Cognitive Effort in Large Language Models: A New Mirror for Seeing Human Reasoning

A recent capability released in one large language model lets the user set the amount of computational "effort" the model spends, independently of which model is chosen. For the first time, this feature separates two axes of reasoning that have always been intertwined in humans, presenting them as distinct parameters: the ceiling of capability, and the quantity of resources spent on a given problem. This paper uses that technical separation as an analytical tool for seeing the mechanism of human reasoning more clearly. We first describe machine reasoning on the basis of evidence and propose a three-dimensional framework: capability, effort, and calibration (the degree to which confidence matches correctness). We then adopt an explicit methodological position, namely that the difference between machine and human traces back to their material substrate while the structure of their reasoning is assumed, within reason, to be alike, and we look for these three dimensions in humans. We argue that all three are identifiable in humans, but with one difference: what is independent and neutral in the machine is, in humans, entangled with motivation and identity. The focus of this entanglement lies in the dimension of effort, which we call "the lever problem". In the machine, an external and impartial agent sets the lever of effort, whereas in humans an interested mind holds it. Key Highlights of the Paper: Motivated Reasoning Redefined: Reframes motivated reasoning not as a flaw in the quality of logical steps, but as an asymmetric allocation of effort in favour of a desired conclusion. The Lever Problem: Explores how the separation of the "lever of effort" from the reasoning process in machines acts as a mirror to expose the hidden, interest-driven nature of human cognitive effort. Historical Perspective: Argues that the history of wisdom can be read, in part, as a history of the unequal distribution of the opportunity for intellectual effort, a scarce resource that machines are now making universal. The paper claims no proof; it offers an analytical framework and several open questions. Perhaps the value of the machine lies not in its resemblance to us, but in that very difference: a mirror of another kind, in which we see, in a decomposed state, what in us was always unified and hidden.