Crocodile tears: Can the ethical-moral intelligence of AI models be trusted?

As AI becomes increasingly embedded into every aspect of our lives, there is evidence that people are turning to these systems for guidance on complex issues and moral dilemmas. Whether or not one agrees that people should do so, the fact that they are necessitates a clearer understanding of the moral reasoning of these systems. To address this gap, this paper introduces an ethical-moral intelligence (EMI) framework for evaluating AI models across dimensions of moral expertise, sensitivity, coherence, and transparency. While this paper focuses on moral sensitivity as an initial empirical test, we argue for a comprehensive framework across all four dimensions. We present findings from a pre-registered experiment testing the moral sensitivity of four AI models (Claude, GPT, Llama, and DeepSeek) using ethically challenging scenarios. While models demonstrate moral sensitivity to ethical dilemmas in ways that closely mimic human responses, they exhibit greater certainty than humans when choosing between conflicting sacred values, despite recognizing such tragic tradeoffs as difficult. This discrepancy between reported difficulty and decisiveness raises important questions about their coherence and transparency, potentially undermining trustworthiness. The research reveals a critical need for more comprehensive ethical evaluation of AI systems. We discuss the implications of these specific findings, how psychological methods might be applied to understand the ethical-moral intelligence of AI models, and outline recommendations for developing more ethically aware AI that augments human moral reasoning.