AI Hallucination as Epistemic Proliferation

Abstract

In generative artificial intelligence, hallucination is predominantly framed as a factual error or retrieval failure. This article conceptualises hallucination as an ethical failure of restraint, that is, the tendency of fluent models to keep producing confident, expert-sounding language after evidential grounding has become weak or unverifiable. Drawing on the Theravāda Buddhist concept of papañca (conceptual proliferation) and the Abhidhamma analysis of the mental factors (cetasikas), the paper develops the Epistemic Proliferation Model. Operating strictly within a non-anthropomorphic paradigm, the framework analyses how language models structurally reproduce the linguistic artefacts of human craving (lobha), conceit (māna), rigid views (diṭṭhi), and delusion (moha) inherited through training data. It distinguishes epistemic proliferation from Frankfurtian bullshit, sycophancy, and overconfidence, and proposes a five-dimensional diagnostic rubric covering evidential drift, unsupported elaboration density, authority inflation, inferential concealment, and abstention failure. The wholesome cognitive factor sati (mindfulness) and the mode of attention known as yoniso-manasikāra (wise attention) are translated into the governance principle of epistemic anchoring. The framework shifts mitigation focus from post-hoc sentence correction to the structural restraint of ungrounded generative expansion, supplying actionable criteria for responsible AI design and oversight.