AI Outputs Are Representations, Not Facts, Says New Framework.

The outputs generated by AI systems, whether text, images, or actions, are not direct reflections of reality or established facts. Instead, they are engineered representations. This paper introduces a new semantic framework designed to help evaluate the correctness and justification of these AI-generated representations. The framework distinguishes between what is supported by accepted domain knowledge, what specific reference sources state, and what the AI system currently has access to and uses. This clear differentiation allows for precise definitions of common AI failures. Terms like "extrapolation," "refuted or unsupported assertion," "sources versus knowledge mismatch," "stale or refuted source," "added hypotheses," and "unsupported use" are formally defined. The authors hope this framework provides a valuable vocabulary for specifying, verifying, and auditing AI systems, especially those whose outputs, citations, tool calls, or real-world actions must be explicitly justified by reliable claims and authority, rather than merely appearing fluent or plausible.