Abstract
Artificial Intelligence (AI) is increasingly embedded in publicly funded social science research, reshaping how questions are formulated, evidence is generated, and policy recommendations are constructed. Although Responsible Innovation (RI) is institutionalised within UK research governance, it is frequently framed as a set of normative principles or compliance criteria. Such approaches are insufficient for AI-enabled research, where socio-ethical issues emerge dynamically within evolving analytical processes. This paper reconceptualises responsible AI as a multilevel organisational capability that develops through sustained engagement with persistent innovation dilemmas. We develop a multilevel, dilemma-informed framework that locates responsibility across governance, organisational, and methodological levels. Governance structures articulate expectations around ethics, inclusion, transparency, and impact; research teams interpret and translate these expectations within project contexts; and responsibility is ultimately enacted through technical configurations, including model design, bias mitigation, explainability, and data governance practices.
We argue that AI-enabled research generates enduring tensions—between analytical performance and transparency, scale and inclusivity, anticipation and normative restraint, and speed and reflexive responsiveness. These tensions are structural rather than episodic, and responsible innovation emerges through iterative negotiation rather than resolution. Empirically, we examine a large-scale portfolio of ESRC-funded AI-enabled projects using text mining and computational analysis of Gateway to Research Plus data, complemented by qualitative case analysis. This mixed-methods approach provides initial validation of the framework by identifying how responsibility is articulated and operationalised across funding, project, and methodological layers. The paper contributes to innovation theory by reframing responsible AI as a dynamic capability embedded within research systems.