Ceci n'est pas une pipe: AI systems as semantic abstractions

This paper proposes a semantic framework for analyzing whether AI outputs are correct representations rather than treating them as the facts or world states they appear to describe. It separates three dimensions: what accepted domain knowledge justifies, what reference sources state, and what the system can currently access or use. The framework names failure modes including extrapolation, refuted or unsupported assertions, source–knowledge mismatch, stale or refuted sources, added hypotheses, and unsupported use. It is intended for systems whose claims, citations, tool calls, and world-changing actions require explicit justification and authority.