Answers are grounded in this paper and cite their sources.
Theoretical analysis reveals systematic blind branches in AI decision workflows, highlighting the need for branch-specific verification and contestation.
AI-enabled workflows are often described as human-in-the-loop (HITL) even when people review only some output paths. This article calls that pattern asymmetric human oversight (AHO), also describable as an asymmetric human-in-the-loop arrangement: the assignment, timing, depth, or authority of meaningful review varies with the system's output or the route it creates. In automated hiring, for example, applications above a score threshold may reach a recruiter, while applications below it may be rejected without recruiter review. A person participates in the workflow, but not in every consequential branch; an unqualified HITL label therefore overstates the safeguard. The same architecture shapes what the organization can learn. If rejected applications are neither independently checked nor subject to practicable contestation, qualified job applicants who were wrongly screened out may remain invisible. The article therefore distinguishes case-review coverage—whether a person meaningfully reviews a case—from reference-verification coverage—whether an independent determination can test the system's output. It defines a blind branch as a consequential path lacking meaningful case review, adequate evidence about relevant errors, and practicable contestation. Drawing on research concerning human factors, selective labels, verification bias, algorithmic governance, and EU law, the article argues that asymmetry is not inherently defective: proportionate, risk-based review may be preferable to universal checking. Governance should instead be branch-specific, stating which actions can occur without review, how each branch is tested, which errors the resulting evidence can reveal, and how affected people can challenge and correct decisions.
No takes yet. Share an insight, caveat, or question.
George Evgrafov (2026) studied this question.