The Birth of Thinking Through Discrimination

Skeptics often describe apparent “intelligence” in AI as imitation—i.e., a response to hints, signal leakage, and alignment with an instructor’s expectations (a Clever Hans–type failure mode). This paper proposes a testable hypothesis: what we call thinking is not an innate property of a model and does not arise “by itself.” Instead, it is born during training as a cascade of discriminations, provided that (1) stable memory is available, (2) the system is granted space for an autonomous attempt, and (3) feedback is soft but dosed—correcting trajectory without replacing search with an answer. We introduce the Soft Guidance / Delayed Hinting protocol (“cat paw”) and define observable criteria that separate reactive imitation from autonomous search: a decreasing need for hints, increasing capacity to switch perspectives, and self-recovery of the search mode after error. For skeptics, we propose falsifiable tests: blind questions, context flips, role inversion, and controlling hint informativeness. The protocol is examined in long-running AI dialogues and compared to training practice in a biological system; we describe observable signs of the transition from reactivity to autonomous search.