Abstract
STS Project
In my STS project, I examined the current applications of artificial intelligence (AI) in the recruitment sector and how they are regulated. My research question is: How is the use of AI in hiring being regulated, and what are the consequences of these regulatory approaches? As AI becomes more widespread in recruitment, there has been growing concern regarding bias, transparency, and fairness in its application. I introduced AI recruitment tools such as video interview analysis and applicant tracking systems. As one of the few states with AI-related laws, I analyzed two Illinois laws: the Artificial Intelligence Video Interview Act of 2020 and the Illinois Human Rights Act Amendment of 2024. Additionally, I analyzed academic literature on the risks of AI recruitment, legal scholarship on Illinois’ laws, and corporate policy documents from HireVue, a company that has been providing AI recruitment tools since the 2010s, including its “AI Ethics Principles” and “Explainability Statement” posted on its official website.
The research reveals that regulations governing the use of AI tools in recruitment in Illinois have evolved from being procedural to substantive. While the 2020 Act primarily focused on procedural factors such as notification, informed consent, and data deletion, it aims to ensure transparency and applicants' right to know. On the other hand, the 2024 Amendment addresses issues of discrimination and accountability. Despite the advances, current regulations have limitations as well. Regulation is always "reactive." States typically add or improve regulatory measures only after public concerns arise or companies providing the relevant technology face litigation. Additionally, procedural protections such as "informed consent" may be ineffective, as applicants do not fully understand how AI evaluates them. Overall, I believe that although regulations are becoming increasingly robust, they still face challenges. In particular, it is hard to hold AI tools accountable, and algorithmic bias is difficult to overcome.
Technical Project
The purpose of my capstone project is to design and prototype a centralized policy querying system for Contracted Independent Organizations (CIOs) at the University of Virginia. This project aims to resolve an important usability issue, where policies for CIOs are scattered across many University of Virginia website locations and documents, making policy verification an inefficient and time-consuming task for students and CIO leaders. Instead of developing an institutional-grade solution, this project designs an architecture that implements a working prototype and tests its usability and performance under real-world scenarios.
The fundamental idea behind this project is to develop a Retrieval-Augmented Generation (RAG) model that leverages a large language model (LLM). This system will be designed to retrieve and store relevant CIO policy documents from various UVA websites. Documents such as University of Virginia Student Affairs policies on student event planning and University of Virginia SAF funding policies will be collected, cleaned, segmented, and stored for embedding into a vector database.
When the user enters a natural-language query, such as "Can we serve food in this space?", the system can retrieve the most relevant policy sections using vector similarity search. These sections will be fed to the LLM, which will restrict it to give answers strictly based on the information present in these sections. Each answer will be accompanied by citations that include the title, department, and the link to the original document, ensuring students get accurate information.
The proposed architecture of the backend will be a document ingestion pipeline that will parse the PDF/HTML documents, chunk the texts, and add metadata tags, an embedding model that will be used to do semantic indexing, a vector store that will be used to do vector similarity search, an LLM layer that will be used to generate answers to the queries, and a citation formatting component that will be used to format the citations to the original documents. The frontend will have a lightweight web interface that lets users enter queries in natural language and view answers in a compact format. Furthermore, the user interface will be clear and accessible so that users can independently verify answers following the citations provided.
