GAILC Working Paper 002: The Intervention Test — When Human Oversight Meets AI-Mediated Decision-Making, Can a Human Actually Change the Outcome?

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

As artificial intelligence systems become increasingly capable of generating recommendations, making classifications, initiating actions, allocating resources, and influencing consequential decisions, organizations increasingly describe these systems as operating with a human "in the loop," "human oversight," or "human review." Yet the existence of a human in an AI-mediated workflow does not establish that meaningful human control exists. A human may be present but lack authority to intervene. They may possess authority but lack the technical capacity to alter the system. They may possess both authority and technical capacity but receive insufficient information to understand what requires intervention. They may understand the system but receive the relevant signal too late to prevent an irreversible consequence. They may intervene without sufficient competence. And an intervention may occur without sufficient provenance to establish afterward what was changed, why it was changed, by whom, and with what effect. This Working Paper introduces the Intervention Test: a GAILC research framework for examining whether claimed human oversight constitutes genuine intervention rather than merely human presence. It develops a distinct line of inquiry from GAILC Working Paper 001, which examined the economics of authenticated human participation in synthetic-media contexts. WP002 asks a different question: not whether a real human originated a piece of content, but whether a human formally responsible for supervising an AI-mediated decision can actually alter its outcome. The framework investigates eight dimensions: Human Intervention, Intervention Authority, Intervention Capacity, Intervention Timing, Intervention Information, Intervention Competence, Intervention Provenance, and Intervention Confidence. Central distinction: Human presence is not human oversight. Human oversight is not necessarily human intervention. Human intervention is not necessarily effective control. The framework does not attempt to determine legal liability. Instead, it provides an evidentiary architecture through which organizations, researchers, regulators, auditors, courts, and other investigators may interrogate claims of human oversight in AI-mediated systems. The intervention question is proposed to remain empirically open to refinement through application across GAILC's Counter-Brief series. Framework status: Evolving research methodology; not a legal standard, certification, regulatory determination, or liability test. Version 0.2 (Revised), August 2026.