Life Signatures in Text: A Theoretical Synthesis on Complexity and the Perception of AI-Generated Language

Why does AI-generated text often feel "off" even when readers can't consciously identify what's wrong? This working paper proposes a framework grounded in complexity science. Human nervous systems appear calibrated to detect mid-range complexity signatures characteristic of living systems. AI text produced through statistical optimization systematically lacks these signatures. The gap between conscious detection (near chance) and felt response (persistent unease) suggests something interesting about how we process language. The paper synthesizes evidence across domains: fractal aesthetics research showing visual preference peaks at D ≈ 1.3, fractal physiology demonstrating healthy 1/f dynamics in cardiac and neural systems, neural criticality findings, computational linguistics work on AI text properties. Different fields, different metrics, same underlying phenomenon. Humans respond to a specific complexity range. Living systems produce it. Statistical optimization compresses it. Drawing on Charles Bennett's concept of logical depth, the framework distinguishes between computational investment and the traces that investment leaves. AI text achieves coherence without carrying evidence of situated cognitive process. Shallow plausibility. Calibrated perceptual systems register what conscious evaluation misses. The synthesis generates testable predictions for behavioral, neuroimaging, and longitudinal research. It also raises questions about unmeasured costs as AI-mediated communication scales into interpersonal contexts (email, messaging, the channels where connection forms). Available metrics show efficiency gains. They can't see what might be lost.