Search

In subjective decision-making, where decisions are based on contextual interpretation, Large Language Models (LLMs) can be integrated to present users with additional rationales to consider. The diversity of these rationales is mediated by the ability to consider the perspectives of different social actors. However, it remains unclear whether and how models differ in the distribution of perspectives they provide. We compare the perspectives taken by humans and different LLMs when assessing subtle sexism scenarios. We show that these perspectives can be classified within a finite set (perpetrator, victim, decision-maker), consistently present in argumentations produced by humans and LLMs, but in different distributions and combinations, demonstrating differences and similarities with human responses, and between models. We argue for the need to systematically evaluate LLMs' perspective-taking to identify the most suitable models for a given decision-making task. We discuss the implications for model evaluation., Comment: Accepted at CHI Conference on Human Factors in Computing Systems (CHI '25), April 26-May 1, 2025, Yokohama, Japan

While a lucrative and growing field, low levels of gender and racial diversity in CS remain prevalent. Education and workforce support programs with the intention to promote underrepresented students' persistence in CS exist, which teach skills, inform of career options, and grow students' network in CS. Studies demonstrate these programs' effectiveness as it relates to changes in affective outcomes, such as participants' confidence in CS skills and attitudes towards CS jobs. However, programs' longitudinal impact on participants' build-up of social capital in CS, and the resulting social capital's influence on their persistence in CS, remain unexplored. Motivated by the literature that associates demographic identifiers with access to social capital, and students' access to developmental relationships and career resources (social capital) in CS with their persistence, this study explores a CS support program's impact on persistence through capital building. We focus on Google's CSSI, which provided graduating high school students with a 3-week-long introduction to CS. We use interviews with participants who are now 2-5 years out of the program to study CSSI's impact on their social capital and long-term CS persistence. Thematic analysis reveals three program elements that influenced students' build-up of social capital, and that the resulting persistence was realized through students' progress towards internships and goals for paying-it-forward in CS. These findings inform our recommendations that future support programs and educational settings consider mentorship centered on socioemotional support, opportunities for collaboration, and time for fun social activities. Additional suggestions center on engaging socially-oriented individuals with CS support programs. These insights inform CS educators on design choices that can encourage the persistence of underrepresented students in CS., Comment: Presented at the 2024 Annual Conference of the American Society for Engineering Education

An essential aspect of evaluating Large Language Models (LLMs) is identifying potential biases. This is especially relevant considering the substantial evidence that LLMs can replicate human social biases in their text outputs and further influence stakeholders, potentially amplifying harm to already marginalized individuals and communities. Therefore, recent efforts in bias detection invested in automated benchmarks and objective metrics such as accuracy (i.e., an LLMs output is compared against a predefined ground truth). Nonetheless, social biases can be nuanced, oftentimes subjective and context-dependent, where a situation is open to interpretation and there is no ground truth. While these situations can be difficult for automated evaluation systems to identify, human evaluators could potentially pick up on these nuances. In this paper, we discuss the role of human evaluation and subjective interpretation to augment automated processes when identifying biases in LLMs as part of a human-centred approach to evaluate these models., Comment: HEAL Workshop at CHI Conference on Human Factors in Computing Systems, May 12, 2024, Honolulu, HI, USA

As large language models (LLMs) advance to produce human-like arguments in some contexts, the number of settings applicable for human-AI collaboration broadens. Specifically, we focus on subjective decision-making, where a decision is contextual, open to interpretation, and based on one's beliefs and values. In such cases, having multiple arguments and perspectives might be particularly useful for the decision-maker. Using subtle sexism online as an understudied application of subjective decision-making, we suggest that LLM output could effectively provide diverse argumentation to enrich subjective human decision-making. To evaluate the applicability of this case, we conducted an interview study (N=20) where participants evaluated the perceived authorship, relevance, convincingness, and trustworthiness of human and AI-generated explanation-text, generated in response to instances of subtle sexism from the internet. In this workshop paper, we focus on one troubling trend in our results related to opinions and experiences displayed in LLM argumentation. We found that participants rated explanations that contained these characteristics as more convincing and trustworthy, particularly so when those opinions and experiences aligned with their own opinions and experiences. We describe our findings, discuss the troubling role that confirmation bias plays, and bring attention to the ethical challenges surrounding the AI generation of human-like experiences., Comment: Presented at the Trust and Reliance in Evolving Human-AI Workflows (TREW) Workshop at CHI 2024

