My research focuses on understanding how self-presentation acts on media. In 3 previous projects, I have examined how people present themselves by comparing their offline personality with self-presentations on Snapchat, Facebook, Instagram, Texting, and Video calls. I use mixed methods, combining Big-5 personality surveys and interviews to probe self-presentational differences between media. I have found reliable differences between media and offline self-presentations in those studies. In particular, that Neuroticism is always lower on media across multiple studies. Furthermore, some differences aren’t due only to the affordances of the media, but social norms on that media, which I found when looking at different types of Instagram accounts.

I expand on this work through two additional studies using new methods to explore different research questions. The first study examines how users of multiple media make decisions about where to make a post. I used scenarios to understand how users of multiple media decide between the media that they use and propose a mental model of media choice. First, if there is a quick match between an existing social norm on a medium and their intended post, they will post it there. If there isn’t, then the user must consider the affordances and other social norms of the media they use before deciding where to post.

The second study examines how outside observers interpret an Instagram profile. I presented observers with a dynamic webpage, similar to Instagram, to gather behavioral and survey data on how said observers looked at and interpreted these Instagram profiles. Based on the results, observers don’t see an Instagram specific self-presentation but are more accurate to the profile owner’s offline personality. I finish by presenting my doctoral research's technical and theory-based implications.

Prior literature on online communities proposes an important function for social roles. Theoretical models argue that specific roles are critical for community success, and claim that individual roles shift in predictable ways over the lifetime of the community. However these models are currently hard to assess, as they do not provide systematic definitions of roles, nor empirical work evaluating the impact of roles on community success. Further they do not specify interrelations between different roles. This thesis addresses these questions in the context of enterprise online communities. Using gold-standard systematically-defined social roles, we explore role behaviors and their impacts using both quantitative and qualitative measures of community success. We find evidence contradicting prior theoretical claims about role shifts and the division of labor between roles. We also examine language styles within communities, finding a complex relationship in the types of support language that engender success. To develop new models of relations between roles, we utilize graphical methods to discover important community subgroups centering around leadership. These subgroups are important predictors of community success, significantly extending the explanatory power of existing network theories. Furthermore, subgroups serve as sources of significant topic-setting language. This work elucidates the workings of online communities at a user level, increasing theoretical understanding and suggesting how new community tools might be developed that promote community success.

Artificially intelligent systems play an increasingly large role in our everyday lives. These intelligent systems make decisions big and small—from who gets a mortgage to what articles we read online. However, these systems are often designed out of convenience; they are built using data that is available, they do not explain the decisions they make, and sometimes they are simply inaccurate. At best, these problems result in systems that are difficult to use; at worst, these problems result in reinforcing societal biases at scale. Rather than designing systems from a standpoint of convenience, my work centers the human experience in the design of intelligent systems. I show how to work with users to build maximally accurate systems. I demonstrate that common intelligent systems are biased and not equally usable by everyone but also that this can be fixed through careful data collection. I show what users need when a system explains itself and how we can build our systems to explain the right thing at the right time. Finally, I move this work out of the lab and into the wild through a field study with expert users of an explainable AI system. Together, I use these studies to make recommendations that ensure we meet the needs of our human users in the intelligent systems that we build.

Personal informatics (PI) technologies allow unprecedented opportunities to track and analyze complex data about ourselves. However, a concern is that these technologies can make normative assumptions about user goals and ideal outcomes. Such assumptions could be especially problematic for Affective PI, as there is a risk that technologies which reflect implicit goals for users be more positive or reduce stress could ironically decrease well-being (Mauss et al., 2011). Furthermore, users could actively avoid PI data if they feel unable to meet the demands of the system (Duval & Wicklund, 1972), running counter to the view that users will engage data for beneficial insights (Kersten-van Dijk et al., 2017). We tested whether Affective PI systems that reflect goals for particular emotion outcomes (Improvement) have counterproductive effects for well-being and user engagement. These outcomes were contrasted against systems that instead reflect goals for Self-Knowledge, a top user interest (Hollis et al., 2018). Study 1 examined the effects of implicit goals in the context of an automatic stress detection and feedback system used during an exam. Participants viewed instructions that either describes the system goal as stress reduction (Improvement), stress reduction with a relaxation strategy (Self-Efficacy), accurate self-knowledge (Self-Knowledge), or saw no system goal (Control). Study 2 was a 21-day field study during which participants used a manual emotion-tracking web app that either emphasized a goal of increased positivity (Improvement), a goal of accurate self-knowledge (Self-Knowledge), or only completed pre-post surveys (Control). For each study, participants completed measures of well-being and engagement with the experimental systems. Across both studies, there were no significant condition differences in well-being. However, participants in the Self-Knowledge conditions of both studies considered themselves significantly more successful at achieving the system goals. As a result, Self-Knowledge participants were also more engaged with the stress- and emotion-tracking systems. Unlike prior work showing the ironic effects of emotional positivity goals, we show such negative impacts do not occur in this real-world context. We discuss these results with design implications for self-tracking systems and deepen the theoretical understandings of how users engage with PI.

