As the use of chatGPT has grown over the past year, there have been growing concerns about its impact on academia and the creative industries. The popularity of the tool has coincided with a growing number of tools aimed at detecting AI-generated text.
A recent study conducted by Stanford University sheds light on potential issues with these tools. The findings reveal a prevalent tendency among these programs to erroneously classify articles penned by non-native language speakers as being of AI origin.
Inherently unreliable
As such, the researchers caution against the utilization of such AI text detectors due to their inherent unreliability, thereby highlighting potential adverse consequences for individuals, including students and job applicants.
“Our current recommendation is that we should be extremely careful about and maybe try to avoid using these detectors as much as possible,” the researchers explain. “It can have significant consequences if these detectors are used to review things like job applications, college entrance essays or high school assignments.”
In order to assess the performance of seven prominent GPT detectors, researchers conducted an empirical investigation. They subjected 91 English essays, authored by non-native English speakers as part of the renowned Test of English as a Foreign Language (TOEFL), to the scrutiny of these detectors.
Alas, the results were disconcerting, as the detectors incorrectly classified over half of the essays as products of AI, with one detector erroneously attributing nearly 98% of these essays to AI authorship. Conversely, the detectors exhibited a higher degree of accuracy in recognizing human-generated text, successfully identifying over 90% of essays authored by eighth-grade students from the United States.
Text perplexity
The authors suggest that many algorithms use an approach known as “text perplexity”, which gauges the level of linguistic surprise in the words used in an essay. This is then the basis for its classification.
“If you use common English words, the detectors will give a low perplexity score, meaning my essay is likely to be flagged as AI-generated. If you use complex and fancier words, then it’s more likely to be classified as human written by the algorithms,” they explain.
The researchers believe this is likely to be because systems, like ChatGPT, are designed to generate text with relatively low perplexity as this is believed to be how humans talk. As such, the simpler word choices used by non-native English speakers risk placing them in the same bracket.
Raising the complexity
This was demonstrated when the researchers input a number of human-authored TOEFL essays into ChatGPT, and asked it to edit each essay to include more sophisticated language. When the resultant essays were checked, they all tended to pass as human-written.
“We should be very cautious about using any of these detectors in classroom settings, because there’s still a lot of biases, and they’re easy to fool with just the minimum amount of prompt design,” the authors explain.
Similarly, the researchers believe that other sectors that are using AI detectors, such as search engines, should be wary that they don’t inadvertently silence non-native English speakers.
