"If you start with the soul of a space, you’ll always find the reality. But if you start with reality alone, the soul of it will stay a mystery." - Danielle Krettek
My research interests are at the intersection of human-computer interaction, artificial intelligence, and affective computing. I aim to advance frameworks for designing computational systems that are aligned with human values and inform the development of policies/technologies that help people thrive. I am particularly interested in exploring:
1) Alignment: How can we ensure machine learning models conform to human values? How can we investigate and mitigate discrimination of ML models?
2) Explainability: What mechanisms can be built for algorithmic transparency and interpretability in high-stakes sociotechnical systems?
3) Human-AI Interaction: What interaction paradigms can make AI systems more accessible, intuitive, and beneficial for diverse users?
Looking ahead, I seek to investigate these questions with a particular focus on neurodiverse populations.
In my previous research, I have conducted quantitative data analysis and qualitative inquiry to measure youth's relationship with emerging technologies, as well as its impact on well being, emotional resilience, and social dynamics.
RESEARCH PAPERS
From Novelty to Norm: An 18-Month Study on Gen Z’s Evolving Relationship with ChatGPT
Jee M, Khan A, Khan I. Computer-Supported Cooperative Work & Social Computing (CSCW). Oct 2024. Manuscript in Review.
Abstract: This study investigated the evolving relationship between Generation Z and ChatGPT, focusing on changes in their behaviors, attitudes, and concerns over an 18-month period. We conducted two surveys with n=131 and n=156 participants in January 2023 and July 2024, respectively. Next, we led four focus groups with a total of 24 participants, divided into two equal groups at each time point. Our findings reveal that ChatGPT has largely become a normalized tool in the lives of Gen Z, evidenced by increased adoption and anticipated use. Key drivers of ChatGPT’s normalization include benefits from Productivity, Emotional Connection, and Versatility, which emerged from participants’ reflections on ChatGPT’s usefulness. Additionally, our findings indicate that Gen Z’s expectations for their relationship with ChatGPT are deepening, with users increasingly relying on it for emotional support and expressing a desire for greater personalization. However, concerns surrounding ChatGPT’s societal impact have also grown over time, centering on Job Displacement, Loss of Human Autonomy, and Data Misuse. Overall, our findings provide actionable insights for AI developers, educators, and employers seeking to tailor AI tools, curricula, and workplace strategies to better align with Gen Z’s evolving needs.
Understanding Adolescents’ Perceptions and Aspirations Towards Their Relationship With Personal Technology: Survey Study. Link 🔗
Jee M, Khan A, Nazneen N. Journal of Medical Internet Research (JMIR) Form Res. December 2021.
Abstract: Understanding adolescents' relationship with technology is a pressing topic in this digital era. There seem to be both beneficial and detrimental implications that originate from use of technology by adolescents. Approximately 95% of adolescents have access to a smartphone, and several studies show a positive correlation between screen addiction and trends of anxiety and depression. At the same time, research shows that two-thirds of adolescents believe that technology is a necessity for connecting and making new friends. The aim of this formative study was to understand adolescents' perception of their own and others’ relationship with personal technology. A survey was conducted with 619 adolescents ranging in age from 13 to 19 years. Adolescents were asked how they perceived the relationship with their personal technology, how they perceived others' (parents, siblings, or friends) relationship with personal technology, and how they wish to relate to their personal technology in the future. "Essential,” “Distractive,” and “Addictive” were the most commonly selected descriptors to describe both adolescents' own relationship with technology and others’ relationship as well. Adolescents selected “Provides an escape” more to describe their own relationship with technology. Whereas, they selected “It's just a tool” and “Creates Barrier” more to describe others' relationship with technology. These trends are consistent across ages and genders. In addition, adolescents' aspirations for their relationship with their personal technology varied across ages: 13 to 15-year olds' top choice was “best friend”, 16 to 17-year olds’ top choice was “I don't believe in personal connection with mobile technology,” and 18 to 19-year olds’ top choice was “My personal assistant.” Our 3-lens method allows us to examine how adolescents perceive their relationship with personal technology in comparison to others, as well as their future technological aspirations. Our findings suggest that adolescents see both communalities as well as differences in their own and others' relationships with technology. Their future aspirations for personal technology vary across age and gender. These preliminary findings will be examined further in our follow-up research.
PAPERS IN PROGRESS
Little Minds, Big Laughs: Prompting Strategies and Multi-Step Reasoning Improves Smaller LLMs Performance on Humor Detection.
Jee M, Motulsky A, Greene J, Cubeddu A. (2024). Manuscript Submission Pending.
Abstract: As language models become integral to daily life, their ability to understand humor is essential for creating engaging experiences. Humor poses challenges in NLP due to its context-dependent nature, subjectivity, and reliance on cultural cues. While the latest large language models sometimes grasp these nuances, their computational costs and demands limit may their use in low-resource regions. This paper explores methods to enhance humor detection in smaller-sized LLMs, such as Mistral 7B and Llama 3 8B, through prompting by incorporating context and multi-step reasoning: n-sentence context, Retrieval-Augmented Generation, Chain-of-Thought Prompting and multi-step agentic reasoning for humor detection. Experiments are conducted using two datasets - a Multimodal Humor Dataset consisting of dialogues from the sitcom The Big Bang Theory and the New Yorker Cartoon Caption Dataset. Results show that prompting strategies and multi-step reasoning techniques improve the performance of smaller LLMs on humor detection tasks. However, we also find that the incorporation of context alone does not consistently enhance performance. The effectiveness of context depends on factors such as the method of context selection and the amount of context provided. Future work will explore advanced reasoning frameworks, attention pattern analysis, and the impact of humorous context on performance on humor detection tasks.