The Mirror Problem: AI Bias as Reflected Cognition

The engineering framing of AI bias — find it, measure it, remove it — quietly assumes bias is a contaminant: something foreign to the system that better data hygiene could filter out. The paper argues the assumption is wrong in an instructive way. A model trained on the written output of human civilisation does not acquire bias by accident; it acquires cognition, and human cognition arrives biased. The system is a mirror. What it reflects is the statistical shape of how we actually think.

The paper organises reflected cognition into two categories. Externalised biases live in the data itself — the stereotypes, omissions and skews that human authors wrote down, which models absorb as fact-shaped patterns. Embodied biases are deeper: structural features of human reasoning, the availability and anchoring heuristics catalogued since Kahneman and Tversky, which models reproduce because the reasoning they learned from embodies them. Sixteen documented bias phenomena are mapped across the two categories, each tied to its human counterpart in the cognitive-science literature.

The distinction carries the governance payload. A contaminant can be certified absent before deployment; a reflection cannot, because the mirror keeps reflecting whatever passes in front of it — new data, new contexts, new interactions. Pre-deployment certification, the instinctive regulatory tool, is structurally mismatched to the problem. The paper argues for continuous monitoring as the organising principle of AI bias governance, and closes on the wider point: every bias found in the mirror is a finding about us, which makes AI bias research an unexpected instrument for civilisational self-knowledge.