiness compares these sentences to one another, determining their uniformity. This is helpful because human writing has more burstiness as we tend to write with more sentence variation.

GPTZero is currently in beta. While still in development, the app has the potential to meet a need within schools and universities as AI technologies improve. Universities can use this app (or one of the other bot detection apps on the market) as a standard procedure for submitting coursework and vetting admissions essays to ensure that applicants have done the work themselves.

However, ChatGPT could easily be taught to evade GPTZero using a generative adversarial network, perhaps even using a variation of GPTZero as the discriminator network.

An experiment - can GPTZero detect something written by ChatGPT?

We asked ChatGPT – “How can you tell if something is written by AI?”.  Here is its response:

“There are several ways to determine if something was written by an AI model like GPT-3. One way is to look for patterns or inconsistencies in the text that may indicate that it was generated by a machine. For example, AI-generated text may have a lack of coherence, repetition of phrases, or inconsistencies in style or tone. Additionally, an AI-generated text may lack the complexity of the human language, such as idiomatic expressions or cultural references.

Another way to tell if something is written by AI is by analysing the text's grammatical structure, syntax, and vocabulary. AI-generated text may have less nuanced grammar, use of unusual words and syntax that deviate from the standard language.

It's also possible to use software tools to detect AI-generated text, for example, OpenAI's GPT-3 has a API that can be used to generate text, so you could use it to verify if the text is generated by it.

It's important to note, however, that AI-generated text can be difficult to distinguish from human-written text, especially if the AI model has been fine-tuned on a specific task.”

We then fed ChatGPT’s response into GPTZero, to see if GPTZero could detect that the response had been written by AI.  Its analysis included some indicators that the text was more likely to be generated by AI, but ultimately GPTZero’s answer was that more data was needed to determine if the text was human generated or generated by AI.

1-0 to ChatGPT! (but it was close).

We then asked ChatGPT to write a 1000 word essay on Macbeth by William Shakespeare.  We will spare you the essay (which was quite good by the way), but GPTZero came back with a strong indication that the text was most likely to be AI generated.

That’s 1-1!

Final thoughts on GPTZero and ChatGPT

For the time being, it appears that GPTZero is a useful tool to help spot AI generated text, but it may depend on the length and type of response.  Both ChatGPT and GPTZero are continuing to be developed, so it will be very interesting to see how the competition between ChatGPT and GPTZero unfolds.

Authors: Michael Caplan, Claire Harris, Ali Khan