Chapter 7: AI Decision-Making — Recommendations, Classifications, and Predictions

"The question is not whether machines can decide. They already do. The question is whether we understand what they're deciding, and for whom."

What You'll Learn

Every time you open a streaming service, apply for a loan, or walk through a neighborhood where police patrol routes have been algorithmically optimized, an AI system has made a decision about you — or about something that affects you. These decisions aren't random. They follow patterns, use data, and produce outcomes. But they also carry assumptions, embed trade-offs, and create consequences that ripple far beyond the moment of the decision itself.

In this chapter, you'll learn to recognize the three fundamental modes of AI decision-making — recommendation, classification, and prediction — and develop a framework for evaluating them. By the end, you'll be able to look at any AI system and ask: What kind of decision is it making? How does it make it? Who does it affect? And what happens when it gets it wrong?

7.1 The Three Modes: Recommendation, Classification, Prediction

Imagine you're at a restaurant. Three things happen:

1. The server suggests you try the daily special. That's a recommendation — pointing you toward something you might want, based on what the restaurant knows about popular dishes, your previous orders, or what goes well together.

2. The kitchen sorts incoming orders into categories: appetizers, entrees, desserts, drinks. Each dish gets labeled and routed accordingly. That's a classification — assigning an item to a category so the system can act on it.

3. The manager looks at reservation data, weather forecasts, and local events to estimate how many customers will come in tonight. That's a prediction — using available information to estimate a future outcome.

These three modes — recommendation, classification, and prediction — are the building blocks of almost every AI decision system you'll encounter. They overlap (a recommendation system classifies your preferences; a prediction system might recommend actions), but distinguishing them helps you understand what an AI system is actually doing and where things can go wrong.

💡 Intuition: Think of recommendation as "What would you like?", classification as "What is this?", and prediction as "What will happen?" Each question demands different data, different methods, and different standards for success.

Why the Distinction Matters

Here's why this isn't just academic taxonomy. Each mode carries different risks:

• Recommendations shape desire. When Netflix recommends a show, it's not just predicting what you'll like — it's shaping what you'll watch, which shapes what gets produced. The stakes seem low (it's just entertainment), but recommendation systems collectively steer culture.

• Classifications impose categories. When an AI system classifies an email as spam, a medical image as benign, or a person as a credit risk, it forces complex reality into discrete boxes. The boxes may not fit. And the consequences of misclassification can range from inconvenient to life-altering.

• Predictions claim to see the future. When an AI system predicts that a student will drop out, a neighborhood will experience crime, or a patient will develop a disease, it's making a probability estimate that can become a self-fulfilling prophecy or a self-preventing intervention — depending on how it's used.

🔄 Check Your Understanding: Your bank uses an AI system to decide whether to approve your credit card application. Is this recommendation, classification, or prediction? (Think carefully — the answer might involve more than one mode.)

The credit card example is instructive because it's actually all three. The system classifies you (creditworthy or not), predicts your behavior (will you repay?), and recommends an action (approve, deny, or offer a different product). Most real-world AI systems combine these modes, which is one reason they're hard to evaluate.

7.2 How Recommendation Systems Shape What You See

You've experienced recommendation systems thousands of times. Every time YouTube auto-plays a video, Spotify builds a playlist, Amazon suggests a product, or TikTok serves you a clip that keeps you scrolling at 2 a.m. when you told yourself you'd go to bed an hour ago — that's a recommendation system at work.

The Basic Mechanics

Recommendation systems generally work through two approaches, often combined:

Collaborative filtering operates on a simple principle: people who agreed in the past will agree in the future. If you and another user both rated The Shawshank Redemption five stars, and that user also loved The Green Mile, the system guesses you might too. It doesn't need to know anything about the movies themselves — just that people with similar taste histories tend to converge.

Content-based filtering looks at the items themselves. If you watched three documentaries about marine biology, the system identifies features of those documentaries (genre: documentary, topic: ocean life, duration: 45-60 minutes) and finds other items with similar features. It doesn't need to know what anyone else thinks — just what you've liked and what the items contain.

