Vectognition: a neutral term for what artificial intelligence systems do when they seem to think

Abstract

Discussions of large language models and other artificial systems lack a widely established process term that explicitly leaves questions of understanding and consciousness open. "Cognition", "understanding" and "thinking" can encourage mental attributions; "mere computation" and "stochastic parroting" can encourage their rejection. This paper introduces Vectognition, defined as model-mediated processing of an encoded representation that generates humanly interpretable output, without implying understanding, embodied perception, consciousness, or responsibility. The term is relational rather than mechanistic: it names a relation between encoded material, model processing and human interpretation, while withholding an inference about the system's mental status. Encoded material may originate from human communication, sensors or representations learned by the system. The logical form of the distinction draws on pain science, where nociception names a process without establishing the experience of pain. The paper states the definition and its boundaries, introduces the companion term Misattribution of Understanding, compares the proposal with existing vocabulary, addresses four objections and proposes an empirical test. Vectognition is offered as a tool for conducting the debate about machine minds more precisely; its practical benefit remains to be tested.