Delegating research functions to artificial intelligence: a risk-control framework for validation, human control and accountability

Artificial intelligence is moving from communication support into consequential research functions, but governance often focuses on disclosure after use rather than the prior decision to delegate. This conceptual, evidence-informed paper proposes a 12-gate Research AI Delegation and Risk-Control Framework for prospective decisions about whether, when and under what controls a defined research function should be performed by an AI-enabled system. The framework evaluates the specific AI-function-context combination and links purpose, permission and task definition to inherent risk, task-specific validity, reliability, human control, control effectiveness, residual risk, provenance, reporting and accountability. Its central distinction is between assistance and delegation: delegation occurs when AI performs a function that would otherwise require independent human judgment or materially determine a consequential research decision. Complementing frameworks concerned with proposal evaluation, institutional governance, oversight design and documentation, the proposed architecture treats delegation itself as the governance decision and requires evidence and controls proportionate to the consequences of error. It also operationalises provenance through an AI Research Notebook and distinguishes AI-based error surveillance from external validation. The framework is intended as a practical decision architecture for researchers, evidence-synthesis teams, institutions, editors and governance bodies. It is not a validated appraisal instrument. Prospective validation should test whether its gates improve decision consistency, error detection, auditability and safe substitution of human research functions.