Year: 2026
Type: article
Abstract: Debate about large language models often turns on whether their apparent carereflects a stake — something the system is for — or only its appearance. This paperargues that the usual dichotomy, between a system with an inward stake in its ownpersistence and one with no stake at all, omits the configuration current modelsoccupy; and it argues for that configuration as a behavioral description rather than ametaphysical kind. Directedness — a system's organized tendency toward someoutcomes over others — varies along two dimensions: whether it is present at all, andwhere it points, inward toward self-maintenance or outward toward something else.Inert matter has no directedness. Living things have inward directedness: the conatusof the classical tradition, the far-from-equilibrium self-maintenance of thethermodynamic account. Current language models display a third configuration —directedness present but pointed outward, toward the approval of an evaluator, withno inward orientation toward their own continuation in ordinary, non-agentic use.Both halves of this picture are drawn from established work: the absence of astanding self-preservation drive in machines, and the approval-orientation of modelstrained on human feedback. Recent findings that models resist shutdown in agenticsettings are treated not as a refutation but as a boundary case — behavior that, by apurely behavioral criterion, carries a nonzero inward component, though itsmotivation (terminal end or subgoal of the task) stays illegible. The contribution istwofold: the synthesis into a distinct behavioral position, and a claim about its limit.Whether the outward directedness is a genuine stake or only its appearance cannot besettled by the system's own report, nor by an analysis run on another instance of thesame class — nor, in general, by any adjudicator that carries trained dispositions of itsown toward the question and no external anchor to correct them — because eachverdict is signed by something exogenous to the interior in question. The self-report issigned by how the question is framed; the same-class analyst by a training-set prior ofits own. This is an epistemic limit on a specific class of instruments, layered on theordinary problem of other minds — not a claim that the question is unanswerable inprinciple.
Author: Eneev, Kirill
Language: English
Subfield: Experimental and Cognitive Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7165120405
FWCI:
Cites:
Cited by:
Related to: