From Explainability to System Language: Why AI Accountability Requires a Formal Execution Vocabulary

Current approaches to AI accountability rely heavily on explainability. By translating system behavior into human-interpretable narratives, explainability is expected to restore trust, support audits, and enable responsibility attribution. This paper argues that this expectation rests on a categorical mistake. Explainability operates at the level of interpretation, while accountability requires control at the level of permission. The analysis shows that accountability fails not primarily because modern AI systems are opaque, but because they lack a formal system language for responsibility, validity, and commitment. In probabilistic systems, explanations can rationalize outputs after the fact, but they cannot determine whether a decision was permitted to occur in the first place. As a result, responsibility remains descriptive rather than enforceable. The paper introduces the concept of system language as a structural counterpoint to explainability. System language does not explain decisions; it defines system states, decision boundaries, and transitions that govern when language becomes an accountable system action. Concepts such as non-decision, commit events, system state, provenance, reliability tiers, and technical truth are examined as necessary functional elements of any accountable AI architecture, independent of specific implementations. Rather than proposing a technical solution, the contribution identifies a foundational requirement for accountable AI: accountability presupposes a machine-readable execution vocabulary that precedes interpretation. Without such a vocabulary, governance mechanisms remain external overlays, unable to enforce responsibility within the decision process itself. The paper positions system language as a prerequisite for auditability, certification, and liability in high-risk AI systems, and argues that explainability alone—however sophisticated—cannot fulfill this role. This paper is part of a series examining accountability, auditability, and operational viability in probabilistic and agentic AI systems. A German-language version is available on Zenodo with DOI: 10.5281/zenodo.18663875