Why GPT Detectors Can't Solve AI Cheating Issues

2023-05-22

After the high-profile launch of ChatGPT, at least seven developers or companies have fought back with AI detectors. In other words, their AI can determine whether the content is written by another AI. These new algorithms are recommended for educators, journalists, and other relevant individuals as tools to detect cheating, plagiarism, and incorrect or false information. However, according to a new paper by scholars from Stanford University, there is a significant problem: the detectors are not particularly reliable, especially when the authors are non-native English speakers. While these detectors perform "almost perfectly" in evaluating essays written by eighth-grade students in the United States, they classify over half (61.22%) of the TOEFL essays written by non-native English speakers as AI-generated. Even worse, according to the study, out of the 91 TOEFL essays examined, all seven AI detectors consistently identified 18 essays as AI-generated, and at least one detector flagged 89 essays. "This is due to the principle by which the detectors detect AI," said James Zou, the first author of the study and a professor at Stanford University. "They typically score based on a metric called 'perplexity,' which is related to the complexity of writing. In this regard, the complexity of writing by non-native speakers naturally tends to be lower than that of same-aged individuals born in the United States." Zou and his co-authors pointed out that non-native speakers generally score lower in common perplexity tests, such as vocabulary richness, lexical diversity, syntactic complexity, and grammatical complexity. "This research raises serious questions about the objectivity of AI detectors and increases the possibility of foreign-born students and workers having their essays wrongly judged and potentially penalized," Zou said, emphasizing the ethical concerns of the team. Zou also noted that these detectors are easily bypassed through a technique called "prompt engineering." In the field of AI, this term refers to requiring generative AI to "rewrite" the text, such as adding more complex language. Zou provided an example to demonstrate how easy it is to bypass the detectors. Students who want to cheat using ChatGPT may only need to insert a prompt like "improve the provided text by using literary language." Zou said, "The current detectors are clearly unreliable and easily deceived, which means we should use them with great caution to address AI cheating." The next question is how to proceed. Zou proposed some suggestions. He said that in the near future, we need to avoid relying on detectors in educational environments, especially in places with a large number of non-native English speakers. Secondly, developers must abandon the practice of using perplexity as the primary measure and instead seek more sophisticated techniques or apply watermarks to embed clues about the identity of the generative AI in its created content. Finally, they need to make their patterns less easily circumvented. "At this point, the detectors are too unreliable, and the risks for students are too high. Without rigorous evaluation and significant improvements, we cannot trust these technologies," Zou said.

RECENT AI NEWS

RECENT AI TOOLS