Does Human Oversight of AI-Assisted Decisions Work? A Systematic Review of the Evidence and a Framework for Meaningful Human Control

Authors:

Abstract
Human oversight has become the default safeguard for AI-assisted decisions: the EU Artificial Intelligence Act, data protection law and many organizational policies require a human to review or override algorithmic outputs. Whether such oversight gives decision makers meaningful control, and whether it improves decisions, is far less clear. We conducted a PRISMA 2020 systematic review of 186 studies from Scopus and Web of Science on the governance of AI-assisted decision-making and graded each study by the strength of evidence its design can provide. Governance was conceptualized mainly through accountability (96% of studies) and transparency (94%), and human-centered control appeared in 70% of studies, yet its measurement was fragmented and human control was rarely measured as a single construct. Although 50 studies reported positive relationships between governance and outcomes, only three (1.6%) used designs able to support causal claims; both causal studies measured perceived legitimacy rather than decision quality, and neither isolated override, calibrated reliance or contestability. The evidence for human oversight is therefore largely normative and associational. We decompose meaningful human control into five measurable components (information and understanding, decision authority, capacity to override, competence and timing, and answerability), propose an eight-component framework with seven testable propositions, and set out a research agenda for testing whether oversight works in practice.