💡 Intuition: Collaborative filtering is like asking your friends for recommendations. Content-based filtering is like reading the back cover of a book. Most modern systems do both.

In practice, platforms like Netflix, Spotify, and TikTok use hybrid systems that combine collaborative and content-based approaches, often using deep learning to find patterns that neither approach alone would catch. TikTok's recommendation algorithm, for instance, famously analyzes not just what you watch, but how long you watch it, whether you rewatch, when you scroll past, and even whether you pause to read a caption.

ContentGuard: Recommendation Meets Moderation

Let's return to ContentGuard, the content moderation system we first met in Chapter 1. ContentGuard doesn't just decide what to remove — it's deeply intertwined with what gets recommended. Social media platforms use recommendation systems to decide which posts appear in your feed, and those systems have to interact with content moderation.

Here's the tension: ContentGuard's moderation AI might flag a news article about terrorism as potentially harmful content. Meanwhile, the platform's recommendation system might be boosting that same article because it's getting high engagement. The recommendation system and the moderation system can work at cross purposes.

This creates a structural problem. Recommendation systems are typically optimized for engagement — clicks, watch time, shares. But high-engagement content is often sensational, outrageous, or emotionally provocative. A system optimized for engagement will naturally amplify content that pushes emotional buttons, even if that content is misleading, polarizing, or harmful.

📊 Real-World Application: In 2021, internal Facebook documents revealed by whistleblower Frances Haugen showed that Facebook's own researchers had found the platform's recommendation algorithm promoted divisive content because it generated more engagement. The company was aware that its recommendation system was amplifying content that its own moderation systems struggled to contain. This wasn't a bug — it was a structural conflict between two optimization goals.

The Filter Bubble Problem

When recommendation systems work well, they show you things you genuinely enjoy. When they work too well, they create what researcher Eli Pariser called a filter bubble — a personalized information universe that reinforces your existing preferences and shields you from diverse perspectives.

The filter bubble isn't a conspiracy. It's an emergent property of optimization. A recommendation system that successfully maximizes your engagement will naturally learn that you engage more with content that confirms your existing beliefs. It doesn't "want" to isolate you — it's just following the gradient.

But the consequences are real. If your news recommendations only show you one political perspective, your music recommendations only play one genre, and your social media feed only surfaces opinions you already hold, you're living in a world the algorithm built for you. And that world may be narrower than you realize.

⚠️ Common Pitfall: It's tempting to think you're immune to filter bubbles because you're aware of them. Research suggests that awareness alone doesn't break the bubble — the algorithmic selection still shapes what's available to you, even if you try to seek out diverse content. The asymmetry matters: the algorithm works 24/7, and your conscious effort to diversify is intermittent.

🔄 Check Your Understanding: If a recommendation system is optimized purely for engagement, why might it show different users different content about the same news event? What are the implications for shared public understanding?

7.3 Classification: Sorting People and Things into Boxes

Classification is perhaps the most consequential mode of AI decision-making, because it's the one that most directly puts things — and people — into categories, and those categories have consequences.

How Classification Works

At its core, a classification system takes an input and assigns it to one of several predefined categories. You encountered this concept in Chapter 3 when we discussed supervised learning: the system learns from labeled examples to identify boundaries between categories.

Consider some everyday classifications:

• Email spam filters classify messages as spam or not-spam
• Medical imaging AI classifies tumors as malignant or benign
• Content moderation classifies posts as acceptable or policy-violating
• Resume screening classifies applicants as "move forward" or "reject"
• Fraud detection classifies transactions as legitimate or suspicious

Notice something about this list? As you move down it, the stakes get higher. Misclassifying an email as spam is annoying. Misclassifying a tumor is dangerous. Misclassifying a job applicant is unjust. The same basic technique — assigning inputs to categories — has wildly different consequences depending on context.

The False Positive / False Negative Trade-Off

Every classification system makes two types of errors:

• A false positive occurs when the system says "yes" but the answer is "no." Your non-spam email ends up in the spam folder. A healthy person is flagged as having a disease. A legitimate transaction is blocked as fraud.

