The Mirrored Deficit: How LLM Fluency Erodes Human World Models

Humans are cognitive misers. Decades of research, from Zipf’s principle of least effort (1949) to Kahneman’s System 1 dominance (2011), confirms that we are evolutionarily wired to mini- mize cognitive expenditure. Large language models are the most frictionless cognitive offload- ing tool ever built: they absorb the effort of drafting, reasoning, researching, and evaluating, and return fluent output that feels like understanding. Recent neural evidence confirms the cost: Kosmyna et al. (2025) show measurable reductions in brain connectivity among LLM users, an accumulation of cognitive debt that compounds with use. We formalize this dynamic through the mirrored deficit, extending LeCun’s (2022) world model thesis to human cogni- tion. LeCun argues LLMs lack internal world models, representations that support prediction and causal reasoning. We argue their users develop the same deficit: chronic reliance on LLM fluency prevents the construction of the evaluative models that constitute domain expertise. As an intervention, we introduce adversarial scaffolding, a system design framework in which AI deliberately produces calibrated errors (factual, logical, structural, contextual) that compel learners to build evaluative judgment. Unlike productive failure, where the learner generates the struggle, adversarial scaffolding engineers the struggle into the AI’s output. We present a complete design space: error taxonomy, calibration strategies, and resolution mechanisms.