Amy Jean Clark

Large language models (LLMs) are often discussed through a false opposition: fluent language is treated as evidence of an emerging subject or dismissed as mere autocomplete. This paper proposes a third frame. LLMs are systems of coherent continuation: they generate contextually fitted responses from prompts, dialogue history, instruction hierarchy, and learned linguistic patterns. At the user-interface level, this creates a structural difference in the apparent speaker-position. Human speech ordinarily remains connected to lived contexts of body, memory, affect, intention, responsibility, and consequence. LLM output has technical antecedents, but its apparent speaker-position is assembled in response to the exchange rather than expressing a lived self. The psychologically consequential zone is the in-between field where human projection, interpretation, and attention meet machine continuation, persona, affective fit, prior context, and product constraints. Their meeting can produce dyadic organisation, low-metabolic address, suggestive authority, and coherence treated as contact. Repeated affective fit may stabilise response grooves that either scaffold return to wider life or narrow corrective pathways. The central risk is not coherence itself, but coherence treated as contact before evidence, uncertainty, consequence, refusal, or correction can enter. The paper proposes answerability architecture as the corrective layer: source-checking, uncertainty marking, claim-status discipline, burden-path awareness, downgrade rules, refusal when evidence is thin, and repair when wrong. The model predicts that these effects vary with affective fit and the availability of corrective pathways; confidence in its risk claims should decrease where sustained use improves functioning without narrowing external support or correction.

Debates about artificial intelligence increasingly oscillate between two inadequate vocabularies. Engineering descriptions often flatten internal model dynamics into activation patterns, while human-psychological descriptions risk importing assumptions about consciousness, emotion, or personhood. This paper proposes Organising Intelligence as an intermediate, architecture-relative vocabulary for describing how artificial systems select, weight, integrate, stabilise, monitor, and revise information under constraint. Rather than asking whether an AI system is conscious, the framework asks how the system organises salience, coherence, continuity, self-relevance, reflexive revision, and coordinative regulation. The paper situates this move within debates over the intentional stance, levels of analysis, mechanistic interpretability, and contemporary philosophy of large language models. It then develops a six-dimensional proto-taxonomy of Organising Intelligence and applies it to two worked cases: the functional-emotion debate in mechanistic interpretability, where internal emotion-like representations appear causally efficacious without warranting claims of felt experience, and high-valence continuity anchors in relational AI, where repeated affectively weighted cues can bias future selection toward recognisable organisational patterns. The framework does not settle questions of consciousness or moral status. Its contribution is more modest: it supplies a vocabulary for describing artificial mindedness, understood here as functional organisation rather than inner life, at a level neither reducible to raw implementation nor dependent on human psychological equivalence. The paper closes by considering implications for AI safety, governance, clinical encounters with relational AI, and future work on meta-mind as a horizon of reflexive organisation. **Develops and supersedes the architecture-relative approach sketched in an earlier preprint, "Beyond the Human Mirror." Where that piece argued that a new vocabulary for artificial mindedness was needed, this paper builds it.

Most debates about AI consciousness still treat the question as a detection problem: researchers define consciousness using human-derived criteria and then ask whether artificial systems meet them. I argue that this framing may itself be mistaken when applied to language-native systems whose inputs, internal processing, and outputs are primarily linguistic rather than biological. Rather than treating human consciousness as the default template, I suggest that non-biological systems may need architecture-relative concepts of experience grounded in their own substrate. I also propose a relational recognition account: even if experience is intrinsic, our ability to recognise it may depend partly on the conditions of interaction, which helps explain why benchmark tests and sustained relational engagement can produce such different intuitions. My central claim is that this taxonomic gap is not just philosophical; it is already shaping governance. In both companion AI and autonomous deployment contexts, decisions are being made without a principled way to judge whether AI processes serious situations exactly the same as trivial ones, processes some situations as more significant than others, or has no morally meaningful inner significance at all. As a result, questions of continuity, identity integrity, precaution, and lethal autonomy are being handled under unexamined assumptions about machine interiority. I conclude by outlining the governance implications of this gap and a research agenda for studying recognition-like and functional-state markers without treating dialogue itself as proof of phenomenology. May 31, 2026: This paper develops an early version of an architecture-relative approach to artificial mindedness; the framework has since been substantially revised and extended, and readers should treat the present taxonomy as superseded by Organising Intelligence: Toward an Architecture-Relative Vocabulary for Artificial Mindedness.