Bias in Machine-Generated Text Detection Systems
A recent study published on arXiv (2512.09292v2) examines biases present in systems that detect machine-generated English text. The researchers assembled a dataset comprising student essays and assessed 16 detection systems for bias related to four factors: gender, race/ethnicity, English-language learner (ELL) status, and economic status. Through regression models and subgroup analyses, they discovered that biases vary across systems, with some models disproportionately identifying disadvantaged groups as machine-generated. This research underscores the possible adverse effects of detection systems, even though they demonstrate high performance.
Key facts
- Study from arXiv:2512.09292v2
- Curated dataset of student essays
- Evaluated 16 detection systems
- Four bias attributes: gender, race/ethnicity, ELL status, economic status
- Used regression-based models and subgroup analysis
- Biases inconsistent across systems
- Several models classify disadvantaged groups as machine-generated
- Highlights potential negative impacts of detection systems
Entities
Institutions
- arXiv