Research Ethics Committees (RECs) perspectives on large language models and AI ethics review: a South African case

The rapid integration of Large Language Models (LLMs) into health research presents a dual promise: to serve as an epistemic equalizer by lowering barriers to scientific participation, and a significant governance challenge, risking a "validity gap" where output volume outpaces integrity. Research Ethics Committees (RECs), tasked with safeguarding research, face a critical "purview weakness" in evaluating these opaque technologies. This tension is most acute in African contexts, where the need for innovation intersects with vulnerability to algorithmic colonization and where empirical data on REC preparedness is absent.

A qualitative study employing in-depth, semi-structured interviews was conducted with a purposively sampled cohort of 14 REC chairs, members, and research ethics office staff from health science institutions across South Africa. Data were collected between January and July 2024, transcribed verbatim, and analysed using inductive thematic analysis.

Analysis yielded five central themes namely: (1) understanding of LLMs, (2) perceived benefits of LLMs use in health research, (3) perceived challenges and concerns in the use of LLMs in health research, (4) human oversight - augmentation, not automation, and (5) fragmented AI literacy within the research. Respondents recognized significant benefits in administrative efficiency, research lifecycle support, and democratizing writing skills. Crucially, there was unanimous consensus that LLMs must only augment, not automate, ethics review. Human oversight was deemed irreplaceable for contextual understanding, empathy, and complex moral deliberation qualities respondents implicitly aligned with relational, Ubuntu-informed ethics.

South African ethics gatekeepers perceive LLMs as powerful but risky tools. Their insistence on human-centric governance, rooted in contextual and communal values, provides a vital counter-narrative to purely technocratic oversight models. These findings provide an urgent empirical foundation for developing context-specific AI governance guidelines in African and other LMIC health research systems.