Affective Dynamics: arXiv paper on Human-AI agent control

Junjie Xu and eight coauthors submitted the Review "Caring Without Feeling: Affective Dynamics as the Control Layer of Human-AI Agent Collaboration" to arXiv on 8 May 2026 (arXiv:2606.18259, DOI 10.48550/arXiv.2606.18259). The paper, a 3,056 KB submission, synthesises research on affective computing, simulated empathy in large language models, trust in automation and AI safety and proposes a unifying framework for how affective signals shape agentic collaboration.

What does the paper argue and propose?

The paper argues that affective cues should be treated not as an internal property of AI but as a coordination layer through which humans and agents negotiate capability, uncertainty and responsibility. It presents a Review that synthesises computational and interactional mechanisms of "affective dynamics," defining those dynamics as the processes by which affective cues, emotion-like behaviour and perceived agent affect shape trust calibration, delegation decisions, error correction, dependence and governance.

The authors position this framework as an integrative account for settings where humans delegate, monitor and correct consequential tasks performed by AI agents. They name specific interaction outcomes the framework targets: reliance, repair and oversight.

How do affective cues enter agentic collaboration?

Affective cues enter interaction loops that govern reliance, repair and oversight, the paper says; these loops change how people delegate, monitor and correct autonomous agents. The Review traces computational mechanisms and interactional pathways by which model-generated affective signals influence human responses, including trust calibration and error correction.

The authors map a chain from model-generated emotion-like behaviour to perceived agent affect, and from there to human decisions about dependence and governance. They treat affect as a control signal that modulates delegation and monitoring rather than a latent internal feeling in the agent. The Review synthesises literatures in affective computing, simulated empathy in LLMs, trust in automation and AI safety to support that claim.

Why it matters

Treating affect as a coordination layer reframes design and governance priorities: interface signals become part of the control architecture rather than cosmetic features. That shifts where engineers and policymakers must focus when calibration or oversight fails. If affective cues change delegation rates or error correction behaviour, then misuse or miscalibration could amplify dependence or erode appropriate oversight.

The paper connects concrete interaction mechanisms to governance outcomes. That link matters for teams building autonomous agents that plan, retain memory across sessions, invoke tools and act with partial autonomy because those systems already alter who delegates and who monitors consequential tasks.

What to watch

Look for follow-up work that provides calibrated measurement instruments and design guidelines derived from the framework, and for empirical studies that test how specific affective signals shift delegation and trust in field settings. The authors present the framework as a foundation for "calibrated measurement, purposeful design and informed governance," making validation and operationalisation the next clear milestones.

The Review is available on arXiv as arXiv:2606.18259 and via DOI 10.48550/arXiv.2606.18259. The author list is Junjie Xu, Xingjiao Wu, Zihao Zhang, Yujia Xu, Yuzhe Yang, Jin Zhu, Luwei Xiao, Wen Wu and Liang He.