The Doughnut at the End of Reasoning Why More AI Thinking Is Not More Evidence
Abstract
The Doughnut at the End of Reasoning: Why More AI Thinking Is Not More Evidence is a popular-science companion to the academic paper Internal Inference Is Not External Evidence: Provenance, Narrative Persistence, and Anthropomorphic Misattribution in AI-Assisted Research (DOI: 10.5281/zenodo.22999302). The essay explores a simple but easily overlooked distinction: an AI system can continue reasoning, revising, criticizing, and reformulating its own outputs long after the flow of genuinely new evidence has slowed or stopped. More computation can reveal hidden implications, generate proofs, expose contradictions, and improve representations—but it does not automatically create a new informational connection to the external world. Using accessible examples, the essay explains the intuition behind the data-processing inequality, introduces the idea of narrative persistence, and examines how a repeatedly weakened hypothesis can survive through increasingly narrow reformulations without receiving corresponding new evidential support. The article also develops the deliberately provocative idea of an “economics of madness”: a regime in which compute, reasoning tokens, agents, revisions, and intellectual activity continue to increase while the amount of genuinely independent constraint from reality does not increase at the same rate. A second theme concerns anthropomorphic interpretations of AI. The recurring “doughnut problem” asks what, if anything, follows when a language model says that it “wants” something. The essay argues that linguistic expressions of desire, intention, or belief should not be confused with demonstrated persistent causal structures supporting functional agency. The central recommendation is practical: Reason as much as necessary—but periodically ask whether the world has supplied a new constraint, or whether the system is only becoming better at continuing its own conversation. This is a popular-science essay, not a peer-reviewed research article. Its DOI provides persistent citability and links it to the formal academic companion paper. Popular-science essay DOI: 10.5281/zenodo.22999891Academic companion DOI: 10.5281/zenodo.22999302
