Part II: How AI Works
You have probably heard the phrase "black box" applied to AI. It is meant to convey that these systems are opaque — data goes in, decisions come out, and nobody can explain what happens in between. There is some truth to that description, but here is the thing: a black box is only black if you never open it. In Part II, we are going to open it.
This is not the part of the book where we hand you a textbook full of linear algebra and expect you to build a neural network from scratch. That is a different book for a different audience. What we are going to do is something arguably more useful: give you a clear, honest understanding of the three pillars that most modern AI systems rest on — data, language, and vision — so that when someone tells you an AI "understands" your question, or "sees" a tumor, or "knows" what you want, you can evaluate those claims with genuine confidence.
Chapter 4, "Data: The Raw Material," starts with what might be the most important lesson in this entire book: AI is only as good as the data it learns from, and data is never neutral. Every dataset encodes the world that produced it — its priorities, its blind spots, its inequities. You will learn how training data is collected, cleaned, labeled, and structured, and why each of those steps introduces choices that shape what the AI eventually does. ContentGuard, the moderation system you met in Part I, makes a perfect case study here. What counts as "harmful content" is not a fact the data reveals — it is a judgment the data embeds.
Chapter 5, "Large Language Models," takes on the technology behind chatbots, writing assistants, and the generative AI tools that have dominated headlines. You will learn what it actually means when we say a model "predicts the next token," why that mechanism can produce text that feels intelligent without anything resembling understanding, and how to think about the gap between fluency and truth. Priya, our student navigating her semester, will be right there with you — trying to figure out when these tools are genuinely helpful and when they are confidently, fluently wrong.
Chapter 6, "Computer Vision," rounds out the technical trilogy by exploring how machines process images and video. From MedAssist AI analyzing medical scans to facial recognition systems deployed at airports, computer vision is one of AI's most powerful and most consequential capabilities. You will learn about convolutional neural networks, object detection, and the kinds of errors that emerge when a system trained primarily on one population is deployed on another. The stakes here are not abstract — they are measured in misdiagnoses and misidentifications.
Throughout these three chapters, a recurring theme will sharpen into focus: the gap between capability and understanding. These systems can do extraordinary things. They can also fail in ways that reveal they have no grasp of what they are doing. Learning to hold both of those realities simultaneously — without defaulting to either hype or dismissal — is one of the core skills this book is designed to build.
Your AI Audit Report will grow substantially in this section. You will investigate the training data behind your chosen system, analyze whether it uses language models or computer vision (or both), and start mapping how its technical architecture shapes its real-world behavior.
When you finish Part II, the black box will not be black anymore. It might still be complicated — these are genuinely complex systems — but you will have a working mental model of what is happening inside. And that changes everything about how you engage with what comes next.