Answers are grounded in this paper and cite their sources.
As AI systems move beyond one-off model invocations to become long-lived actors that invoke tools and produce real-world effects, safety concerns extend beyond the correctness of model outputs. They also raise a question of continuity: can decisions made through human governance remain attached to the same acting subject over time and reliably constrain the effects that are ultimately realized? Credentials, sessions, runtime environments, and authorization states all change. If those changes sever the relationships among subjects, actions, authority, and history, governance may lose its effectiveness even as systems continue to operate normally and generate extensive logs. We study long-lived AI actors, operationalize persistent subjecthood as a cross-lifecycle governance condition, and define subject attribution closure, accountability addressability, governance referential integrity, and responsibility black holes. An observational adequacy lemma and two instantiations characterize the information required for subject-selective online governance and historical attribution. A further counterexample shows that correct governance references do not imply effective enforcement. Building on these distinctions, we formulate six structural conditions: subject and role resolution, preservation of event-time provenance evidence, a scoped authority basis, governance complete mediation, governance-state freshness and revocation semantics, and admission-to-effect binding. These conditions organize a real-time governance chain and a historical accountability chain, supporting a candidate bounded composition result (P4) under an explicit research contract, trust assumptions, and action scope. A public research artifact enables executable examination of these structures through TLA+ formal models, Rust bounded-state exploration, and an independent engineering test harness. No counterexample to the target properties was found for the reference configuration within the declared bounded model. Eight structural mutations expose failure paths involving subject confusion, provenance loss, authority mismatch, governance bypass, execution after denial, stale governance state, effect substitution, and repeated consumption of a single admission. Engineering scenarios further reproduce four accountability verdicts and bounded instances of P4. We also analyze the longer-term implications: enforcement failures prevent established governance decisions from constraining present actions, whereas responsibility black holes may sever the path through which past effects inform future governance when that feedback depends on the affected relationships. If such ruptures persist across large populations of long-lived AI actors and their collaborative systems, the risks extend beyond isolated technical incidents and may erode institutions' capacity to sustain authorization, accountability, credit, and corrective governance over time. The paper thereby recasts a class of long-lived AI governance failures from undifferentiated concerns about model behavior into structural problems involving subjects, execution, and history that can be separately defined, disrupted, and examined.
No takes yet. Share an insight, caveat, or question.
Kun et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: