Human–AI Collaboration in Occupational Risk Prevention: A Scoping Review of Explainability, Decision Authority, and Safety Governance

Industry 5.0 promotes human-centric industrial safety, yet it remains unclear whether AI-based risk-prevention systems support meaningful human decision-making or merely leave humans nominally in control. Following PRISMA-ScR, this scoping review examined 65 studies published between 2021 and 2026 (20 from an initial search and 45 from an updated search). Human authority was almost universally retained: 98.4% of the 61 implemented systems left the final decision to humans, and none operated autonomously. However, only 24.6% specified the human role, and 34.4% asked humans to decide on black-box outputs. The evidence was predominantly retrospective or simulated: only three studies were evaluated in operating workplaces, one assessed users’ responses, and none measured preventive effect comparatively. A strong reported preventive contribution was associated with operationally grounded evidence (Fisher’s exact p = 0.004; Holm-adjusted p = 0.019), whereas explainability was not (p = 0.602). AI acted mainly at the hazard source (53.8%, mostly through retrospective analytics) or the receptor (32.3%), rarely on the exposure pathway (4.6%). Only 3.1% of studies cited ISO 45001, and 75.4% did not embed any safety-management standard or process, revealing a reporting gap between technical oversight and formal safety governance. We propose a nine-item minimum reporting set to address these gaps.