Potential Abstract:
Abstract: Large language models (LLMs) have gained significant attention in recent years for their ability to generate text that resembles human language. However, the potential biases and limitations of LLMs have also raised concerns, particularly in the context of education where nuanced understanding and inclusive language are essential. This study aims to investigate the intersectional spheres of LLMs and their rhetorical implications, with a focus on capturing nuances in educational text.
Drawing on critical theories and rhetorical analysis, we examine the biases and implications of LLM-generated text in relation to intersectionality, a framework that recognizes the interconnected nature of social identities and their impact on lived experiences. Through a mixed-methods approach, we analyze a diverse corpus of educational texts and LLM-generated outputs, considering the ways in which the LLM models interpret and represent different dimensions of intersectionality.
Our findings reveal that LLMs tend to reproduce dominant discourses and perpetuate biases, thereby limiting the nuanced understanding of intersectionality. Despite their ability to generate text at a large scale, LLMs struggle to accurately capture the complexities of intersectionality due to the inherent biases in the training data and the limitations of the model architecture. We argue that a deep understanding of rhetoric and intersectionality is vital for evaluating and critiquing LLM outputs, as it enables researchers and educators to recognize and address the shortcomings of these models.
Our research contributes to the emerging field of AI in education by illuminating the intersectional implications of LLMs in educational text generation. By highlighting the limitations and biases of LLMs, we advocate for the critical engagement with these models in educational contexts. Our study offers insights into how educators can leverage LLMs responsibly and ethically, while also emphasizing the importance of human judgment and subject matter expertise in order to interpret and refine LLM-generated outputs.
Potential References:
- Bias out-of-the-box: An empirical analysis of intersectional occupational biases in popular generative language models
- Detecting emergent intersectional biases: Contextualized word embeddings contain a distribution of human-like biases
- Unequal representations: Analyzing intersectional biases in word embeddings using representational similarity analysis
- Embodying public feminisms: Collaborative intersectional models for engagement
- Scaling language models: Methods, analysis & insights from training gopher
