Research on Improving Ethical Sensitivity for Ethical Decision-Making in Conversational AI

The development of large language models has significantly advanced the inferential capabilities of artificial intelligence (AI), surpassing human-level performance. Despite the rapid growth in AI's cognitive abilities and the consequent expectations for high-level ethical judgments, ethical issues have increased. This indicates a heightened risk of bias as AI models scale up and train on vast amounts of general data that inherently include social conventions related to gender, race, politics, and religion. This study proposes methods for enhancing ethical sensitivity to social bias. To achieve this, we defined 20 categories of social bias and developed a model that predicts the ethical sensitivity of sentences by leveraging the influence scores of words within these categories. The ethical sensitivity prediction model was validated using a paired-sample t-test, comparing the ethical sensitivity evaluations of 25 AI-generated responses assessed by both AI and human evaluators. The test revealed no significant differences between the two groups, thus confirming the validity of the model. The findings of this study suggest that recognizing and predicting the ethical sensitivity of utterances concerning social biases can enhance ethical sensitivity, mitigate the risk of bias, and contribute to more ethical decision-making in AI interactions.