Scholarlink

<p>Human-in-the-loop (HITL) is widely regarded as a fundamental safeguard in AI system design, ensuring that human judgment remains the final arbiter of consequential decisions. This paper argues that HITL undergoes a structural corruption driven by two independent forces that intensify precisely as AI systems become more capable. The first force, the <i>verification penalty</i>, describes how rising AI accuracy makes the act of human verification increasingly costly relative to its yield, creating organizational selection pressures that systematically eliminate verifying humans from the loop. The second force, <i>pre-skill deprivation</i>, identifies a generational mechanism whereby individuals who have never performed tasks independently lack the experiential foundation needed to evaluate AI outputs—a condition qualitatively distinct from the deskilling of existing practitioners. These two forces converge to produce a paradox: HITL becomes least functional precisely when it is most needed—that is, when AI systems are highly accurate and human dependence is deep. The paper introduces the Bloom-Delegation Ladder as a framework for mapping which cognitive capacities have been delegated to AI systems and which remain as the basis for human oversight. The analysis demonstrates that HITL, absent structural intervention, tends to transition from a guarantee of safety to an illusion of safety.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The structural corruption of human-in-the-loop: how AI competence undermines its own safety architecture

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Kenji Yamada

摘要

Human-in-the-loop (HITL) is widely regarded as a fundamental safeguard in AI system design, ensuring that human judgment remains the final arbiter of consequential decisions. This paper argues that HITL undergoes a structural corruption driven by two independent forces that intensify precisely as AI systems become more capable. The first force, the verification penalty, describes how rising AI accuracy makes the act of human verification increasingly costly relative to its yield, creating organizational selection pressures that systematically eliminate verifying humans from the loop. The second force, pre-skill deprivation, identifies a generational mechanism whereby individuals who have never performed tasks independently lack the experiential foundation needed to evaluate AI outputs—a condition qualitatively distinct from the deskilling of existing practitioners. These two forces converge to produce a paradox: HITL becomes least functional precisely when it is most needed—that is, when AI systems are highly accurate and human dependence is deep. The paper introduces the Bloom-Delegation Ladder as a framework for mapping which cognitive capacities have been delegated to AI systems and which remain as the basis for human oversight. The analysis demonstrates that HITL, absent structural intervention, tends to transition from a guarantee of safety to an illusion of safety.