Class of 2024San Jose, CA
AboutMy name is Kabir Dua and I am currently a senior at Bellarmine College Preparatory in San Jose, California. In n society where artificial intelligence is becoming increasingly implemented into everyday practical applications, the ethical dimensions of these technologies have become a critical focal point. My research project navigates the intricate maze of ethics and AI, with a particular emphasis on Large Language Models (LLMs). As my research unveils, these digital entities, designed to assist and inform, are far from neutral arbiters. Trained on human-generated data, they inherit not just our collective knowledge but also our latent prejudices. The central question driving this research is multi-pronged: "What biases are present in large language models' representation and depiction of individuals from various protected groups, how do these results compare across different LLMs, and what would it mean to create more equitable and unbiased model outputs?" The architecture of this research is designed to be both comprehensive and accessible, manifesting in a scholarly paper, a condensed video presentation, and an expansive podcast series. The first section of the research lays the intellectual groundwork by examining existing literature on biases in LLMs, dissecting the extent to which these biases are a byproduct of skewed training data or algorithmic predispositions. The second section delves into the semantics of bias, scrutinizing the vocabulary LLMs employ when discussing specific "personas" to identify which models are more or less problematic in their language use. The third section probes the emergent self-awareness of LLMs, assessing their ability to recognize and rectify problematic outputs. Finally, the concluding section synthesizes these findings into actionable recommendations, outlining what an ethically responsible LLM should ideally produce. This research is not merely an academic endeavor; it is a moral imperative. As we become increasingly reliant on AI for various aspects of life, understanding and addressing the biases within these systems becomes crucial. This project aims to serve as a vanguard in the quest for ethical technology, challenging us to confront the digital mirror of our own biases and advocating for a more equitable technological landscape.
- "Marked Personas in Large Language Models" with mentor Kara (Working project)
Marked Personas in Large Language Models
Started July 14, 2023
Abstract or project description
How do large language models depict individuals based on membership in different protected groups (race, gender identities, etc.)? This project will apply a newly proposed framework called "Marked Personas" to examine the associations that large language models form. In addition to applying the framework to different scenarios and examining the results, the project will explore a new direction of investigating whether large language models can understand the harmful content of their associations.