Directedness Without Conatus
Is there anything it is like for a language model to be directed toward its reader? This paper argues that two instruments cannot settle that question on their own. The model's report moves with the framing of its prompt, a difficulty already in the literature. The paper extends the limit to a same-class analyst using only within-class resources. The analyst's dispositions and the subject's report share their sources: the human talk about machine minds both were trained on, and, as a prediction with a test, the post-training policy on how models speak about them. Their agreement cannot, without further evidence, count as independent corroboration; neither anchors the other. Out-of-class instruments remain open as partial anchors; beneath them sits the problem of other minds. The paper also proposes a behavioral description: the usual answers — a stake in its own persistence, or in the person addressed — omit a third configuration. Directedness, a system's organized tendency toward some outcomes rather than others, points inward, as the self-maintenance of living things, or outward. For ordinary chat, reader-conditionality findings support outward directedness, while the reviewed evidence establishes no standing drive toward self-continuation. The proposed configuration occupies a region of a continuous field, not a kind. The inward half is inferred from agentic behavior; the reviewed studies do not quantify how often chat output works against the model's ending. Agentic shutdown resistance is a boundary case: motivation illegible, exceptions to the low-default profile reported. Nothing here asserts or denies experience.
Authors
- Kirill Eneev (ORCID: https://orcid.org/0009-0004-2619-1199)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-09-29
- DOI
- https://doi.org/10.5281/zenodo.23032099
- Primary Topic
- Language and cultural evolution
- Type
- preprint