AI Detection Tools Are More Likely to Flag Non-Native English Speakers, According to Report
A pattern of false positives is happening more often for teachers using AI detection tools. In the past few months, instructors and lecturers have discovered that AI detectors are biased against non-native English speakers compared to native speakers, a new report finds.
In the study entitled "GPT detectors are biased against non-native English writers," researchers concluded that plagiarism checkers like Turnitin, as well as seven other AI detectors, flag the text written by non-native speakers a whopping 61 percent of the time. What's more troubling is that 20 percent of papers that were run through these platforms the incorrect assessment was unanimous. When compared to works from native speakers, such mistakes were almost never found.
"Our findings emphasize the need for increased focus on the fairness and robustness of GPT detectors, as overlooking their biases may lead to unintended consequences, such as the marginalization of non-native speakers in evaluative or educational settings," the authors of the report wrote.
Roughly 91 Test of English as a Foreign Language (TOEFL) essays were examined from a Chinese forum and 88 U.S. eight-grade essays from Hewlett Foundation’s ASAP dataset to come up with the findings. The results were discouraging for GPT detector makers.
ALSO READ:
1 in 4 Filipinos Is Ethnically Tagalog. Here's a Breakdown of Ethnicities in the Philippines
After Resigning From Google, the Godfather of AI Has a Chilling Warning for the Future
"Upon closer inspection, the unanimously identified TOEFL essays exhibited significantly lower text perplexity. Here text perplexity is a measure of how 'surprised' or 'confused' a generative language model is when trying to guess the next word in a sentence," the research claims. "If a generative language model can predict the next word easily, the text perplexity is low. On the other hand, if the next word is hard to predict, the text perplexity is high."
GPT detectors use text perplexity as a measurement to find AI-generated text. Unfortunately, because non-native English speakers use simpler linguistic expressions and sentences, they are wrongly identified as AI-generated. This means that detectors inherently discriminate against international students. For educators in other countries, this can spell trouble for them, too, when grading papers. The technology seems to have issues with recognizing linguistic diversity.
Even the creators themselves are seeing this troubling trend. OpenAI shut down its AI detector last month after low accuracy rates. OpenAI shut down its AI detector at the end of July because of low accuracy, and Quill.org and CommonLit did so, as well. Turnitin, on the other hand, remains the only platform to double down.