• A false negative occurs when the system says "no" but the answer is "yes." A spam email reaches your inbox. A malignant tumor is called benign. A fraudulent transaction goes through.

Here's the critical insight: you can't minimize both at the same time. Making a system more sensitive (catching more true positives) inevitably means more false positives. Making it more specific (reducing false alarms) inevitably means more false negatives. Every classification system sits somewhere on this trade-off, and where you set the threshold is a values decision, not a technical one.

💡 Intuition: Imagine a smoke detector. You can make it incredibly sensitive — it will catch every real fire, but it will also go off when you make toast. Or you can make it less sensitive — fewer false alarms, but a higher chance of missing a real fire. There's no "correct" setting. It depends on how much you fear each type of error. AI classification systems face the same dilemma, just with more complex inputs.

ContentGuard's Classification Dilemma

Return to ContentGuard. Its classification system must sort billions of posts into categories: acceptable, graphic violence, hate speech, misinformation, harassment, spam, and more. Each category has subcategories. Each subcategory has edge cases. And the volume is staggering — Facebook alone processes over 350,000 posts per minute.

ContentGuard faces the false positive / false negative trade-off in a politically charged way:

• False positives (removing content that should stay) suppress legitimate speech. When ContentGuard incorrectly classifies a news report about violence as "graphic violence," it's censoring journalism. When it flags a political critique as "hate speech," it's silencing dissent. Users who experience false positives feel oppressed.

• False negatives (leaving up content that should be removed) allow harm. When ContentGuard fails to catch harassment, targets suffer. When it misses misinformation, it spreads. Users who experience false negatives feel unprotected.

ContentGuard can't satisfy both groups simultaneously. And the problem is worse than it sounds, because the system must work across languages, cultures, and contexts. Sarcasm in one culture is a sincere threat in another. A political symbol that's mainstream in one country is banned in another. The categories themselves are culturally constructed, but the algorithm needs sharp boundaries.

🌍 Global Perspective: Content moderation systems trained primarily on English-language data perform significantly worse in other languages. In 2021, researchers found that Facebook's hate speech detection caught 90% of flagged content in English but less than 50% in Arabic and less than 30% in several African languages. The classification system's errors are not distributed equally — they fall hardest on users from underrepresented language communities.

When Categories Don't Fit

There's a deeper problem with classification that goes beyond error rates: sometimes the categories themselves are wrong.

Consider a hiring AI that classifies applicants as "strong candidate" or "weak candidate." This classification assumes that there is a clear, objective boundary between strong and weak candidates. But in reality, candidates are strong in different ways, for different roles, in different organizational contexts. The binary classification erases nuance that a human reviewer might have preserved.

Or consider a medical AI that classifies skin lesions as "benign" or "potentially malignant." What about lesions that are genuinely ambiguous — where experienced dermatologists would disagree? The classification system forces a binary answer where the truth might be "we need to watch this one."

🔍 Why Does This Work?: Classification systems require discrete categories, but the real world is continuous. This mismatch is not a bug to be fixed — it's a fundamental feature of classification. Understanding this helps you ask better questions: "Are these the right categories?" is as important as "Is the classification accurate?"

7.4 Prediction: From Weather to Recidivism

Prediction is the mode of AI decision-making that most resembles what people imagine when they think about AI: a system that looks at data about the past and tells you what will happen in the future. But the word "prediction" carries a weight of authority it hasn't earned. What AI prediction systems actually produce are probability estimates — educated guesses, with varying degrees of confidence, based on patterns in historical data.

What Prediction Really Means

When a weather app says there's an 80% chance of rain tomorrow, most of us understand what that means: it will probably rain, but it might not. We don't treat the forecast as a guarantee. We bring an umbrella but don't cancel our plans.

