Research

Understanding and Designing for Empathetic Human-AI Interactions Powered by Large Language Models

EmotionAIze: Empathy-Driven Interactive Human-AI System for Autistic Individuals

In this project, we explore the potential of generative AI models in supporting mental well-being of neurodiverse individuals, with a primary focus on individuals on the autism spectrum. Many autistic adults confront co-occurring mental health challenges, such as depression and anxiety. These challenges can be magnified by life transitions, from forming friendships to entering college or navigating a new job. Such transitions can lead to an increase in negative self-talk (NST), potentially worsening their mental health conditions. Traditional talk therapies might not always be the best fit for those with verbal communication differences, highlighting a significant gap in support. How can the latest advancements in generative AI models be harnessed to address this gap? To answer this question, we aim to build an interactive Human-AI system that can generate contextual, empathetic counter-responses to NST, expressed through a variety of modalities, and provide an interactive system that respects the unique needs of individuals on the autism spectrum. EmotionAIze cannot only reframe unhelpful self-talk, but also serve as an additional source of support that is available whenever and wherever it is needed outside of therapy sessions.

Understanding Motivations & Barriers in Online Counterspeech Engagement and the Potential Role of AI

In a digitally interconnected world, social media platforms are both conduits for public dialogue and breeding grounds for hate speech. While prior research has often focused on the effectiveness of online counterspeech, little is known about motivations and barriers to engage in it. Based on a survey of 458 U.S. participants, we develop and validate a multi-item scale for understanding counterspeech motivation and barriers, revealing that past experiences and demographic differences strongly influence counterspeech engagement on social media. Our work also examines how motivation and barriers affect user satisfaction and perceived effectiveness of their counterspeech. Additionally, our work is the first to examine people’s willingness to use AI technologies like ChatGPT for writing online counterspeech. Our findings contribute to a nuanced understanding of factors shaping online counterspeech engagement and provide key insights around how people perceive the potential role of AI assistance in countering online hate.

Countering Hate in the Age of Artificial Intelligence: Evaluating the Effectiveness of Human vs. Machine-Generated Counterspeech

The increasing prevalence of counter-speech as a tool to challenge online hate speech has gained popularity, primarily due to its respect for freedom of speech and its capacity to empower individuals to openly oppose hate speech. However, the nuances that define an effective counter-speech and the potential role of AI in crafting such responses remain largely unexplored. In this project, we delve into the following questions: What are the defining characteristics of an impactful counter-speech? How do individuals who disseminate offensive content react attitudinally and behaviorally when confronted with counter-speech? And how does human-crafted counter-speech compare to AI-generated responses in terms of effectiveness? To address these questions, we measure the shifts in attitudes and behaviors of individuals posting hateful content when exposed to both human and AI-generated counter-speech. Through this, we seek to uncover not only the nuances of effective counter-speech but also the potential role, involvement, and degree of transparency required when AI is employed in its generation.

Collaborative Co-Writing With AI: Investigating User Strategies and Preferences for Machine-Generated Counterspeech

This user study investigates user interactions and preferences with a system powered by Large Language Models (LLMs) designed to generate counterspeech. Our goal is to understand editing patterns and strategies (tone, rhetorical style, etc.) users prefer when responding to harmful content on social media.

Affecting Real-World Changes Through Improved Discourse Strategies and Conversational Norms

This project uses a multi-step reasoning framework and prompt-based large language models (LLMs) to explore the correlation between social media language patterns and national health outcomes during pandemics. This framework, grounded in fuzzy-trace theory, allows for the extraction and analysis of causal coherence in discussions opposing health measures, providing empirical links between online linguistic patterns and real-world public health trends.

Interpretable NLP in Traffic Stops: Understanding Officer Speech in Escalated Interactions

This collaboration with Stanford University uses NLP and deep learning interpretation techniques to analyze officer speech in traffic stops with Black drivers from police body-worn camera footage. We aim to identify early interaction patterns associated with escalated outcomes and compare model results with human perspectives to better understand the public’s perception and concerns in police interactions. [paper link]

Semantic Polarization in Broadcast Media: Analyzing Impact on Online Discourse

Recent studies reveal limited evidence that filter bubbles and echo chambers fully explain partisan segregation in news audiences. This suggests traditional broadcast media may play an equally important role in polarizing public discourse. This research expands analysis to include both online and traditional media by examining the relationship between broadcast news language on CNN and Fox and corresponding Twitter discourse. Analyzing a decade of closed captions from CNN and Fox and topically related tweets, we provide a framework for measuring semantic polarization between the networks over time. Results show a sharp increase in polarization in how the channels discuss key topics after 2016, with the highest peaks in 2020, where there is barely any linguistic overlap in contextual discussions of the same keywords. We further demonstrate broadcast media language significantly impacts semantic polarity on Twitter. Overall, the work implies media polarization on TV fuels rather than supports online democratic discourse. [paper link]

Understanding Linguistic Signature of Online Racial Microaggressions

In this work, we examine the linguistic signature of online racial microaggressions (acts) and how it differs from that of personal narratives recalling experiences of such aggressions (recalls) by Black social media users. We manually curate and annotate a corpus of acts and recalls from in-the-wild social media discussions, and verify labels with Black workshop participants. We leverage NLP and qualitative analysis on this data to classify (RQ1), interpret (RQ2), and characterize (RQ3) the language underlying acts and recalls of racial microaggressions in the context of racism in the U.S. Our findings show that neural language models (LMs) can classify acts and recalls with high accuracy (RQ1) with contextual words revealing themes that associate Blacks with objects that reify negative stereotypes (RQ2). Furthermore, overlapping linguistic signatures between acts and recalls serve functionally different purposes (RQ3), providing broader implications to the current challenges in content moderation systems on social media.