Redirecting

What this source establishes

The study claims that GPT detectors are not sufficiently reliable or fair, particularly because they can misclassify human writing as AI-generated and can be easily bypassed. This is because over half of non-native English writing samples were incorrectly classified as AI-generated, while better prompt design could help AI-generated text evade detection.

Limitations

Specific 2023 detectors and text samples; newer versions require fresh evaluation.

Supporting excerpt

Introduction Generative language models based on GPT, such as ChatGPT, have gained significant attention in recent times. Within a mere 2 months of its launch, ChatGPT amassed over 100 million monthly active users, marking its place as one of the fastest-growing consumer internet applications in history.1 Despite their immense potential for enhancing productivity and fostering creativity, these powerful models also pose risks, such as the proliferation of AI-generated content masquerading as human written, which may lead to the spread of fake content and exam cheating. Educators, in particular, are increasingly concerned about determining when and where students have used AI and AI writing tools in their work. However, multiple studies have demonstrated the difficulty humans face in detecting AI-generated content with the naked eye,2 thus creating an urgent and pressing demand for effective detection methods. While several GPT detectors have been developed and implemented to mitigate the risks associated with AI-generated content, their accuracy, reliability, and effectiveness remain uncertain due to limited evaluation.3 This knowledge gap is especially worrisome given the potentially harmful consequences ofmistakenly flagging an innocent student’s work as AI generated.4 Given the transformative impact of generative language models and the potential risks associated with their misuse, developing trustworthy and accurate detection methods is crucial. In our recent preprint,5,6 we exposed an alarming bias in GPT detectors against non-native English speakers: over half of the non-native English writing samples were misclassified as AI generated, while the accuracy for native samples remained near perfect. Our analysis further revealed a trend where more literary language was classified as more ‘‘human’’: enhancement of word choice in non-native English writing samples reduced misclassification, while simplifying native writing samples increased it, suggesting that GPT detectors are inadvertently penalizing individuals with limited linguistic proficiency. On the other hand, we found that GPT detectors be easily bypassed by better ChatGPT prompt design. This raises a pivotal question: if AI-generated content can easily evade detection while human text is frequently misclassified, how effective are these detectors truly? 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 themarginalization of non-native speakers in evaluative or educational settings. This paper is among the first to systematically examine the biases present in GPT detectors and advocates for further research into addressing these biases and refining the current detection methods to ensure a more equitable and secure digital landscape for all users.