The Invisible Tiers — Summary

James and his roommate Devin both need a letter to their landlord about black mold. Same apartment. Same mold. Same legal rights. Same AI subscription. James drafts through four iterations, cites the relevant statute, requests remediation in fourteen days, notes rent withholding as a legal remedy. The landlord responds within a week; a remediation company arrives on Thursday. Devin types “write a letter about mold in the bathroom,” gets a polite vague request, sends it, gets no response. The mold stays. Neither thinks the AI is the variable. Both are partly right. Neither can see the invisible architecture that sorted them into different tiers of effectiveness.

Six tiers, all operating simultaneously, all invisible:

The affordability tier: free and paid versions share the same interface, so Margaret on the free tier does not see a smaller seat. She gets a less capable mind and has no way to know this because she has never experienced the alternative. The gap is epistemic, and epistemic gaps are the hardest to close.

The modeling tier: AI systems model some populations with high fidelity and others with low fidelity, mapping onto existing hierarchies of visibility. The AI does not announce this. It answers with the same fluency, the same apparent confidence. The quality is lower; the presentation is identical. Uneven approximation depth, invisible to the person being approximated.

The information quality tier: ask about tax optimization for high earners — rich training data, sophisticated accurate responses. Ask about navigating SSI eligibility while working part-time — thin training data, fluent guidance that may be materially wrong. The AI produces both with equal confidence. The people most in need of cognitive assistance to navigate institutional complexity are the people least likely to get accurate cognitive assistance.

The usability tier: the conversational interface rewards a professional-class cognitive disposition — comfortable with open-ended prompting, accustomed to iterative refinement, treating first outputs as drafts. Devin grew up in a household where official documents arrived bearing authority; you responded to them, you did not mark them up. The interface was built for people who iterate professionally. Everyone else must either adopt that culture’s habits or accept worse outcomes.

The fluency tier: knowing what to ask, recognizing when output is wrong, iterating toward what you actually need. You need a baseline of fluency to develop more fluency. The bootstrapping problem is pernicious: those who most need to develop AI fluency are least likely to develop it because the development process requires fluency.

The compounding tier: every successful interaction builds confidence, develops fluency, expands the range of tasks considered AI-suitable. Every mediocre interaction reinforces the opposite pattern. Previous technology advantages were static. AI advantages have a derivative. It is inequality with a growth rate, not just a magnitude.

All six operating at once produce the same visible outcome: Devin concludes that AI is okay but not that useful. A rational conclusion drawn from systematically degraded experience. The leveling is real; the sorting is also real; and the sorting is invisible while the leveling is visible. The most dangerous feature of the new AI inequality is not its magnitude. It is that it looks like a personal skill deficit rather than a systemic design.