<p>Governments are increasingly experimenting with Artificial Intelligence (AI) innovations as tools for improving the efficiency of governance processes in various aspects of public administration, including information gathering, processing and decision making. While AI has the potential to support the objectives of good governance, it has also raised legal, ethical and democratic concerns about the technology’s opacity, biases, accountability deficits, and adverse impacts on human rights of individuals. In response to these concerns, many governments have established AI regulatory sandboxes as experimental governance frameworks for enabling supervised AI experimentation within controlled regulatory environments. However, the conventional regulatory sandboxes were designed as experimental regulatory regimes for testing private sector innovations seeking market entry under regulatory uncertainty, rather than public sector innovations seeking institutionalization in governance frameworks. The question explored in this paper is whether traditional regulatory sandboxes (which are designed for balancing the innovation promotion and consumer protection objectives of State regulation of private market innovations) are adequate experimentalist governance frameworks for public sector AI experimentation. This paper argues that adopting AI regulatory sandboxes as experimentalist governance frameworks for public sector AI experimentation requires substantial normative, legal and institutional adaptation of the conventional regulatory sandboxes, to ensure that the legal, ethical and democratic concerns of public sector AI experimentation are adequately addressed. The paper explores this argument by analysing the challenges that public sector AI experimentation poses for the traditional AI regulatory sandboxes. It then proposes a framework for adapting the conventional regulatory sandboxes to support legally and democratically legitimate AI experimentation in the public sector.</p>