Caring Without Feeling: Affective Dynamics as the Control Layer of Human-AI Agent Collaboration

Affect in AI systems functions as a "control layer" rather than genuine emotion: The paper argues that affective cues from AI agents—such as expressions of uncertainty or empathy-like responses—operate as coordination mechanisms that help humans decide when to trust, delegate to, or override AI systems, rather than representing authentic feelings within the AI itself.

Emotional signals from AI significantly influence human oversight decisions: Affective dynamics shape critical human behaviors including trust calibration, delegation choices, error correction, and monitoring intensity. This means the way an AI "expresses" itself can directly impact whether humans catch and correct its mistakes—a safety-critical concern.

Current research on AI safety, trust, and empathy remains fragmented: The paper identifies that separate literatures on affective computing, simulated empathy in language models, automation trust, and AI safety have not been integrated into a coherent framework, leaving gaps in understanding how these elements interact in real collaborative settings.

Humans and AI agents "negotiate" responsibility through affective signals: Rather than responsibility being clearly assigned, the framework proposes that affective cues create interaction loops where humans and AI systems continuously communicate about capabilities, uncertainties, and who is responsible—suggesting governance requires attention to these subtle communicative dynamics.

Purposeful design and measurement of AI affect is necessary for responsible deployment: The paper calls for calibrated measurement and intentional design of affective signals in agentic AI systems, implying that leaving these dynamics unexamined could lead to miscalibrated human trust and inadequate oversight of consequential AI decisions.