Machine Learning from Zero
by Sergey Okinchuk
What LLMs Can and Cannot DoFairness and Bias

Chapter Ten · Where Models Meet People

ML in the Real World

Nine chapters judged models by their predictions. This chapter judges them by their consequences. When a model scores a customer, ranks a courier, or learns from your data, the question stops being "is it accurate?" and becomes "is it fair, is it private, can it explain itself, what can't it see — and what does it do to the work we get paid for?" Five topics, one honest tour of where the machinery touches people.

5 topics

Everything so far happened inside the machinery: data in, training, metrics, deployment, generation. But Plateful's models were never really pointed at dashboards — they were pointed at people. The courier who gets the good dinner slot or doesn't. The customer whose behavior became training data without anyone quite asking. The loyal regular declined a payment option with no reason given. Each of those is a person on the receiving end of a file of numbers, and each one raises a question no accuracy score can answer.

This chapter asks those questions in order. How bias gets into a model no one intended to bias — and what it has already done out in the world. What training on people's data means for privacy, from both seats you sit in. Why "the computer said no" is not an acceptable sentence for decisions that matter. The two limits baked into every prediction: correlation is not causation, and deployed predictions reshape the world they predict. And finally the question Iris gets at family dinners — what all this does to jobs — answered evenly, without utopia or doom.

One model, and everyone it touches — the questions of this chapter
Plateful's courier-assignment model
The courier & the customeris it fair to them?can it say why?
The data it learned fromdid people consent?
The regulatoris it accountable?
The companywhat are its limits?what happens to the work?

Topics in This Chapter

Topic 45
Fairness and Bias
Nobody wrote a biased rule — the model distilled yesterday's patterns and shipped them as tomorrow's decisions. How bias gets in, what it has already done, and why "fair" is a choice, not a number.
Concept
Topic 46
Privacy and Your Data
Training data is people. What consent means when records become training sets, when models can leak what they learned, and why a prompt is a transmission — you sit in both seats of this page.
Concept
Topic 47
Why Did the Model Say No?
A loyal customer, declined with no reason given. Explainability: which models can show their reasoning, how black boxes get examined, who needs the explanation — and when opacity is too expensive.
Concept
Topic 48
The Limits of Prediction
Ice-cream orders and pool accidents rise in perfect sync — and dessert is innocent. Correlation versus causation, prediction versus intervention, and the feedback loop where models reshape the world they predict.
Concept
Topic 49
ML and Your Job
"Is AI taking the jobs?" The honest answer is smaller and more useful: ML automates tasks, not titles. What actually changed for four people at Plateful — and the career compass hiding in this book.
Concept