Who Gets Approximated — Summary

Not everyone benefits equally from AI that approximates human understanding. Some people are approximated accurately because they match the patterns in training data. Others are systematically misunderstood because they do not.

AI systems trained predominantly on WEIRD populations — Western, Educated, Industrialized, Rich, Democratic — approximate those populations well and fail others in predictable ways. A health AI trained on clinical data that over-samples white patients will miss atypical presentations more common in other populations. A language AI trained on formal English will struggle with code-switching and dialect. A financial AI trained on mainstream banking patterns will flag non-standard behavior as suspicious even when it is simply unfamiliar.

These failures compound. Each misapproximation adds friction to a life that already carries more. And people learn to adapt — describing symptoms in ways the system recognizes, modifying financial behavior to avoid algorithmic suspicion, code-switching to match what AI expects. The burden of translation falls on those already marginalized.

The feedback loop turns sinister here. As systems train on adapted behavior, they learn the adaptations. They get better at understanding the translated version and never improve at understanding the original. The gap widens while the system appears to improve.

The deeper problem is that we struggle to even measure these failures. Average accuracy metrics obscure systematic inequality. If a system is 95 percent accurate for one population and 70 percent for another, the aggregate looks like success.

This is not a technical problem with a technical solution. It is a question of political choice: whose understanding counts, whose experiences constitute valid training data, and who bears the burden when approximation fails.