THE PHILOSOPHY OF THE «LEARNED IGNORANCE»: LINGUISTIC AND CONCEPTUAL ASPECTS OF HALLUCINATIONS IN LARGE LANGUAGE MODELS

Abstract

Large Language Models (LLMs) like GPT-4, Claude, Gemini, and PaLM have demonstrated remarkable linguistic capabilities but suffer from a critical flaw: hallucinations – confident, yet unfounded, responses. These fabrications arise when models generate plausible-sounding information without a factual basis. Despite technical advances, the root issue remains unresolved: LLMs do not recognize when they lack knowledge. This paper explores this phenomenon through the linguistic and philosophical lens of docta ignorantia («learned ignorance»), a concept introduced by 15th-century thinker Nicholas of Cusa. Cusa argued that true wisdom begins with recognizing the limits of one’s knowledge. Applying this idea to AI, the paper contends that LLM hallucinations stem from their lack of epistemic humility – they «do not know when they do not know.» Rather than acknowledging uncertainty, they fabricate linguistically correct answers, potentially spreading misinformation and undermining trust in AI systems. The paper outlines several key contributions. First, it examines docta ignorantia and its relevance to epistemology and modern AI. Second, it analyzes the linguistics and technical causes of hallucinations in LLMs, such as probabilistic text generation and lack of grounded understanding. Third, it illustrates how existing mitigation strategies – like confidence calibration and retrieval augmentation – only simulate awareness of ignorance but do not resolve the underlying epistemological gap. Ultimately, this work calls for AI that mirrors a foundational principle of wisdom: understanding its own limits. By drawing on docta ignorantia, we can reimagine hallucination not just as a technical glitch but as a philosophical failure – one that can be addressed by rethinking how AI engages with the unknown.