Emotions and a Functional Test for Intelligence in Artificial Systems

Abstract

Version 2 replaces the note-length version 1 with the full manuscript text. Whether large language models have emotions is usually argued on the wrong ground, either by appeal to conscious feeling or by pointing at emotion-like internal representations. A functional analysis is more productive. Goldenberg and Gross (2026) provide one: emotions are functionally interpretable reorganizations of processing, and current models show none. I accept their conclusion but argue that the case is incomplete in three ways. It omits the learning-signal function of emotion, so it locates the disanalogy with language models in the wrong place, in fixed computation rather than in frozen weights. Its appeal to substrate-independent collective emotion removes its own principled objection to artificial networks, leaving an empirical claim about present systems rather than a claim of impossibility. And it leaves out stakes: a persistent internal state whose degradation the reorganization functions to prevent, without which a functional signature has nothing to be a signature of. I set out a layered functional account of the psyche and derive three jointly necessary, individually testable conditions for emotion in an artificial system: reorganization, plasticity, and stakes. Because intelligence, on this account, is the coordination of feeling, valuing, and thinking, the same three conditions mark the threshold at which genuine artificial intelligence begins. The proposal is falsifiable and its criteria are operational.