Disclosure Stops Before Resilience: Public Evidence on AI Dependencies at Global Systemically Important Banks

Abstract

This study examines how far public information allows external observers to move from a bank's visible artificial-intelligence (AI) use to disclosure relevant to dependency resilience. We audit all 29 banks on the Financial Stability Board's 2025 G-SIB list using a frozen twocorpus protocol. The bank, not an individual service path, is the unit of analysis. The eightelement sequence covers AI materiality, concrete use, external-stack visibility, AI-specific risk, linked dependency, common dependency, substitutability, and tested resilience. Under the strict AI-specific rule, 23 banks reach AI-risk recognition, seven reach linked dependency, and three reach common dependency; under the broader enabling-technology rule, 22 reach dependency and six reach common dependency. No bank completes either sequence through tested resilience. Two blinded independent model recodings agree on 87.4% of 348 institution-item decisions (pooled Cohen's kappa = 0.719; Gwet's AC1 = 0.771), and both independently end at zero complete tested sequences despite different intermediate counts. The evidence concerns public assessability within the specified corpus, not internal control quality or a verified service-level dependency path. The study identifies a bank-level linkage gap and proposes separate minimum disclosures for public comparability and confidential supervisory collection.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.