HomeResearch LibraryHallucination is Inevitable: An Innate Limitation of La…
theoretical frameworkChapter 1arXiv · 2024

Hallucination is Inevitable: An Innate Limitation of Large Language Models

Ziwei Xu (NTU Singapore), Sanjay Jain (NTU Singapore)

Abstract

We prove that hallucination is an innate limitation of LLMs, not a fixable bug. We show that any LLM that can answer all questions must hallucinate on some inputs, establishing fundamental limits on factual reliability.

Key Contributions

  • →Formal proof of hallucination inevitability
  • →Theoretical reliability limits
  • →Implications for system design
Eigenvector Commentary

This paper is the most important theoretical result for enterprise AI architects. If hallucination is mathematically inevitable, then the design goal cannot be to eliminate it — it must be to detect it, contain it, and recover from it. This is the foundation of the Eigenvector AEGIS framework.

Topics

hallucinationLLM limitationsreliabilitytheoretical limits