Accuracy of Azure AI Language Service PII Detection: an analysis for future implementation into the Digital Trails App; Hey Siri, You’re Sexist—An Actor-Network Theory Analysis on the Gendering of AI 2 views

Artificial intelligence has rapidly evolved into a foundational technology shaping how individuals interact with information, systems, and one another. As AI becomes embedded in everyday tools, it introduces both technical challenges and ethical concerns. My research examines two critical issues: the detection of personally identifiable information (PII) and the reinforcement of gender bias in AI voice assistants. Each component of my work addresses one of these concerns through a technical evaluation and sociotechnical analysis. In the technical portion of my research, I evaluated the effectiveness of Microsoft Azure Foundry’s AI Language PII detection system in identifying and redacting sensitive personal information. As AI systems process growing volumes of user data, failure in PII detection pose significant risks to privacy and regulatory compliance. Using the SPY dataset, which contains diverse examples of PII such as names, phone numbers, and addresses, I measured the model’s recall across multiple categories. My findings highlight both the strengths and limitations of automated redaction systems, demonstrating that Azure’s model performs well in identifying name, email, and address entities but struggles with phone number and ID number entities. These results suggest that Azure’s PII detection requires further refinement or data preprocessing before it can be fully relied upon. In my STS research, I examined the widespread feminization of AI voice assistants and the social forces that sustain this design pattern. Using Actor-Network Theory (ANT), I analyzed how human actors (developers, users) and non-human actors (datasets and cultural stereotypes) interact to produce and stabilize gendered AI systems. I found that the dominance of female-coded assistants to be the result of a self-reinforcing network consisting of user preferences, market incentives, developer biases, and biased training data, all of which continuously interact to reproduce feminized AI personas. This dynamic not only reflects existing gender stereotypes but actively normalizes them through everyday human-AI interaction. Considering both components together provides a comprehensive understanding of AI as a socio-technical system. While technical solutions like PII detection aim to protect users, they operate within broader social contexts that shape the design and deployment of AI. By pairing a technical evaluation of data privacy with a sociotechnical analysis of gender representation, this project illustrates the crucial role that fairness and inclusivity play in AI development. As AI becomes increasingly pervasive and shapes how we communicate and interact, its future will depend on both technological precision and a collective commitment to ethical and equitable design.

BS (Bachelor of Science)

ANT; AI; Gender

School of Engineering and Applied Science Bachelor of Science in Computer Science Technical Advisor: Mark Rucker STS Advisor: Sean Murray

All rights reserved by the author (no additional license for public reuse)

2026-05-08