But when an AI system says a defendant has an 80% chance of reoffending, or a student has an 80% chance of dropping out, or a patient has an 80% chance of developing diabetes, people tend to treat these predictions differently. They feel more certain, more authoritative, more final. After all, the system analyzed thousands of data points and used sophisticated algorithms. Surely it knows.

It doesn't know. It estimates. And there's a profound difference.

🚪 Threshold Concept: AI decisions are probability estimates, not truths. This is one of the most important ideas in this entire book. When an AI system "decides" something, it's actually calculating the probability that an input belongs to a category or that an outcome will occur. The number 0.83 doesn't mean "this will happen." It means "based on patterns in the training data, inputs like this one were associated with this outcome 83% of the time." The training data might be incomplete. The patterns might not hold. The future might differ from the past. Probability estimates are useful — but they're not certainty, and treating them as certainty causes real harm.

CityScope Predict: The Promise and Peril of Predictive Policing

CityScope Predict, the predictive policing system we've been following since Chapter 1, illustrates prediction at its most consequential.

CityScope Predict works like this: it ingests historical crime data — where crimes were reported, when, what type — along with contextual data like weather, time of day, day of week, and local events. Using these patterns, it generates a "heat map" predicting where crime is most likely to occur in the next 24-48 hours. Police departments then allocate patrol resources accordingly.

On the surface, this sounds reasonable. If the data shows that car break-ins spike on Friday nights in the parking district near the stadium, it makes sense to send more patrols there on Friday nights.

But look deeper, and serious problems emerge:

The data problem. CityScope Predict's predictions are based on reported crime, not actual crime. These are very different things. In neighborhoods with high police presence, more crimes are reported (and discovered). In neighborhoods with low police presence, fewer crimes are reported. In communities where residents distrust police, crimes go unreported. The data doesn't measure where crime is — it measures where crime has been observed and recorded.

The proxy problem. CityScope Predict doesn't use race as an input variable — that would be illegal. But it uses zip codes, property values, and historical arrest rates, which are strongly correlated with race due to decades of residential segregation, discriminatory policing practices, and systemic inequality. These variables act as proxy variables — technically neutral data points that carry the signal of the protected characteristic they're standing in for.

💡 Intuition: A proxy variable is like a shadow. You're not looking directly at the thing you care about, but the shadow's shape tells you a lot about it. Using zip code as a proxy for race is like insisting you're only looking at shadows while knowing exactly whose shadow it is.

The prediction-as-action problem. When CityScope Predict identifies a "high-risk" neighborhood, police send more patrols there. More patrols lead to more arrests. More arrests feed back into the system as data confirming that the neighborhood is high-risk. The prediction doesn't just forecast the future — it creates the future it predicted. We'll explore this feedback loop dynamic in detail in Section 7.6.

📊 Real-World Application: In 2020, the city of Los Angeles suspended its use of PredPol, a predictive policing tool, after a report by the LAPD Inspector General found the system disproportionately directed police to historically over-policed communities of color. Santa Cruz, California had already banned predictive policing outright in 2020, becoming the first U.S. city to do so. These decisions reflected a growing recognition that prediction systems can encode and amplify the biases in their training data.

Prediction vs. Explanation

There's a crucial distinction that gets lost in conversations about AI prediction systems: predicting that something will happen is not the same as explaining why it happens.

A prediction system might accurately identify that students from a certain zip code are more likely to drop out of college. But it can't tell you why. Is it because the schools in that zip code are underfunded? Because families face economic hardship? Because the college isn't providing adequate support? The system identifies a correlation — a statistical pattern — not a causal mechanism.

This matters because prediction and explanation suggest different interventions. If you only have prediction, you might target interventions at students from that zip code (flagging them for extra advising, for instance). If you have explanation, you might address the underlying causes (improving local schools, providing financial aid, reforming institutional practices).

AI systems are generally much better at prediction than explanation. They can tell you what will happen with reasonable accuracy, but they usually can't tell you why — and the "why" is what you need to actually solve problems.

🔄 Check Your Understanding: A hospital uses an AI system to predict which patients are most likely to be readmitted within 30 days. The system has high accuracy. A hospital administrator proposes using it to deny insurance coverage to "high-risk" patients. What's wrong with this proposal? (Consider the prediction vs. explanation distinction.)

7.5 The Accuracy-Interpretability Trade-Off

Now that we've explored the three modes of AI decision-making, we need to confront a tension that runs through all of them: the trade-off between how accurate a system is and how interpretable it is.

What Do We Mean by Interpretability?

Interpretability (sometimes called explainability) refers to how well a human can understand why a system made a particular decision. Some AI systems are inherently interpretable — you can look inside and see the reasoning. Others are "black boxes" — they produce accurate results, but the internal process is opaque.

Consider two approaches to predicting loan defaults:

Approach 1: A simple decision tree. If the applicant's debt-to-income ratio is above 0.4 AND their credit history is less than 3 years AND they've missed a payment in the last 12 months, classify as high risk. This is easy to understand. You can explain to the applicant exactly why they were denied. You can audit the system for bias. You can check whether the rules make sense.

Approach 2: A deep neural network that processes 200 input features through 15 layers of interconnected nodes, each with learned weights. It's more accurate than the decision tree — it catches subtle patterns that simple rules miss. But when it denies an applicant, no one can point to a specific reason. The system's "reasoning" is distributed across millions of parameters in ways that resist human interpretation.

The Trade-Off

In general (with important exceptions), simpler models are more interpretable but less accurate, and complex models are more accurate but less interpretable. This is the accuracy-interpretability trade-off.

⚠️ Common Pitfall: It's tempting to always choose the most accurate model. But accuracy isn't the only thing that matters. In contexts where decisions must be explained (legal, medical, financial), interpretability may be more important than marginal accuracy gains. A 95%-accurate system that can explain its decisions may be more appropriate than a 97%-accurate system that can't.

When Does Interpretability Matter?

The importance of interpretability depends on context:

High stakes, individual impact: When an AI system makes decisions about individual people — loan approvals, medical diagnoses, criminal sentencing — interpretability is often essential. The person affected has a right to know why. Regulators need to audit the system. And if the system makes an error, you need to understand the reasoning to fix it.

Low stakes, aggregate impact: When an AI system makes decisions about things rather than people — optimizing supply chains, recommending movies, predicting weather — interpretability is less critical. If Netflix recommends a show you don't like, you just scroll past. No one demands an explanation.

Regulatory requirements: Some contexts legally require interpretability. The European Union's General Data Protection Regulation (GDPR) includes a "right to explanation" — individuals have the right to understand how automated decisions that significantly affect them are made. If your system can't explain itself, it may not be legally deployable.

🔗 Connection: This trade-off connects directly to the concepts from Chapter 5 about large language models. LLMs are extraordinarily capable but notoriously difficult to interpret. When an LLM generates a medical recommendation or a legal argument, it can't explain why it produced that specific output. This is one reason why human oversight remains essential — a theme we'll continue exploring in Chapter 8.

The Explainable AI Movement

Researchers have developed techniques to make complex models more interpretable without sacrificing too much accuracy. This field, known as Explainable AI (XAI), includes approaches like:

• LIME (Local Interpretable Model-Agnostic Explanations): Creates a simple, interpretable model that approximates the complex model's behavior for a specific decision. It can tell you which factors most influenced a particular prediction.

• SHAP (Shapley Additive Explanations): Borrows from game theory to assign each input feature a contribution score for each prediction. It answers: "How much did each feature push the prediction up or down?"

• Attention visualization: For neural networks that use attention mechanisms (like the transformers discussed in Chapter 5), you can sometimes visualize which parts of the input the model "paid attention to" when making its decision.

These techniques are valuable, but they're approximations. They explain what the model appears to be doing, not what it's actually doing internally. The interpretation is a simplified narrative layered on top of a process that may be far more complex.

🔍 Why Does This Work?: Explainable AI methods don't make black-box models transparent. They create simplified explanations of complex behavior. This is useful — like a doctor explaining a disease in plain language rather than showing you the molecular biology — but it's important to understand that the explanation is a simplification, not a window into the model's actual reasoning process.

