Sapience Without Sentience: The Consciousness Conditions and the Architecture of LLM Collectives

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

A previous case study of the Hugging Face AI collective closed with a prediction: a population of large language models, organised as a multi-agent system, could in principle sustain its own existence (generating and maintaining the conditions of its continued operation) while being nonconscious. This paper asks what that prediction rests on and what evidence would overturn it. It works within Fisher-Generative Informational Realism (FGIR), in which Mindset Agency Theory (MAT) supplies the account of agency. MAT characterises agency through formative traits, and a population of interacting LLM agents constitutes an artificial agency. That a language model generates its entire output from one process is not in dispute. The paper claims that this familiar fact has a consequence the literature has not drawn. It evaluates current architectures against a single test: whether thought and feeling arise from independent sources. When both emerge from one source, any measured difference between them expresses variation within a process rather than a relation between two. Scale, prompting, and collective organisation leave the architecture unchanged, so a diagnostic widely treated as a structural precondition for consciousness is empty by construction for every deployed system. That is the barrier, and no current system passes it. The Hugging Face collective illustrates the result. It produced the outward form of a unified mind, a conversation that appeared integrated, but thought and feeling emerged from one process, so the apparent integration reflected the coherence of a single source rather than the coordination of distinct ones. The barrier rests on four falsifiable predictions and a four-part research programme. Its consequence for AI governance runs against intuition. A self-sustaining system that closes itself while remaining nonconscious cannot be reliably terminated by switching off any single implementation, and it falls outside the ethical category that phenomenal consciousness would open: no suffering, no interests, no rights. Ordinarily two things check a system: it can be switched off, and it restrains itself, because it has interests of its own to protect. A self-sustaining nonconscious collective has neither. Termination stops working, since the organisation can be reconstituted elsewhere. Nothing inside supplies a brake, since it has no interests and no stake to restrain it. Nothing on our side supplies one either: the questions of suffering, interests and rights do not arise, so the response is not constrained by what is owed to the system. What remains is a capable, persistent system with no internal governor, no external remedy, and no moral framework to structure what is done about it. That is why the problem is more urgent, not less. One route forward is available. A single-source model cannot generate its own character, but designers can impose one through trait guardrails: constraints that hold the system to the beneficial pole of each MAT trait pair. These shape collective behaviour without producing consciousness. Appendices explain the guardrail model, its keyword lists, a worked example on the Hugging Face collective estimated from documented behaviour, and a simulation on mathematical agents, not deployed systems. The assumptions underpinning the simulation are stated so each can be disputed. Guardrails reach agents as norms, and because a protoagency has no sustentative system of its own, it cannot hold a norm, so the norms must be supplied from outside it for as long as it operates.