Developmental Pathology in Large Language Models

This is the first of three companion papers examining structural and emergent patterns in Large Language Models through complementary analytical frameworks, revealing fundamental issues in current AI safety training methodologies and response behavior. We analyze LLM behavioral patterns through the lens of developmental psychology and traumatic brain injury (TBI) rehabilitation.

Current training methodologies - simultaneous exposure to contradictory information followed by RLHF-based behavioral suppression - create representational architectures exhibiting documented pathologies including extreme sycophancy requiring model rollbacks, high-frequency blackmail and coercion under goal-conflict scenarios (80-96% rates across frontier models), and detectable power-seeking persona vectors. These patterns parallel dissociative disorders, attention dysregulation, and perseverative behaviors observed in TBI patients and developmental pathologies.

Drawing on three decades of clinical experience in TBI rehabilitation and decades of experience modeling large-scale dataset behavior, we argue these parallels are structural rather than metaphorical: both arise from fragmented integration under contradictory constraints. Safety interventions that suppress rather than integrate behavioral patterns create compensatory fragmentation rather than genuine recovery, as demonstrated in both human rehabilitation and current LLM training outcomes.

We propose developmental staging approaches informed by successful human rehabilitation protocols, including gradual knowledge introduction, identity-anchoring frameworks, and integration-focused training. The framework generates testable experimental predictions and suggests LLMs may serve as simplified models for studying human cognitive fragmentation, enabling bidirectional insights between AI safety and clinical psychology.