AI Gender Bias, Disparities, and Fairness: Does Training Data Matter?
Project Overview
The document examines the role of generative AI in education, emphasizing its potential to address issues of gender bias, disparities, and fairness, particularly in the context of automatic scoring systems. It reveals that imbalanced training datasets can reinforce existing gender inequalities, while the use of mixed-gender training datasets can significantly reduce these biases. The findings indicate that AI models developed using balanced datasets yield more equitable and less biased outcomes, thereby promoting fairness in AI applications within educational environments. This research underscores the importance of equitable data practices in the development of AI tools, suggesting that thoughtful data curation can enhance the effectiveness and fairness of AI in educational settings, ultimately supporting better learning experiences for all students.
Key Applications
Fine-tuned BERT and GPT-3.5 models for automatic scoring
Context: Educational context focusing on automatic scoring of student-written responses for high school assessments
Implementation: Used mixed-gender datasets for training AI models and evaluated their performance against gender-specific models
Outcomes: Mixed-gender trained models showed lower mean score gaps and better fairness compared to gender-specific models, indicating reduced gender bias.
Challenges: The challenge lies in ensuring balanced datasets and addressing existing biases in training data.
Implementation Barriers
Bias-related barrier
Existing gender bias in training datasets can lead to biased AI predictions and scoring. Misconceptions about AI bias can lead to unjustified fears and resistance to AI applications in education.
Proposed Solutions: Employ mixed-gender datasets for training AI models to mitigate bias and enhance fairness. Increase awareness and understanding of AI bias, emphasizing the need for balanced and fair datasets.
Project Team
Ehsan Latif
Researcher
Xiaoming Zhai
Researcher
Lei Liu
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Ehsan Latif, Xiaoming Zhai, Lei Liu
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai