Researchers have found that detectors used to identify AI-generated work, such as essays and job applications, can unfairly discriminate against non-native English speakers. The study tested seven popular AI detectors on over 90 essays written by non-native English speakers for the Test of English as a Foreign Language (TOEFL). In the results, over 60% of the TOEFL essays were wrongly flagged as AI-generated. - When essays written by native English-speaking eighth graders in the U.S. were tested, over 90% of them were correctly identified as human-written by the same tools.
- According to the researchers, the bias is likely rooted in the concept of "text perplexity," which evaluates the predictability of words in a sentence.
- Low perplexity is typically associated with AI-generated text, and non-native speakers, who often rely on common words, are more likely to exhibit this predictability.
- The study emphasizes the potential harm this misclassification can cause, including discrimination against non-native English speakers in academic and job settings.
|
