Influence Is Not Warrant: An Architecture-First Account of Human Semantic-Cognitive Intelligence in Human–AI Interaction

Influence is not warrant, and assisted performance is not retained capability. An AI output can change what a person finds accessible and worth considering without giving that person any checked reason to believe it, and a person can perform better with AI assistance while retaining less once the assistance is withdrawn. This conceptual article presents Human Semantic-Cognitive Intelligence (HSCI), an architecture that keeps these quantities apart. HSCI is a route, not a new equation: it orders registered objects of an open equation library (Toledo) from a person's readout of an encountered difference, through meaning, experience, retention, problem formation, accessibility, discovery and discriminating action, into the human–AI dialogue, and on to live possibility, choice, agency, Human Return (unaided retained capability) and Human Conversion (change in a person's capability in the world). At the dialogue the route splits into an influence lane, along which an output moves what is accessible and live for a person, and a warrant lane, along which content acquires a checked status as knowledge. Movement in one lane does not imply movement in the other. One running example — a published field experiment in which students with an unguarded AI tutor did better on practice problems and, relative to a control group, worse on an unaided exam at the end of the same session — is traced through the main nodes and both lanes, with the points where the example goes beyond the study marked as illustrative. Several of the separations HSCI uses are already established in neighbouring literatures — automation deskilling, the continued influence of corrected misinformation, illusory truth despite knowledge, illusions of competence in learning, and the performance–learning distinction — and are credited as such; HSCI's contribution is their placement on one route and the requirement that they be measured jointly. Six propositions with pre-declared, bounded falsifiers and stated design assumptions are set out as future work. No empirical result of the author's is reported. The main limitations are that no decorrelated independent check has run and that the bridge from in-interaction change to retained capability is an open proposal.