Human–AI Linguistic Safety and Semantic Integrity Framework: A Working Theory for Continuous Per-User and Across-User Evaluation

Human–AI interaction is predominantly mediated through language, yet shared vocabulary does not guarantee shared meaning. Words acquire significance through conventional definitions, culture, lived experience, relationships, behaviour, embodiment, authority, context, and time. When AI systems rewrite, summarize, classify, translate, or act upon language, they may change not only wording but also evidentiary strength, certainty, responsibility, and institutional meaning. This white paper introduces the Human–AI Linguistic Safety and Semantic Integrity Framework, a conceptual approach for tracing and testing those changes. It proposes that systems preserve original language, separate observation from inference, retain uncertainty, support correction, and calibrate clarification to consequence. Examples involving mother, predator, fear, and injury demonstrate mechanisms that can be turned into reproducible evaluations. They do not validate the framework across populations. The paper concludes with falsifiable propositions and an invitation for interdisciplinary testing.