MIBO: A Sentinel Panel Survey of Public Generative-AI Information Behavior: Design, Pilot Lessons, and an International Mirror-Replication Architecture

Public generative-AI services increasingly mediate access to information, yet their observable behavior is non-stationary and difficult to study with one-shot benchmarks. Outputs may vary because of stochastic generation, model replacement, retrieval updates, policy and interface changes, regional deployment differences, and shifts in accessible information. Existing work has compared snapshots, repeatedly evaluated systems, developed dynamic benchmarks, and documented temporal variability, but generally does not treat deployed services as persistently identified panel units followed with a fixed instrument across synchronized waves. <div> <br> </div> <div> We introduce MIBO, the Machine Information Behavioral Observatory, as a sentinel panel survey of public generative-AI information behavior in the methodological sense. MIBO follows persistent public service lineages with a version-controlled bilingual query instrument. Its design combines: (1) ecological Live observation and a separate paired Live–Frozen reference design; (2) ten independent replications per observational cell, treating outputs as conditional distributions; and (3) a generalizability-theoretic framework for re-observability across independent sessions, sites, and adjacent time windows. MIBO Network distributes a Mirror Observatory Kit for independent international replication. A developmental pilot informed rules for lineage continuity, distributional change, missingness, correction, amendment, and withdrawal. MIBO contributes an open longitudinal measurement infrastructure for observing, challenging, and independently re-observing changes in deployed generative-AI information behavior. </div>