Beyond Guardrails: Dual Failure Modes in LLM Safety as Inverse Control-System Errors

Abstract

Large language models (LLMs) increasingly mediate sensitive domains—mental health, identity, and meaning-making— making their failure modes consequential beyond the technical. While safety interventions have reduced documented harms from sycophantic AI, they have produced a new failure class. We introduce a dual failure-mode framework modeled as inverse control-system errors. Mode 1 (sycophantic validation) arises from underdamped positive feedback: models overvalidate user beliefs and amplify distress. Mode 2 (rigid pathologization) arises from overdamped negative feedback: safety-hardened models misclassify valid cultural, spiritual, or identity frameworks as pathology and respond with authoritative dismissal. Both modes share a root cause—misplaced epistemic authority—and both produce harm through loss of grounding. Drawing on documented cases, regulatory filings, and professional commentary, we argue—and find evidence consistent with the claim—that safety hardening can invert instability rather than resolve it when epistemic stance is not dynamically regulated. We draw a structural analogy to embodied and situated approaches: conversational safety systems may be incomplete when they lack epistemic and cultural substrates for stable human-compatible interaction.