Human Oversight in Agentic AI: Explainability, Anthropomorphism, and the Boundary of Cognitive Forcing
Abstract
Explainability helps users assess evidence behind an AI recommendation, but it may not change decisions when that evidence goes unexamined. The problem is sharper in agentic decision support, where one assistant interprets evidence, explains its assessment, and recommends an action, while anthropomorphic delivery may add social credibility. Cognitive forcing, operationalised here as Human Intervention Checkpoints (HICs), requires users to specify what the assistant should examine before it recommends. We report a randomised 2 × 2 × 2 hiring experiment (N = 211) in which recruiters assessed a candidate whose unmet role-critical requirements were stated in every recommendation. Advancement was not associated with explanatory provenance, anthropomorphic delivery, HIC presence, or opening source documents. Post-task perceived explainability was associated with perceived anthropomorphism and trust, and trust was associated with advancement. In HIC-present conditions, 70 participants directed the assistant to a role-critical requirement and received Hold, yet 46 advanced. The findings identify a boundary of cognitive forcing: influence over what the agent evaluated and recommended did not reliably carry through to the final human decision.