The data created from virtual communication platforms presents the opportunity to explore automated measures for monitoring team performance. In this work, we explore one important characteristic of successful teams - Psychological Safety - or the belief that a team is safe for interpersonal risk-taking. To move towards an automated measure of this phenomenon, we derive virtual communication characteristics and message keywords related to elements of Psychological Safety from the literature. Using a mixed methods approach, we investigate whether these characteristics are present in the Slack messages from two design teams - one high in Psychological Safety, and one low. We find that some usage characteristics, such as replies, reactions, and user mentions, might be promising metrics to indicate higher levels of Psychological Safety, while simple keyword searches may not be nuanced enough. We present the first step towards the automated detection of this important, yet complex, team characteristic., Comment: Published as an extended abstract in the ACM Conference on Human Factors in Computing Systems (CHI) 2024

While distributed workers rely on scheduled meetings for coordination and collaboration, these meetings can also challenge their ability to focus. Protecting worker focus has been addressed from a technical perspective, but companies are now attempting organizational interventions, such as meeting-free weeks. Recognizing distributed collaboration as a sociotechnical challenge, we first present an interview study with distributed workers participating in meeting-free weeks at an enterprise software company. We identify three orientations workers exhibit during these weeks: Focus, Collaborative, and Time-Bound, each with varying levels and use of unstructured time. These different orientations result in challenges in attention negotiation, which may be suited for technical interventions. This motivated a follow-up study investigating attention negotiation and the compensating mechanisms workers developed during meeting-free weeks. Our framework identified tensions between the attention-getting and attention-delegation strategies. We extend past work to show how workers adapt their virtual collaboration mechanisms in response to organizational interventions, Comment: Published at the ACM Conference on Human Factors in Computing Systems (CHI) 2024

The growing field of machine learning (ML) and artificial intelligence (AI) presents a unique and unexplored case within persistence research, meaning it is unclear how past findings from engineering will apply to this developing field. We conduct an exploratory study to gain an initial understanding of persistence in this field and identify fruitful directions for future work. One factor that has been shown to predict persistence in engineering is belonging; we study belonging through the lens of confidence, and discuss how attention to social belonging confidence may help to increase diversity in the profession. In this research paper, we conduct a small set of interviews with students in ML/AI courses. Thematic analysis of these interviews revealed initial differences in how students see a career in ML/AI, which diverge based on interest and programming confidence. We identified how exposure and initiation, the interpretation of ML and AI field boundaries, and beliefs of the skills required to succeed might influence students' intentions to persist. We discuss differences in how students describe being motivated by social belonging and the importance of close mentorship. We motivate further persistence research in ML/AI with particular focus on social belonging and close mentorship, the role of intersectional identity, and introductory ML/AI courses., Comment: Published in the 2023 Annual Conference of the American Society for Engineering Education

Machine Learning (ML) and Artificial Intelligence (AI) are powering the applications we use, the decisions we make, and the decisions made about us. We have seen numerous examples of non-equitable outcomes, from facial recognition algorithms to recidivism algorithms, when they are designed without diversity in mind. Thus, we must take action to promote diversity among those in this field. A critical step in this work is understanding why some students who choose to study ML/AI later leave the field. While the persistence of diverse populations has been studied in engineering, there is a lack of research investigating factors that influence persistence in ML/AI. In this work, we present the advancement of a model of intentional persistence in ML/AI by surveying students in ML/AI courses. We examine persistence across demographic groups, such as gender, international student status, student loan status, and visible minority status. We investigate independent variables that distinguish ML/AI from other STEM fields, such as the varying emphasis on non-technical skills, the ambiguous ethical implications of the work, and the highly competitive and lucrative nature of the field. Our findings suggest that short-term intentional persistence is associated with academic enrollment factors such as major and level of study. Long-term intentional persistence is correlated with measures of professional role confidence. Unique to our study, we show that wanting your work to have a positive social benefit is a negative predictor of long-term intentional persistence, and women generally care more about this. We provide recommendations to educators to meaningfully discuss ML/AI ethics in classes and encourage the development of interpersonal skills to help increase diversity in the field., Comment: Presented at the 2022 Annual Conference of the American Society for Engineering Education