GPT detectors are biased against non-native English writers

By Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, James Y. Zou. Patterns, 2023

585 citations recorded by the source corpus.

Summary GPT detectors frequently misclassify non-native English writing as AI generated, raising concerns about fairness and robustness. Addressing the biases in these detectors is crucial to prevent the marginalization of non-native English speakers in evaluative and educational settings and to create a more equitable digital landscape.

Subfields: Health Informatics, Safety Research

View Paper: https://www.semanticscholar.org/paper/e2352eb8e69d7cc0d98f9b13eae33f9c02310ad4

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Health Informatics

1. Cecilia Ka Yuk Chan
2. Katherine K. W. Lee
3. Dragan Gasevic

In Safety Research

1. Katherine K. W. Lee
2. Lixiang Yan
3. Maarten Sap
4. Cecilia Ka Yuk Chan
5. Louis-philippe Morency
6. Dragan Gasevic

Citations in the corpus

• Can AI-Generated Text be Reliably Detected? (2023, background): “In addition to reliability issues of detectors that we have studied in this paper, a recent work by Liang et al. [30] has shown that these detectors can be biased against non-native English writers.”
• Practical and ethical challenges of large language models in education: A systematic scoping review (2023, background): “These details are especially pertinent when considering the diverse cultural backgrounds of students and the fact that most LLMs are primarily trained on English datasets, which could potentially introduce biases towards non-native English students (Liang et al., 2023).”
• Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text (2024, methodology): “(3) Note that · denotes the dot product between two vector-valued quantities.”
• The AI generation gap: Are Gen Z students more interested in adopting generative AI such as ChatGPT in teaching and learning than their Gen X and millennial generation teachers? (2023, result): “…unreliable, in line with other recent studies warning of risks of false positives and false negatives when using these tools (Dalalah & Dalalah, 2023) and the tendency of current AI detectors to misclassify and exhibit bias against the writing of non-native English speakers (Liang et al, 2023).”
• SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents (2023, background): “However, we remind readers that LLMs are known to have biases and problems for evaluation, including positional bias (Wang et al., 2023), factual inconsistency (Luo et al., 2023), favoring native speakers (Liang et al., 2023).”
• DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature (2023, unknown): “However, these mask-filling models are used off-the-shelf, without any fine-tuning or adaptation to the target domain.”
• Provable Robust Watermarking for AI-Generated Text (2023, unknown): “These works primarily leverage hand-crafted or learned “statistical patterns” of generated text, thus their performance is not robust to distribution changes (e.g., by prompting / conditioning), prone to biases [Liang et al., 2023], and vulnerable to adversarial attacks.”
• Out of Style: RAG's Fragility to Linguistic Variation (2025, unknown): “Failing to account for these linguistic variations risks excluding a broad segment of users from effective interaction, especially for users whose linguistic expressions fall outside the narrow patterns these systems are tuned on (Liang et al., 2023).”