When a language model exhibits capabilities that cannot be traced to any single component, or when a multi‑agent system outperforms each agent alone, researchers label the phenomenon as “synergy” or “emergence”. Such claims are usually informal and lack a common theoretical basis, making comparison or dispute difficult. We argue that information decomposition offers a unifying formal framework to operationalize assessments of synergy in AI systems. The framework distinguishes three qualitatively different types of information: unique information available from individual components, redundant information shared across components, and synergistic information that can only be accessed when components are considered together. According to this view, a system’s behavior is emergent only if it carries synergistic information about task‑relevant outcomes. We illustrate the approach on attention heads, autonomous agents, and human‑AI teams, showing how it organizes existing knowledge across scales and substrates while generating cross‑level conjectures that siloed fields would miss.
Review: Information decomposition provides a measurable, falsifiable criterion for AI emergence, paving the way for more systematic, interdisciplinary investigations.