Emotionally Aware AI Models: An Evaluation-and Safety-Centric Survey
Published in ResearchGate Preprint, 2026
This survey reframes emotionally aware AI around evaluation and safety. We analyzed existing generative models, datasets, and benchmarks for evaluating emotional intelligence in affective computing systems, and based on the analyses, propose the claim that emotional awareness and manipulation are deeply connected, and evaluations of such models should include more critical details which are currently overlooked.
Specifically, the mechanisms and metrics that certify supportive behavior can certify sycophancy. The open problem is specifying honest, non-dependence-fostering behavior precisely enough to evaluate.
Recommended citation: Nicole Hao. (2026). "Emotionally Aware AI Models: An Evaluation-and Safety-Centric Survey." ResearchGate Preprint.
Download Paper