Chapter 03: How Machine Learning Amplifies — and Can Fix — Bias

Authors: Angel Zhang & Charlie Cao

You asked: “Is the model in the middle?” The answer is yes — and there is more than one model. A typical content system contains at minimum a representation model, a ranking model, and a feedback model.

3.1 How Models Amplify Bias

3.1.1 Representation Bias

Training data carries historical biases. What the model learns as “correct” is only “what was most measured historically.”

For example: search for “CEO” images and you predominantly see Western men — not because the model is biased, but because that is what the training data reflected.

3.1.2 Feedback Loop Amplification

User clicks → Model learns → System serves more of the same → More clicks → Stronger bias

This is a positive feedback system. Once started, bias self-accelerates.

3.2 How Models Can Fix Bias

3.2.1 Constrained Optimization

Add fairness constraints to the objective function:
“While maximizing user satisfaction, ensure that no demographic group or viewpoint receives less than X% of total exposure.”

3.2.2 Exploration Mechanisms

Recommendation systems can deliberately surface content with “high uncertainty and low prior exposure” to break bubble boundaries.
This is the real-world application of ε-greedy strategies or Thompson Sampling.

3.3 The Model Is the Middle Layer — The Objective Is the Soul

$$\text{Matrix (state space)} \times \text{Objective Function} = \text{Model Behavior}$$

Change the objective, and the model trains into completely different behavior.
This is why regulation must act at the objective function layer, not only at the model layer.

3.4 Chapter Summary

Machine learning is not a neutral tool. It is an amplifier of human objectives.

Feed it one objective, and it amplifies that outcome at scale.

This is the basic cognitive literacy all of us need — now.