From disclosure to epistemic accountability: preserving meaningful human control across the AI-assisted research lifecycle
Year: 2026
Type: article
Abstract: Artificial intelligence (AI) is becoming embedded across the scientific research lifecycle, while prevailing governance remains centered on disclosure of tool use and the assignment of responsibility to human authors. We argue that disclosure is necessary but insufficient because it does not establish that researchers understood, verified, challenged, and authorized the AI-assisted processes that shaped scientific claims. We define this disconnect as the epistemic accountability gap and develop the Lifecycle Epistemic Accountability Framework (LEAF) as a testable theory of how meaningful human control can be preserved across AI-assisted knowledge production. Drawing on a structured integrative review and critical interpretive synthesis, LEAF links four human-control capacities—competence, contestability, intervention authority, and traceable ownership—supported by evaluability by design, to five lifecycle checkpoints spanning evidence identification, study design and ethics, data processing and analysis, interpretation, and reporting or knowledge translation. The framework proposes that accountability improves when oversight is exercised prospectively at epistemically consequential decision points, documented through an audit trail, and calibrated to epistemic materiality: the extent to which AI assistance can alter the validity, interpretation, or consequences of a scientific claim. Six propositions specify testable relationships among disclosure, epistemic materiality, meaningful human control, lifecycle checkpoints, auditability, and defensibility of scientific claims. LEAF therefore shifts AI research governance from retrospective visibility to prospective, risk-proportionate control. We identify boundary conditions, implementation pathways, and an empirical research agenda for testing the framework across disciplines and resource settings. LEAF's distinctive explanatory mechanism is conditional: epistemic materiality escalates oversight, evaluability enables four human capacities, lifecycle checkpoints interrupt error propagation, and retained audit evidence makes consequential decisions answerable.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,418
OpenAlex ID: W7225744117