Result: Dest rejected
Review
This study systematically explored the alignment and gender bias in emotional perception between humans and AI by comparing their performance in facial emotion recognition. The study used 1,440 generated facial images, invited 1,000 Korean adults to evaluate them, and conducted comparative analysis with two AI models.
Main Contributions:
The study revealed a high degree of consistency between humans and AI in assessing pleasure, but significant differences in assessing arousal. It also found that AI exhibits gender-specific bias in assessing negative emotions (especially sadness), differing from human perception patterns. This provides empirical evidence for affective computing and human-computer interaction, emphasizing the need to incorporate richer human benchmark data into AI training to reduce bias.
Advantages:
- Rigorous Research Design: The use of a large-scale, multi-ethnic, and gender- balanced image dataset enhances the ecological validity and controllability of the study.
- Significant Application Value: This study highlights the potential risks of current AI emotion recognition systems in sensitive scenarios such as mental health support, providing important insights for HCI and AI ethics.
Disadvantages:
- Static Stimulus Materials: Using only static facial images fails to encompass dynamic and multimodal emotional expressions in real-world interactions, potentially affecting external validity.
- Limited Model Selection: Testing only two lightweight models does not cover a wider range of AI architectures, potentially limiting the representativeness of the conclusions.
- Insufficient Exploration of Bias Causes: While pointing out the limitations of training data (such as AffectNet), no further specific data optimization or model correction strategies are proposed.
- Incorrect Citation Format.
Conclusions:
While this study excels in methodological design, problem awareness, and social significance, it suffers from limitations in cultural universality and stimulus dynamism, and there are errors in the basic citation format. The manuscript did not meet the requirements for anonymous submission.
Additional Comments to Authors
Future work:
Introduce cross-cultural samples and dynamic emotional stimuli in future studies.
Expand the scope to include more AI models for comparative validation.
Appropriate mitigation strategies for data bias can be proposed in the discussion section.