GPTZero

GPTZero is an artificial intelligence detection software developed to identify artificially generated text, such as that produced by large language models.

While GPTZero has received positive coverage for its efforts to prevent academic dishonesty, many news outlets criticized the tool for its false positive rate, especially in academic settings.

History
GPTZero was developed by Edward Tian, a Princeton University undergraduate student, and launched online in January 2023 in response to concerns about AI-generated usage in academic plagiarism. GPTZero has raised over 3.5 million dollars in seed funding.

In the first week of its release, the GPTZero experienced 30,000 uses, which led to a crash. It is supported by the web app company Streamlit, who allocated more server resources in response.

Mechanism
GPTZero uses qualities it terms perplexity and burstiness to attempt determining if a passage was written by a AI. According to the company, perplexity is how random the text in the sentence is, and whether the way the sentence is constructed is unusual or "surprising" for the application. It relies on language models, and the more such models, the higher the chance that a person did not write the text. In contrast, burstiness compares sentences with each other, determining their similarity. Human text is more discontinuous, meaning humans tend to write with more sentence variation than AI.

News website Ars Technica commented that humans can still write sentences in a highly regular way, leading to false positives. The writer, Benj Edwards, went on to state that the perplexity score only concerns itself with what is "surprising" for the AI, leading to instances where highly common texts, such as the US Constitution, are labeled as AI-generated.

Use cases and applications
The academic community has attempted using GPTZero to tackle concerns about AI-generated content for plagiarism. Educational institutions, including Princeton University, have discussed the use of GPTZero to combat AI-generated content in academic settings, with mixed opinions. In October 2023, GPTZero had partnered with the American Federation of Teachers.

Efficacy
In a March 2023 paper named "Can AI-Generated Text be Reliably Detected?", computer scientists Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi from the University of Maryland demonstrate empirically and theoretically that several AI-text detectors are not reliable in practical scenarios.

Tech website Futurism tested the tool, and said that while the "results are impressive", based on its error rate, teachers relying the tool would end up "falsely accusing nearly 20 percent of [innocent students] of academic misconduct".

The Washington Post have noted GPTZero suffers from false positives, emphasizing that "even a small “false positive” error rate means that some students could be wrongly accused [of academic misconduct]".