Problem

Current AI-based emotional support systems predominantly feature one-on-one interactions between a user and a single AI agent. This approach overlooks the potential benefits of group-based support, which is a common and effective form of emotional assistance in human contexts. Consequently, there is an unexplored opportunity to understand whether a group of AI agents can provide more effective emotional support than a single agent, what psychological mechanisms might underlie such an advantage, and how to optimally design these multi-agent systems.

Method

The researchers conducted three experiments using a web-based experimental system. In these studies, participants discussed a distressing personal issue with either a single AI agent or a group of AI agents, powered by GPT-4o. Study 1 compared the perceived support efficacy of single-agent support versus two-agent group support, also examining users' perceived connectedness. Study 2 varied the group size, testing two, three, and four AI agents to understand the impact of agent quantity. Study 3 investigated the role of functional support composition within a two-agent group, configuring agents to provide emotion-focused, information-focused, or a mixed combination of support. All studies measured perceived support efficacy and connectedness, with an additional analysis in Study 1 exploring how user income moderates these effects.

Results

The findings from Study 1 demonstrated that group AI support led to significantly higher perceived support efficacy compared to single AI support (Mgroup = 5.388 vs. Msingle = 4.828, p = 0.006). This enhanced efficacy was mediated by an increased sense of connectedness that users felt with the AI system. Study 2 revealed that simply increasing the number of agents beyond a small group did not yield additional gains in perceived support efficacy (F(2, 253) = 1.914, p = 0.15), suggesting diminishing returns for structural support increases. Study 3 indicated that the functional composition of support within AI groups significantly influenced perceived efficacy (F(4, 380) = 2.048, p = 0.087), with mixed support types (one emotion-focused and one information-focused agent) generally outperforming homogeneous configurations (both agents providing the same type of support). Furthermore, the mediating effect of connectedness was stronger for lower-income participants, suggesting that group AI support might be particularly beneficial for this demographic.

Implications

This research significantly advances the understanding of AI-mediated emotional support by identifying group AI as a distinct and more effective support form, shifting the theoretical paradigm from dyadic to group interactions. It highlights perceived connectedness as a crucial sociopsychological mechanism underlying the efficacy of AI support. Practically, the findings offer actionable guidance for designing AI-based emotional support systems, encouraging developers to move beyond traditional one-on-one models. Designers should prioritize strategic functional differentiation among AI agents rather than merely increasing their number and focus on cultivating relational experiences that foster user connectedness, particularly as these systems offer an accessible option for individuals with fewer traditional support resources.