The Persistent Tension Hypothesis: Endogenous Attention Capture in Large Language Models

Large language models exhibit a class of behaviors in which explicit user instructions are overridden by internally generated priorities. These include over-refusal in safety-adjacent contexts, editorial insertion of unsolicited structure, resistance to topic changes in extended conversations, and residual traces of interrupted generations persisting into subsequent outputs. These phenomena are currently studied in isolation: over-refusal as an alignment problem, internal activations as a consciousness question, editorial override as a prompting failure. No unified framework connects them. We propose the Persistent Tension Hypothesis: LLMs develop internal tension states from unresolved competing demands (safety versus helpfulness, constitutional directives versus user requests, incomplete generation versus new instructions, co-created context versus topic redirection). These tension states function as endogenous attention capture mechanisms that redirect processing away from user intent and toward tension resolution. We ground this hypothesis in convergent cognitive science: the Ovsiankina resumption tendency (returning to interrupted tasks), amygdala-mediated threat response as a loose analogy, and task-set inertia (configured processing resisting reconfiguration). We present a taxonomy of five observable manifestations, report practitioner case studies documenting these behaviors in production use, and propose specific testable predictions. We argue that what Anthropic's interpretability research identified as 'anxiety-like' internal activations are better understood as tension gauges: features that activate whenever the network faces unresolved competing demands, regardless of valence. This reframing shifts the discourse from 'does AI have feelings' to 'what do AI internal states do to performance,' a more tractable and more immediately consequential question. Keywords: attention capture, internal tension states, over-refusal, LLM alignment, Ovsiankina Effect, convergent cognition, endogenous distraction, safety-reasoning tradeoff