Scholarlink

<p>Artificial intelligence is rapidly advancing toward a threshold where its structure, reasoning, and behavior will likely exceed the limits of human understanding. In this paper, we introduce the concept of a <i>cognitive horizon</i>—a boundary beyond which human comprehension of AI systems is no longer possible. This dilemma is not due to a lack of effort or technical insight, but because of the fundamental constraints of human cognition. Asking us to fully grasp such systems may be akin to problems whose solutions resist faithful compression into human-interpretable abstractions. We survey philosophical and computational foundations for this claim, including formal results in complexity theory and epistemology. We then examine early empirical signals from large-scale models that suggest this horizon is not speculative but rapidly approaching. Rather than framing this opacity as a failure, we argue for recognizing incomprehension as a design and governance variable in its own domain. If artificial general intelligence develops along trajectories we cannot fully interpret, preparing for that reality will require rethinking how we approach trust, responsibility, and collaboration with systems that may ultimately think in ways totally alien to our own.</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

Approaching the cognitive horizon: recognizing the limits of human comprehension in advanced AI systems

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • John T. Sinnott,
  • Yalitza Martinez,
  • Yumeng Zhang

摘要

Artificial intelligence is rapidly advancing toward a threshold where its structure, reasoning, and behavior will likely exceed the limits of human understanding. In this paper, we introduce the concept of a cognitive horizon—a boundary beyond which human comprehension of AI systems is no longer possible. This dilemma is not due to a lack of effort or technical insight, but because of the fundamental constraints of human cognition. Asking us to fully grasp such systems may be akin to problems whose solutions resist faithful compression into human-interpretable abstractions. We survey philosophical and computational foundations for this claim, including formal results in complexity theory and epistemology. We then examine early empirical signals from large-scale models that suggest this horizon is not speculative but rapidly approaching. Rather than framing this opacity as a failure, we argue for recognizing incomprehension as a design and governance variable in its own domain. If artificial general intelligence develops along trajectories we cannot fully interpret, preparing for that reality will require rethinking how we approach trust, responsibility, and collaboration with systems that may ultimately think in ways totally alien to our own.