Google DeepMind wants to know if chatbots are just virtue signaling (www.technologyreview.com)

Google DeepMind is emphasizing the need for rigorous evaluations of the moral behavior of large language models (LLMs), similar to how their coding and computational capabilities are assessed. As LLMs increasingly play sensitive roles—acting as companions or advisors—understanding their trustworthiness in moral decision-making becomes crucial. Research scientists William Isaac and Julia Haas argue that while LLMs have shown impressive moral competence in studies, distinguishing genuine moral reasoning from mere imitation poses a significant challenge. The team proposes new, more stringent testing methodologies to explore the robustness of moral reasoning in LLMs. These include tests that assess whether models flip their answers based on rephrased moral dilemmas, indicating a lack of genuine reasoning. They also highlight the need for models to articulate their reasoning processes and adapt to various cultural and ethical perspectives, considering the diverse values of global users. By addressing these challenges, DeepMind aims to enhance the moral competency of LLMs, which could improve overall AI systems' alignment with societal values and ensure their responsible deployment.