Authority, Silence, and Failure Modes in AI-Driven Systems

Authority, Silence, and Failure Modes in AI-Driven Systems examines a class of failure in which autonomous systems, distributed systems, and AI-driven systems remain technically correct and operational while the conditions supporting legitimate execution authority have changed. The paper focuses on failures that occur at authority boundaries, where degraded coordination, silence, stale state, uncertain control relationships, changing trust, or incomplete evidence can invalidate assumptions that previously supported legitimate execution. Building on the Stable Authority Boundary (SAB), the paper distinguishes technical capability, confidence, optimization, access, and control from recognized execution authority. It develops the Boundary Authority Problem as an under-modeled governance failure pattern and examines authority contraction, refusal, authority restoration, authority leakage, degraded operation, and illegitimate persistence as architectural concerns in autonomous and AI-enabled systems. A central issue is whether the evidence supporting execution remains current and trustworthy. Modern autonomous systems increasingly depend on information produced by sensors, external repositories, AI pipelines, distributed services, and other automated processes. This makes machine-readable provenance, dataset provenance, dataset integrity, metadata generation, data validation, automated transformation, and computational reproducibility relevant to authority decisions. A system may consume an open research dataset, sample dataset, derived dataset, or information maintained within a research data archive. It may execute an automated transformation, use a data processing example, or rely on an artifact identified through a dataset DOI or data DOI. The continued availability of that information does not establish that it remains valid, current, or authorized for the intended action. Where provenance becomes uncertain, validation fails, metadata is incomplete, or a transformation can no longer be reconstructed, a trustworthy autonomous system may need to contract authority, refuse execution, or enter a governed silence state. Tamper-evident data, persistent identifiers, machine-readable records, and reproducible processing can help establish whether the evidence supporting execution remains trustworthy. The analysis therefore argues that uncertainty cannot create authority, probability is not permission, and evidence can support reassessment without itself authorizing execution. A trustworthy autonomous architecture must preserve governed paths for refusal and non-action, constrain retries and recovery to the same authority boundary, and require recognized authority before consequential action proceeds. The result is a governance-first approach to AI governance, autonomous-system safety, resilient systems, runtime governance, authority verification, verification records, and evidence-bearing system governance. Related BLOCK VECTOR research Stable Authority Boundary (SAB) v1.0 — Conformance Specificationhttps://doi.org/10.5281/zenodo.21939806 BLOCK VECTOR Research Maphttps://doi.org/10.5281/zenodo.22102620 Interactive BLOCK VECTOR Research Maphttps://blockvectortech.com/research-map.html Author ORCID — David B. Forbeshttps://orcid.org/0009-0003-8833-3853