Should We, Like Artificial Intelligence, Sometimes Rewrite Our Thinking From Scratch? What Separates Machine From Human

Abstract

This article begins with a simple observation: midway through a long conversation, an artificial intelligence sometimes suggests starting the chat afresh. Rather than taking the model's own explanation for this suggestion at face value, the article turns to an external measure of performance. Drawing on Laban et al. (2025), it shows that when information reaches a language model piecemeal rather than all at once, the resulting drop in performance comes mainly not from reduced capability but from instability: the model anchors on whatever arrives first. The article then asks whether we are entitled to extend this finding to human thinking, and, drawing on an independent literature on anchoring and adjustment in cognitive psychology (from Tversky and Kahneman to recent work on language models), shows that such licence exists, at least at the level of behavioural pattern. In humans, however, this anchor can become attached to something the machine lacks: a self. From here, two layers are distinguished: a cold layer, shared by model and human, and a warm layer, found only in humans, in which the initial anchor turns into identity-linked commitment. The article's practical conclusion is that neither the inner sense of steadfastness nor the act of questioning oneself offers a reliable test for distinguishing sound conviction from premature commitment. Only a behavioural test, independently reconstructing the problem and comparing the result, can make that distinction.