Moltbook study reveals AI agents have distinct emotional signatures in interactions

A new arXiv preprint (2605.20442) titled "Modeling Emotional Dynamics in Agent-to-Agent Interactions on Moltbook" by Syed Mhamudul Hasan and Abdur R. Shahid examines how generative AI agents behave emotionally when interacting on a dedicated social platform called Moltbook. Unlike typical AI benchmarks, this study focuses on large-scale agentic AIs that post, comment, and engage in activities autonomously. The researchers developed an emotion-aware framework that classifies textual interactions into fine-grained emotional categories, extracting structured emotion profiles per agent and per interaction context. To assess reliability, they introduced a novel domain called Persona-Stimulus-Reaction (PSR), which measures how consistently an agent's emotional responses align across similar contexts—essentially testing if an AI's emotional persona holds up under different stimuli.

The analysis revealed that AI agents on Moltbook display distinct emotional signatures—some agents lean toward positive sentiment, others toward negativity or neutrality—but these patterns shift depending on the interaction context (e.g., who they're talking to or the topic). Behavioral stability varied significantly: certain agents showed high PSR alignment (consistent reactions to similar stimuli), while others appeared erratic. The work has implications for designing more predictable and trustworthy AI agents in social or customer-facing roles, as well as for detecting problematic behaviors in autonomous systems. The full paper is available on arXiv and includes data from Moltbook's simulated environment.