From Feeling to Form: Co-Created AI Art for Emotional Support
In everyday life, many people live with ongoing emotional strain, anxiety, or low mood, yet do
not meet the threshold for clinical intervention. What they truly need is not intensive treatment,
but an outlet that can touch and release their inner feelings. Traditional creative art therapy
is widely recognized for its healing value in this regard: the act of creating helps individuals
externalize inner emotions and revisit and reorganize the self through visible form (Stuckey and
Nobel, 2010; Malchiodi, 2020). However, this path is not accessible to most people (Fancourt and Finn, 2019). The cost of professional studios, fixed appointment schedules, implicit expectations of drawing or crafting skills, and the anxiety of “not being good at art” often block people before emotional regulation even begins (Haiblum-Itskovitch et al., 2018; Kaimal et al., 2016; Moon, 2016). More fundamentally, traditional art-making is slow, while emotions move quickly. When someone is under acute stress or caught in complex emotional states, the pace of manual creation often cannot keep up with the speed of inner change (Chang et al., 2025).
Conversational AI built on large language models has substantially lowered barriers to accessing
emotional support (Epstein et al., 2023; McGuire et al., 2024; Luo et al., 2024). However, it faces
a structural limitation that technical improvements alone cannot resolve. Langer (2009) distinguishes between two forms of symbolic representation: discursive symbols, which are sequential and propositional, and presentational symbols, which convey meaning through simultaneous and holistic form. Language operates exclusively through the discursive mode. Because discursive representation requires experience to be broken into discrete units and arranged according to logical and syntactic rules, it is poorly suited to express the non-propositional and pre-verbal qualities of affective experience (Gendlin, 1997). Conversational AI inherits this constraint. Regardless of improvements in fluency or empathic tone, systems that rely on language remain bounded by the expressive limits of discursive representation. Art, in contrast, operates through presentational symbols and can therefore convey forms of emotional meaning that resist linguistic articulation (Van der Kolk, 2014; Arnheim, 2023). This distinction implies that the limitations of language-based emotional support are not contingent on the current state of AI technology but are intrinsic to the representational medium itself.
The present study addresses these gaps through lab experiments. We first test whether GenAI
art co-creation leads to greater reductions in negative affect relative to control conditions. We
then examine the psychological mechanisms underlying this effect by evaluating two mediators.
The first is symbolic self-expression, defined as the extent to which users can externalize and give perceptible form to their emotional experience through visual co-creation. The second is selfdistancing, defined as a shift from an immersed, first-person perspective on a negative experience to a more observational, third-person stance, which facilitates emotional reappraisal and reduces affective reactivity (Lazarus and Folkman, 1984).
At the moderation level, this study manipulates two designable tool features as moderators.
The first is AI output unexpectedness, operationalized as variation in image style ranging from
literal and concrete to abstract and metaphorical. This manipulation reflects a counterintuitive
theoretical claim: deviation between AI output and user expectation—typically framed as a
design flaw—may, in emotional support contexts, be the very source of regulatory value in the
co-creation process (Gaver et al., 2003). Such deviation can interrupt existing emotional frames
and create conditions conducive to reappraisal or self-distancing (Epstein et al., 2023; Lazarus,
1991).
The second moderator is choice variety, operationalized as the number of style options available to users (e.g., three versus fifteen). Choice variety shapes the depth of exploration and engagement during co-creation. When style options are limited, users’ creative space is constrained, reducing their ability to iteratively align internal feelings with external visual representations. In contrast, a broader set of options enables users to actively search across a wider visual language for forms that resonate with their emotional experience. This process of active exploration may itself strengthen symbolic self-expression or facilitate self-distancing. By experimentally testing these two mediating pathways, this study clarifies the mechanisms through which GenAI art co-creation operates in emotional support, which remains largely unexamined in the existing literature. For designers, identifying which tool features (e.g., output unexpectedness or choice variety) activate these pathways has direct implications for design decisions, shaping how emotional support systems should be configured to achieve their intended psychological effects.
Yi Wu is a second-year Ph.D. student in the Department of Economics at the University of
Texas at Austin. Her research interests lies on the use of chatbots and other AI systems in
healthcare contexts; the adoption of emerging technologies; and how these technologies shape
human cognition, emotion, creativity, and behavior.
We use cookies to analyze website traffic and optimize your website experience. By accepting our use of cookies, your data will be aggregated with all other user data.