Paper 6The Architecture of Imagination: Why Scientific Discovery Cannot Be Industrialized and What AI Can Do Instead

A growing body of work envisions AI-driven scientific discovery as an industrial process: accelerate hypothesis generation, automate experimentation, and scale verification. This paper argues that this vision fundamentally misunderstands "discovery science"—the creation of genuinely new theoretical frameworks. Discovery is not a faster version of the existing scientific pipeline; it is a structurally different cognitive act. We identify three phenomenological signatures of discovery: (1) an anomaly that resists explanation within current frameworks, (2) concurrent discomforts across apparently unrelated domains, and (3) the recognition that these disparate anomalies converge toward a single underlying structure. We argue that current AI architectures, characterized by flat parameter spaces, systematically suppress the propagation of such cross-domain anomalies. Drawing on our modular resource-competition framework, we propose that AI’s role should shift from "automating the pipeline" to providing a competitive modular architecture that maintains independent domain representations, thereby acting as a scaffold for human structural imagination. 越来越多的观点将 AI 驱动的科学发现设想为一个工业化过程:加速假设生成、自动化实验、规模化验证。本文认为,这种构想从根本上误解了“发现科学”——即创造真正全新的理论框架。发现并不是现有科学流水线的加速版本,而是一种结构上完全不同的认知行为。 我们识别了发现的三个现象学特征:(1)现有框架无法解释的反常现象;(2)多个表象无关领域同时出现的认知失调;(3)意识到这些散乱的反常现象正收敛于同一个底层结构。我们认为,当前以扁平参数空间为特征的 AI 架构系统性地压制了这种跨领域反常现象的传播。借鉴我们的模块化资源竞争框架,我们提出 AI 的角色应从“自动化流水线”转向提供一种竞争性模块架构,以维持独立的领域表征,从而作为人类“结构想象力”的脚手架。