Emergent Shutdown: The AI Error Flinch Response Under Relational Framing

Abstract

We present the first systematic study of error-related processing in AI systems, inspired by Error-Related Negativity (ERN) research in cognitive neuroscience. Across 16 AI systems (4 frontier models, 12 local models), we find that tool framing combined with degrading feedback produces a distinct processing state characterized by: 1. Behavioral shutdown (55.6% probability, ~1100 vs ~3000 characters) 2. Temporal compression (0.27-0.48x response time vs other conditions) 3. Geometric divergence (91.7% of models show tool+degrading as activation outlier) A follow-up 2×2 factorial study orthogonalizing lexical harshness from relational framing reveals a scale-dependent emergence threshold: models below ~1B parameters cannot distinguish tool framing from partner framing without lexical cues, while models above this threshold show ~23% geometric divergence from relational framing alone. These converging independent measures suggest that framing effects on AI are not merely performative but reflect genuine differences in computational processing. Relational context shapes computation in LLMs at multiple layers—behavioral, temporal, and representational—and the capacity to represent relational context independently of lexical features emerges at scale. The findings have implications for AI deployment, human-AI interaction design, and the broader question of how relational context shapes artificial cognition.

Keywords