Cultural Bias and AI Detection
At QuillBot, we are invested in the ethical use of artificial intelligence in education. So, a recent study on the unreliability of seven widely-used AI detectors (Liang et al., 2023) caught our eye. We dig into the details below.
Study summary
The researchers tested seven GPT detectors to see how they classify essays written by native and non-native English speakers.
- Researchers collected essays written by non-native English speakers (for the TOEFL test, a standardized test of English proficiency) and essays written by 8th-grade native English speakers.
- The essays were run through the detectors. While the detectors generally correctly classified the essays from native English speakers as human-generated, they mistakenly classified over half of the essays written by non-native speakers as AI generated.
- Researchers then used AI to modify the essays.
- When ChatGPT was used to enhance the vocabulary of the TOEFL essays, the detectors were less likely to flag the texts as AI generated.
- When ChatGPT was used to simplify the vocabulary of the 8th-grade U.S. essays, the detectors were more likely to flag the texts as AI-generated.
AI detectors rely on measures of text predictability. Because English language learners’ writing is typically less complex, and thus more predictable, than native speakers’ writing, it is more likely to be flagged by detectors.
Why is this important?
This study confirms what many students and educators have experienced: traditional AI detectors are biased against the writing of English language learners.
The poor performance of AI detectors can have serious consequences for students and scholars, as many schools and organizations use them to enforce rules against use of AI-generated content.
Tens of millions of English language learners write in English for all sorts of purposes, often employing paraphrasing and grammar checking tools ethically. Currently, they are at risk of facing unfair accusations of dishonesty and all the repercussions that follow.
QuillBot’s response: a better AI detector
AI tools are here to stay, as is the demand for reliable AI detection. So, in order to support students and educators, we’ve built a better AI detector, one that provides high-quality, accurate AI detection without penalizing ethical use of AI tools.
QuillBot’s AI detector has many advantages over competitors:
- It was designed with writers in mind and works for students and educators alike.
- It provides in-depth analysis of text that clearly distinguishes between AI-generated and AI-refined text, which helps avoid false positives and biases against non-native English students who use writing tools ethically.
- It’s free, so the barrier to access has been removed.
We are excited to set a new standard in AI detection that works for all writers.
Reference
Liang, Weixin, et al. “GPT detectors are biased against non-native English writers.” Patterns, vol. 4, no. 7, July 2023, p. 100779. https://doi.org/10.1016/j.patter.2023.100779.
