Operationalising Harmlessness in Online AI Systems

Although the harmlessness techniques for foundation models have been explored in the industry, existing methods struggle to adapt to increasingly stringent legal requirements. We observe that as concerns from governments and the public about the harmfulness of automated decision-making by foundation models continue to grow, the interpretation of legal requirements for explainability and fairness is evolving. This requires the industry to rethink how to achieve compliance, ensuring that technologies and applications can be effectively implemented in practice. In this paper, we propose an operationalising research framework to enhance the harmlessness of foundation models, consisting of three major research questions: exploring how foundation models amplify the harmfulness of datasets, investigating harmlessness techniques under strict black-box conditions, and examining proactive explanations for foundation models. Based on this framework, we discuss related literature and our advocated methodologies, introducing an initial exploration: a method that ensures the effectiveness and fairness of algorithmic predictions without requiring sensitive information. Finally, we outline future work to be conducted.