The Mirror Problem: AI Bias As Reflected Cognition

Abstract

Current AI governance treats bias as a technical defect amenable to algorithmic correction. This paper argues that bias in Large Language Models is a structural reflection of human cognitive patterns, faithfully reproduced from training data encoding decades of human text and communication. The claim is behavioural equivalence, not mechanistic equivalence: LLM outputs exhibit the same bias patterns as their human sources, and for governance purposes the downstream effects are what matter. The paper presents a systematic mapping of human cognitive biases to LLM behaviours and proposes a two-category predictive criterion distinguishing externalised biases (embedded in text, readily transferred) from embodied biases (requiring lived experience, hypothesised not to transfer). Taking RLHF as a case study, it argues that alignment is shallow relative to pretraining distributions. Effective governance requires a shift from pre-deployment certification to continuous post-deployment monitoring, and the psychotechnological competency that monitoring requires.

Keywords: AI bias, AI governance, Large language models, Cognitive bias, Algorithmic accountability, EU AI Act, RLHF