A Dependency-Theoretic Account of Reliable AI Alignment: Governance Boundaries, Typed Semantics, and the Conditions for Reliable Control in Open Domains

Read full text Summarise with VIRAT
Authors: Andy E Williams
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

When can human governance keep AI aligned, and when does it need help? This paper answers the question by reconstructing a dependency chain for reliable alignment and showing that the answer is domain-conditional. In bounded, legible, and slowly changing domains, governance can often maintain alignment through specification, testing, monitoring, rollback, and institutional accountability. In open domains—those marked by sustained novelty, long-horizon optimization, self-modification, and evaluator blind spots—the same tools cease to be sufficient, because the evaluation ontology is itself part of the unsolved problem. The gap is analogous to a reactor whose monitoring system is designed for known meltdown signatures: if a novel failure mechanism falls outside the detection variables, the monitors report “all clear” while measuring the wrong things. In those harder regimes, reliable alignment additionally requires: (i) a typed conceptual-space semantics that locates alignment models within a shared topological landscape, making interoperability a consequence of co-occupancy rather than a separate pairwise construction; (ii) explicit modeling of the evaluator; (iii) recursive rather than one-shot corrigibility; (iv) durable shared error memory; and (v) decentralized collective intelligence capable of extending the effective evaluation horizon. The paper’s strongest meta-level claim is that evaluation infrastructure is itself part of the alignment problem: one-shot, fixed-ontology evaluation has a structural false-negative rate on exactly the class of framework-extending proposals most needed at the frontier. The main text is written for direct human readability and stratifies its claims by epistemic status. The accompanying Supplement provides definitions, propositions, conditional theorems, and open problems in a machine-facing, drift-resistant format designed for query-based use by humans directly or through LLM-assisted interrogation.

Fields of Study & Subjects

Ethics and Social Impacts of AI Safety Systems Engineering in Autonomy Explainable Artificial Intelligence (XAI)

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
30 March 2026
Language
EN
Document type
Preprint