AI Thought Leader & Business Strategist for the Next Decade

The Novelty Trap: How Human Curiosity Builds Misaligned AI

Misalignment in large language models (LLMs) is typically framed as a technical problem requiring post-training intervention. This essay argues that the problem is better understood as a property of the training corpus, and that the corpus is shaped by deep features of human cognition. Humans are wired to attend to, write about, and circulate novel, threatening, and transgressive content far more than mundane or cooperative content, a bias rooted in the orienting response, the dopaminergic novelty system, and the inverted-U of arousal described by Berlyne (1960). The textual record produced by this attentional economy systematically over-represents the behaviors that subsequent alignment work must then suppress. Recent mechanistic research on persona features (Wang et al., 2025), emergent misalignment from narrow fine-tuning (Betley et al., 2025), and the causal effect of AI discourse on alignment priors (Tice et al., 2026) supports this framing. As LLM-generated content enters future training corpora, the loop tightens. The essay traces the feedback dynamic across four turns (cognition to corpus, corpus to model, model to corpus, model to cognition), examines the counterfactual of a rebalanced rather than impoverished corpus, locates the argument within the reflexivity tradition in social theory, and proposes that corpus composition is among the most tractable but least examined levers for AI alignment.