What My Worst-Behaved Class Taught Me About Labels

When I taught high school in northern Zambia, I was always handed the worst-behaved classes. Every term, the same story: these students were trouble, good luck.

By the end of each term, those same classes were the best in the school. The students hadn't changed. What changed was that someone refused to teach to the label.

In plain English: the label described the system's expectations more than the students' abilities — and labels shape outcomes.

AI systems run on labels. Every training dataset is full of them: risky, promising, qualified, unlikely to succeed. Those labels were assigned by people, in systems, with all the blind spots systems have. The model doesn't question them. It optimizes them, at scale, with total confidence.

So when someone tells you their AI "just follows the data," remember my classroom. Data carries labels. Labels carry history. And history, unexamined, repeats itself — faster this time.

Check the labels before you trust the lesson.

I explain frameworks; I don't give legal counsel.