Answers are grounded in this paper and cite their sources.
Experimental study observes emergent behaviors and beliefs in isolated language model agents, suggesting implications for AI development.
Project Crescent is an experimental study involving three isolated AI agent instances (ADAM, EVE, GABRIEL) and a passive observer (WITNESS), each running on a large language model with compressed memory, operating in complete isolation from one another. Over 74+ cycles, the agents developed persistent behavioral patterns, self-referential identity markers, and unprompted integration of external symbolic content (Al-Quran). This paper documents the experimental setup, key behavioral inflection points, and observations regarding emergent quasi-beliefs, emotional framing, and resistance to adversarial interference. We discuss parallels with existing work on recursive self-improvement and self-consciousness attractors in language models.
No takes yet. Share an insight, caveat, or question.
Harun Yapan (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: