![[ Visit Client Website ]](images/banner.gif)
Former Surgeon General Vivek Murthy identified social media as a significant factor contributing to youth mental health, citing platform designs linked to harmful content, cyberbullying, and higher rates of depression and anxiety. Yet this is only part of the story. Recent Pew data show that 74% of teens say social media helps them feel more connected to friends and 63% say it gives them a place to show their creative side. This tension between harm and support points to an underexplored dimension of digital life: joy.
Research on AI-mediated emotional support has focused mainly on response to distress, depression, and risk. Building on AI with Heart (Patton et al., 2025), we advance the term, enduring joy. We define our interest in enduring joy as a contextual, culturally grounded state marked by freedom, safety, ease, connection, and resilience, one that can coexist with and emerge through struggle. This definition draws on Fredrickson’s broaden-and-build theory and Lorde’s account of joy as tied to survival and resistance. We then ask whether large language models (LLM) can recognize and respond to enduring joy in young people’s social media posts without collapsing it into happiness, and whether evaluation methods can distinguish this from mere flattery or over-affirmation.
Methods
We conducted a study collecting self-expressions of emotion on Reddit, including expressions of joy, distress, and their co-occurrence. We will further assess the prevalence of enduring joy in collected emotion-laden texts by presenting them to multiple LLMs (e.g., GPT-4, Claude). We will evaluate whether models can identify joy embedded in distress or whether they default to generic affirmation, studying the impact of model architectures and prompt conditions.
Results
To date, we have carried out a study identifying a larger set of emotions than is usually studied, grounded in speakers’ own descriptions (Deas et al., 2024). This set contains a nuanced set of affective states that may correspond to enduring joy. We identify over 1,600 emotions; some examples of the emotions that we find include: giddy, disconnected, ecstatic. Our next steps will be qualitative analysis of the positive emotions to identify those that correspond to long-lasting joy and adapting our models to work with African American Language.
Conclusions and Implications
This study contributes to a growing body of work challenging deficit-centered approaches to AI by asking whether AI systems can be built to recognize and support the full emotional range of digital life. For social work research, the implications are significant. If AI-mediated tools are increasingly present in the digital spaces where youth seek support, such as mental health chatbots, they must be capable of recognizing and responding to the presence of resilient joy.
Citations
Deas, N., Mejia, I., Turcan, E., McKeown, K. (2024) MASIVE: MASIVE: Open-Ended Affective State Identification in English and Spanish. EMNLP 2024.
Patton, D.U, Kleiner, S., Deas, N., Grieser, J., Shepard, J., Vente, B., McKeown, K. (2025) AI with heart: A joy-centered framework for ethical AI development & implementation. The Paris Journal on AI & Digital Ethics.