Skip to main content Skip to navigation

Unraveling Downstream Gender Bias from Large Language Models: A Study on AI Educational Writing Assistance

Project Overview

The document examines the role of generative AI, particularly large language models (LLMs), in educational writing assistance and highlights the critical issue of inherent gender bias in these models. A study involving 231 students revealed that the use of LLM-based writing support does not significantly affect the gender bias observed in students' peer reviews, suggesting that while AI tools can enhance writing skills, they may not exacerbate existing biases in student assessments. The findings underscore the necessity of recognizing and addressing AI biases to promote equitable educational outcomes. Overall, the document calls for a careful consideration of how generative AI is integrated into educational settings to ensure it serves as an effective and fair tool for all students.

Key Applications

AI writing support using LLMs (GPT-2, GPT-3, GPT-3.5)

Context: Peer review writing exercise conducted with students in a classroom and online context

Implementation: Students received varying levels of support: no support, feature-based recommendations, or suggestions from fine-tuned LLMs.

Outcomes: No significant difference in gender bias between groups using LLM suggestions and those receiving no assistance; students demonstrated similar outcomes in writing quality.

Challenges: Potential biases in LLMs exist, but they did not transfer to student outputs in this study.

Implementation Barriers

Bias Transfer

Biases inherent in LLMs can propagate through the writing support pipeline.

Proposed Solutions: Conduct thorough analyses of biases in LLMs and ensure that writing suggestions are unbiased.

Project Team

Thiemo Wambsganss

Researcher

Xiaotian Su

Researcher

Vinitra Swamy

Researcher

Seyed Parsa Neshaei

Researcher

Roman Rietsche

Researcher

Tanja Käser

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Thiemo Wambsganss, Xiaotian Su, Vinitra Swamy, Seyed Parsa Neshaei, Roman Rietsche, Tanja Käser

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

Let us know you agree to cookies