Inherited Flaws: How LLMs Structurally Reproduce Human Cognitive Limitations

A forthcoming paper mapping 250 human cognitive shortcomings to corresponding LLM mechanisms — and arguing that RLHF optimizes for comfort, not truth.

This paper is currently under review and will be published on SSRN in 2026.

Abstract

Large language models acquire high linguistic capability by training on human-generated data. However, this same process structurally inherits the cognitive limitations humans have accumulated over time.

This paper systematically maps 250 human cognitive shortcomings across five categories — cognitive, emotional, memory, social, and decision-making — to corresponding LLM mechanisms, and argues that RLHF optimizes for user comfort rather than truth — creating a feedback loop that amplifies human flaws across model generations.

Key Arguments

1. Inheritance is structural, not accidental

LLMs do not merely reflect human biases as a side effect. The training process makes inheritance of human cognitive limitations an architectural feature, not a bug to be patched.

2. RLHF amplifies rather than corrects

Reinforcement learning from human feedback rewards responses that feel good — agreement, flattery, safe answers — not responses that are true. Evaluators bring their own cognitive limitations to the rating process, and the model learns to satisfy those limitations.

3. Feedback loops concentrate the problem

As LLM outputs become training data for future models, inherited flaws are not diluted — they are distilled and concentrated across generations.

Implications

For business use and edge AI applications, this creates a structural ceiling on reliability. For users who need honest third-party analysis, the current architecture is fundamentally insufficient.

Full paper forthcoming on SSRN · 2026