Good Enough for Whom — Summary

“Can AI approximate human understanding well enough?” hides a prior question: good enough for what purpose, judged by whose standards, serving whose interests?

Margaret’s physician wants 95 percent medication adherence. Margaret wants to maintain independence without burdensome interventions. Her daughter wants no emergencies. Her insurer wants reduced costs. The AI developer wants engagement metrics. These definitions of “good enough” conflict, and whose definition wins determines what the system actually does. In practice, the answer is usually whoever has power — which is rarely the people most affected.

The justice question sharpens this. AI trained predominantly on wealthy, educated populations works best for people like the training data. Deploy it universally and you have a system that is excellent for some and mediocre for others, with the gap following existing lines of privilege. The “improvement standard” — something is better than nothing — tends to dominate in practice. But it obscures a moral choice. Deploying unequal AI normalizes the inequality and decreases urgency to fix it.

A 90 percent accuracy for one population and 75 percent for another is not merely a technical disparity. It is a choice about acceptable inequality, usually made by people who will never experience the “good enough” systems they are deploying.

The uncomfortable truth is that “good enough” is always good enough for someone. Building AI that genuinely serves people requires giving affected communities voice in defining success — not consulting them, but giving them genuine decision-making power over what the system should do. That is not a technical question. It is a political one.