HOW DO THE DESIGN CHOICES, DATASET CONSTRUCTIONS, AND EVALUATION PRACTICES USED IN FACIAL EMOTION RECOGNITION RESEARCH LIMIT THE RELIABILITY OF THESE SYSTEMS OUTSIDE CONTROLLED BENCHMARK SETTINGS

Facial emotion recognition systems are commonly evaluated using benchmark datasets and standard classification metrics. High accuracy on these benchmarks is often taken as evidence of reliability. This paper reviews how design choices,dataset construction, and evaluation practices shape that perception. It examines how emotion recognition has remained framed as a single label classification task based on facial appearance, even as modeling approaches have evolved. Through a review of methods, datasets, and evaluation norms, the paper shows that labeling practices compress ambiguity, training objectives enforce decisiveness, and benchmarks reduce variation present in applied use. These factors interact to produce systems that perform consistently under controlled conditions while remaining fragile outside them. The review also examines how interpretation risk increases when model output is treated as emotional inference rather than as label reproduction. By tracing reliability limits across methodological and evaluative stages, this paper clarifies why improvements in model architecture do not translate into dependable behavior beyond benchmark settings.