”TINY” SILENT HALLUCINATIONS IN AGENTIC AI: HIDDEN FAILURE MODES IN AUTONOMOUS SYSTEMS

Mahule Roy, Subhas Roy

Keywords: agentic AI, silent hallucinations, internal reasoning, reliability, introspection, evaluation metrics

TL;DR: Silent hallucinations are internal reasoning failures in agentic AI that evade standard evaluation, requiring new detection methods and architectural approaches.

Abstract: Agentic AI systems extend large language models beyond static content generation to autonomous planning, tool use, memory management, and long-horizon decision-making. In such systems, reliability is a prerequisite rather than a convenience. While hallucinations in foundation models are typically studied as errors in generated outputs, we identify a distinct and underexplored failure mode that arises within agentic architectures: Silent Hallucinations. Silent hallucinations are internally generated false beliefs, assumptions, or intermediate representations that influence an agent’s decisions and actions without being explicitly surfaced to users. Their invisibility renders them difficult to detect, evaluate, or correct, despite their potential to cause compounding downstream errors. This paper formalizes silent hallucinations as a system-level reliability problem, situates them within existing hallucination and uncertainty literature, proposes a taxonomy of their manifestations in agentic workflows, and discusses implications for evaluation, oversight, and human–AI collaboration. Our analysis reveals that silent hallucinations represent a fundamental reliability gap in autonomous AI that requires new evaluation paradigms and architectural approaches.

Submission Number: 18