Abstract
Sociotechnical Synthesis
Introduction
My Capstone project and my STS research explore how AI can be made safer, more trustworthy, and better aligned with human needs in real-world settings. My technical Capstone focused on enhancing small parameter language models' commonsense reasoning through CASEEDIT, a localized knowledge editing framework. My STS project, on the other hand, examined how cognitive science-inspired interpretability methods could help AI models behave more ethically and transparently, especially in high-stakes sectors like healthcare, justice, and education. Together, these projects grapple with the technical and societal challenges of ensuring that AI systems function effectively and behave in ways that align with human values.
Capstone Project
My Capstone addressed the problem of how small, edge-deployed language models struggle with personalized commonsense reasoning. In environments like smart homes, users often need AI to adapt to unconventional but intuitive uses of everyday objects—something large models handle better, but at a cost of impractical computational resources. To address this, we developed CASEEDIT, a knowledge editing framework that enabled small models to integrate personalized commonsense adaptations while minimizing ripple effects on unrelated knowledge. By enabling AI systems to better reflect human behavior patterns, CASEEDIT aimed to make AI assistants more practical, ethical, and useful. Considering the societal impacts of this technology is essential because commonsense reasoning is deeply cultural, context-dependent, and often invisible to model designers. Poorly adapted AI can perpetuate misunderstandings or even create risks in household environments if models act on rigid or inappropriate assumptions.
STS Research Project
In my STS research, I explored how cognitive science and neuroscience-inspired methods, such as mechanistic interpretability and chain-of-thought reasoning, could be used to decode AI systems' internal processes. I critically analyzed how interpretability techniques reveal how AI models function technically and how they reflect broader social norms, biases, and ethical standards. Using STS frameworks, I argued that AI interpretability must be culturally sensitive and sector-specific, especially in healthcare, law, and education, where the risks of AI errors are highest. My research showed that simply improving technical transparency is insufficient; AI systems must also be regulated and designed with an understanding of how different societies perceive fairness, trust, and accountability.
Conclusion
When considered together, my Capstone and STS research reveal that improving AI’s technical adaptability and transparency must go hand-in-hand with broader societal and ethical considerations. CASEEDIT demonstrated how models could be engineered to reason more flexibly within human-specific contexts. However, without STS-driven insights, such technical improvements could still fail by embedding hidden biases or ignoring ethical risks. Bridging the gap between technical and societal understanding ensures that future AI systems are not only functionally capable but also socially responsible, trustworthy, and truly aligned with the diverse ways humans live and reason.
