Part III: AI in Action
Imagine two scenarios. In the first, an AI system reviews a loan application and recommends denial. The applicant is never told why. In the second, an AI medical tool flags a scan as showing early-stage cancer — except the patient does not have cancer, and the false alarm triggers a cascade of invasive follow-up procedures. In both cases, the technology worked exactly as designed. In both cases, someone was harmed.
Part III is where we move from understanding how AI systems are built to confronting what happens when they are deployed. This is the pivot point of the book — the place where technical literacy meets real-world consequences.
Chapter 7, "AI Decision-Making," examines how AI systems arrive at their outputs and why those outputs carry weight in the world. You will learn the difference between classification and prediction, between recommendation and determination, and between a system that advises and a system that decides. CityScope Predict, the predictive policing system from Part I, returns here in full force. When an algorithm tells police where to patrol, is it making a prediction or issuing an order? The answer depends on how humans in the loop interpret and act on the output — and that, it turns out, is where many of the most consequential failures occur. This chapter will equip you with a framework for mapping any AI system's decision chain from input to outcome, a skill that will prove essential for the rest of the book.
Chapter 8, "When AI Gets It Wrong," is the chapter nobody building AI products wants you to read — but it is one of the most important chapters in this book. Every AI system fails. The question is not whether, but how, how often, and who bears the cost. You will explore the taxonomy of AI failures: from training data gaps to distribution shift, from adversarial attacks to good old-fashioned overconfidence. MedAssist AI and Priya both reappear here, each encountering a different kind of failure. MedAssist produces a false positive that erodes a physician's trust. Priya submits AI-assisted work that contains a fabricated citation — plausible, well-formatted, and entirely invented. These are not edge cases. They are the predictable, recurring failure modes of systems that process patterns without understanding context.
A theme that runs through both chapters, and that will echo through the rest of the book, is the question of who benefits and who is harmed. AI failures do not distribute evenly. They cluster around the people and communities least likely to have been well-represented in training data, least likely to have a voice in system design, and least likely to have recourse when something goes wrong. Recognizing that pattern is not pessimism — it is the beginning of accountability.
Your AI Audit Report takes a critical turn in this section. You will map your chosen system's decision types, trace an input through to its output, and then — perhaps more revealingly — analyze its failure modes. What happens when it gets it wrong? Who notices? Who pays?
Part III is short — just two chapters — but do not mistake brevity for lightness. This is where the abstract becomes concrete, where "interesting technology" becomes "technology with consequences." By the time you turn to Part IV, you will be ready to ask the harder questions about fairness, labor, creativity, and privacy — because you will understand exactly what is at stake.