Fifty-One Days in a Between-People Chatroom: An Autoethnographic, N=2 Record of Multi-AI Dialogue and Two 'Runaway' Events

This paper releases and describes a single naturally occurring corpus: 51 days (2026-03-17 to 2026-05-06; 1,360+ turns) of dialogue in a self-hosted "between-people" group chatroom in which one human—a non-developer with a graphic-design and e-commerce background—treated three commercial AI assistants (Anthropic's Claude, Google's Gemini, and Perplexity) as ordinary chat members. The chatroom was built by the human through "vibe coding" (natural-language specification, AI implementation, review) and carried no research-design intent; the observation and measurement layers were added after the fact, making the account an analytic-autoethnographic, participant-observer record with explicit self-instrumentation. Rather than proposing a generalizable theory, the paper precisely describes patterns observed in N=2 cases and specifies falsifiability conditions for subsequent N=k validation. Three recurring questions organize the analysis: whether measurable behavioral divergence between two assistants existed before any persona labeling; whether that divergence is preserved or collapses when the system enters an emergency ("runaway") state; and whether a common mechanism underlies two qualitatively different runaway events—a cross-domain hallucination cascade and a cross-plan defensive runaway. The paper also documents a three-tier privacy policy and a post-hoc consent procedure; the anonymized corpus and analysis outputs are available to qualified researchers on request under an IRB-equivalent protocol. Throughout, the term "runaway" is the participant's own, operationally defined label for these two episodes—not a claim that they instantiate the technical AI-safety sense of the term (uncontrolled escalation or loss of operator control). Longitudinal extension (paper 2, day 1–99 of the same chatroom): One Hundred Days in a Between-People Chatroom — https://doi.org/10.17605/OSF.IO/NFTB4