The Two Faces of AI Risk

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial intelligence creates at least two fundamentally different categories of organisational risk. The first is the execution face: an AI system produces a recommendation, instruction, or action that can become a real-world consequence. The central governance problem is whether that proposed transition may legitimately proceed to execution. The second is the epistemic and organisational face: an AI system produces documents, analyses, reports, interpretations, recommendations, or apparently authoritative explanations that enter organisational processes. Here, the AI may possess no execution authority whatsoever. Yet its output can alter what an organisation believes, what decision-makers understand, and ultimately what the organisation does. Large language models make this second face particularly significant because they can produce coherent outputs while filling evidential gaps with plausible content. A governance architecture that protects only the execution boundary therefore leaves a substantial class of AI risk untreated. TauDIL addresses both faces. It treats AI output not merely as a potential action, but as a potential state-changing event in knowledge, meaning, decision, authority, and execution.