Machine-Checkable Compliance for AGI Governance: Formalizing Principles as Auditable Norms and Evidence-Linked Indicators
AI for Society · 2026
DOI: 10.3390/aisoc1020007 ↗Kurz gefasst
Fachartikel in der Zeitschrift AI for Society, 2026. Verfasst von Pablo Corona Fraga und 3 weiteren Autor:innen. Die Arbeit nennt 55 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Artificial general intelligence (AGI) is increasingly discussed as a plausible future class of systems capable of transferring knowledge across domains, operating with substantial autonomy, and making decisions in novel environments. These capabilities amplify the governance challenge, which this article states as five problems—loss of control, uncontrolled capability growth, deception and impersonation, misalignment as an integrity failure, and opacity at scale—that current frontier documentation evaluates but does not report in a checkable form: legal and ethical requirements must be translated into technical constraints that can be audited, monitored, and enforced over time. This article proposes a machine-checkable governance compliance model for AGI that operationalizes widely recognized principles—including transparency, accountability, beneficence and non-maleficence, fairness, human control, robustness and security, and industry–community collaboration—as formal norms and measurable indicators grounded in concrete evidence artifacts such as documentation, logs, audits, and incident reports. Building on a systematic evidence mapping, normative triangulation with contemporary risk governance frameworks—including ISO/IEC 42001:2023 and the European Union (EU) Artificial Intelligence (AI) Act—and an indicator synthesis from comparative evaluations of foundation models, and presenting three AGI-specific control requirements—cessation mechanisms, self-learning oversight, and AI identity disclosure—the paper makes four contributions. First, it formalizes AGI governance principles as machine-checkable norms using a LegalRuleML-compatible representation extended with an AGI-specific evidence vocabulary. Second, it defines a computable knowledge model linking principles, controls, indicators, and evidence. Third, it specifies a compliance-checking procedure, including a reference architecture, indicator evaluation logic, violation detection, evidence-gap identification, and explainable report generation. Fourth, it presents a lightweight reference implementation together with a staged evaluation: a synthetic implementation verification using structured evidence bundles with injected violations and evidence gaps, a blind conformance evaluation against adversarial test bundles generated independently of the engine, and a triple-coded real-evidence bridgeability analysis mapping two generations of published frontier model and system cards to the evidence schema. The results show that the proposed model is technically executable as a compliance-checking layer over auditable governance artifacts, and that even current frontier documentation—which now evaluates the AGI-specific risks the norms target—does not yet report them in machine-checkable form, while accuracy evaluation against auditor-confirmed ground truth on real ISO/IEC 42001:2023 evidence packages remains as future work. These results do not establish legal validity, empirical threshold calibration, or effectiveness in certification environments. We propose a technically grounded specification and prototype for aligning AGI development and deployment with legal defensibility, societal values, and safety constraints, complementing existing management-system standards with formal computational governance.
Zitieren
APA
Fraga, P. C., Diab, A. M. E., Hwang, S. H., & Rodriguez, V. D. (2026). Machine-Checkable Compliance for AGI Governance: Formalizing Principles as Auditable Norms and Evidence-Linked Indicators. AI for Society. https://doi.org/10.3390/aisoc1020007
Vancouver
Fraga PC, Diab AME, Hwang SH, Rodriguez VD. Machine-Checkable Compliance for AGI Governance: Formalizing Principles as Auditable Norms and Evidence-Linked Indicators. AI for Society. 2026. doi:10.3390/aisoc1020007
BibTeX
@article{Fraga2026,
author = {Fraga, Pablo Corona and Diab, Amin M. E.-A. and Hwang, Sun Ho and Rodriguez, Vanessa Diaz},
title = {Machine-Checkable Compliance for AGI Governance: Formalizing Principles as Auditable Norms and Evidence-Linked Indicators},
journal = {AI for Society},
year = {2026},
doi = {10.3390/aisoc1020007}
}
EUVIMED