Reflection is a systematic process of reviewing memories. Reflection on past positive and negative memories increases well-being, as does reflection that is mediated by technology to provide rich digital records of past personal experiences. Technology mediated reflection (TMR) is rapidly growing in popularity, with many deployed systems, however we know little about how one’s mood when using TMR might influence well-being. I use theories of memory and emotion-regulation to motivate hypotheses about the relationship between reflection, mood, and well-being when using technology. I developed a web-based application called MoodAdaptor to test these hypotheses. MoodAdaptor prompted participants to reflect on positive or negative memories depending on current mood. I evaluated how mood and memory interact during written reflection and measured effects on well-being. Qualitative and quantitative data were collected via surveys, logfiles, and interviews. Follow-up assessments were also administered three months after the initial month-long study to measure potential long-term benefits. When participants reflected on memories with valences opposite to their current mood, their mood became more neutral. However this did not impact overall well-being. My findings also clarify underlying TMR mechanisms, showing that moods and memories competed with each other. When positive moods prevailed over negative memories, people demonstrated classic mechanisms shown in prior work to influence well-being. When negative moods prevailed over positive memories, memories became negatively tainted. My results have implications for new well-being interventions and technologies that capitalize on the interconnectedness of memory and emotion.

Recent research has shown that distrust for a company is not merely the opposite of trust. Rather, it is a distinct construct that exists parallel to trust. Nevertheless, researchers studying distrust for a company have so far typically relied on reverse-worded trust scales which only apply to a single company. I created the Consumer Distrust in Specific Company Scale (CDISCS), a four-subscale instrument which can be deployed flexibly across most companies. Through three studies deploying the scale to participants on Mechanical Turk, I provide evidence for the current reliability and validity of the scale. I demonstrate the scale’s high retest reliability, differential reliability across demographic comparisons, and ability to measure changes in consumer distrust in response to an authentic corporate scandal. I note that the internal consistency, inter-item correlations and inter-scale correlations of the Incompetence, Malevolence, and Intentions subscales are higher than desired, though within the same range as other trust and distrust scales considered validated in prior research. The CDISCS should therefore be considered equally ready to deploy as other published trust and distrust scales that have these features.

Rates of mental health conditions like depression and anxiety have been rising each year, especially during the COVID-19 pandemic. Despite the increasing prevalence of these conditions, barriers to accessing mental health care remain for many people in need of support. Digital mental health technology has proliferated in response to the increased rates of mental illness, as well as with the widespread adoption of smartphones. While prior work has rightly focused on the evaluation of the efficacy and evidential basis of these apps, there is a need for an examination of the subjective user experience of mental health apps. However, many apps are not supported by an evidential basis and, even highly valid apps struggle with low user engagement, as the adoption of technology is not necessarily driven by validity but rather user experience. The flexibility and constant availability of apps provide users with on-demand support not accessible through therapy or other traditional means of support. Conversational agents (chatbots), in particular, are increasingly better at simulating naturalistic interactions similar to those one would receive from a therapist or close friend. This paper describes two thematic analyses conducted on user reviews of 39 health and wellness apps, with a deep dive into six chatbot apps, as these are increasingly developed and downloaded. We discuss user-perceived benefits general to these types of apps – such as the 24/7 availability, social and motivational benefits – as well as benefits specific to apps implementing conversational agents, many pertaining to the development of some sort of advantageous relationship with a chatbot. We suggest implications for the future design and research of mental health apps.

Curiosity is a core drive for learning in humans which is increasingly being looked at in developing robots capable of internally motivated lifelong learning. However, there are also potential risks associated with robots being curious when people do not expect or want robots to be curious, especially when the robot already has work tasks to perform. Further, how robots’ mental states are described may prime expectations inhuman counterparts. This thesis presents a pair of experiments on people’s perceptions of four levels of curious robot behavior, designed by a professional animator. We also explored the impacts of a curious robot being seen as on-duty vs. off-duty. In addition,we examined whether our curious robot behavior matched peoples’ expectations when primed to expect a “curious” vs. “learning” vs. “autonomous” robot. In both studies, as curious behavior increased so to did ratings of the robots’ thinking ability coupled with a decrease in ratings of it as an effective working and social agent, particularly when it was framed as on-duty. Further, each cognitive prime resulted in a unique trajectory of expectation matching showing that robots’ internal mental states should be matched with analogous external behaviors. Our findings have implications for the design and development of robots that engage in learning modeled on human cognition.