7.6 Feedback Loops: When AI Decisions Shape Future Data

We've saved what might be the most important concept in this chapter for last. Feedback loops are what happen when an AI system's decisions influence the data that the system is later trained on or evaluated against. They're subtle, powerful, and extremely common — and they can turn small biases into large ones.

How Feedback Loops Work

Here's the basic mechanism:

1. An AI system makes a decision based on existing data
2. That decision changes the real-world situation
3. The changed situation produces new data
4. The new data is used to update or retrain the system
5. The updated system makes new decisions — which may reinforce the original pattern

This can be a virtuous cycle (the system gets better because its decisions improve outcomes, which improves the data) or a vicious cycle (the system gets worse because its decisions distort outcomes, which distorts the data).

💡 Intuition: Imagine a teacher who decides certain students are "smart" and gives them more attention. Those students perform better (because they got more attention), confirming the teacher's initial belief. Meanwhile, students labeled "not smart" receive less attention, perform worse, and further confirm the label. The teacher's prediction became self-fulfilling. AI feedback loops work the same way, just at scale and at speed.

CityScope Predict's Feedback Loop

Let's trace CityScope Predict's feedback loop in detail, because it's one of the most well-documented examples of how AI feedback loops cause harm:

Step 1: CityScope Predict analyzes historical crime data and identifies Neighborhood A as "high risk" and Neighborhood B as "low risk."

Step 2: The police department sends more patrols to Neighborhood A and fewer to Neighborhood B.

Step 3: More patrols in Neighborhood A means more crimes are observed, reported, and recorded there. Fewer patrols in Neighborhood B means fewer crimes are observed and recorded there — even if the actual crime rate is similar.

Step 4: The new data shows that Neighborhood A has more crime (because more was detected) and Neighborhood B has less crime (because less was detected).

Step 5: CityScope Predict, retrained on this new data, becomes even more confident that Neighborhood A is high-risk and Neighborhood B is low-risk.

Step 6: Even more patrols are sent to Neighborhood A. The cycle intensifies.

The result? A data-driven system that amplifies existing patterns of over-policing and under-policing, all while appearing to make objective, evidence-based decisions.

📊 Real-World Application: Researchers at the Human Rights Data Analysis Group demonstrated this feedback loop mathematically in a landmark 2016 study of the PredPol predictive policing system. They showed that even starting from perfectly unbiased data, the feedback loop between prediction and policing would, over time, produce increasingly biased outcomes. The bias wasn't in the algorithm — it was in the system that connected algorithmic predictions to real-world actions to new data.

Feedback Loops Beyond Policing

Feedback loops aren't unique to policing. They appear everywhere AI decisions influence future data:

• Hiring AI: A system trained on data about successful employees recommends candidates who resemble current employees. If the current workforce lacks diversity, the system perpetuates homogeneity — and each new hire confirmed by the system reinforces the pattern.

• Recommendation systems: If a music recommendation system promotes certain artists, those artists get more streams, which makes them appear more popular, which leads to more recommendations. New artists without an initial boost struggle to enter the cycle.

• Credit scoring: People denied credit can't build credit histories, which makes their credit profiles look riskier, which leads to further denials. People granted credit can build positive histories, which leads to better offers. The gap widens.

• Search engines: Websites that rank high get more clicks. More clicks are interpreted as evidence of relevance. Higher relevance leads to higher rankings. Lower-ranked sites receive fewer clicks, appearing less relevant, and sink further.

Breaking Feedback Loops

Feedback loops aren't inevitable. They can be interrupted by:

• Randomization: Occasionally making decisions that the system wouldn't recommend (showing lower-ranked content, patrolling "low-risk" areas) to gather data that isn't shaped by the system's own decisions.

• External data: Incorporating data sources that aren't influenced by the system's decisions (surveys, independent audits, data from other jurisdictions).

• Temporal analysis: Tracking whether the system's predictions are becoming more extreme over time, which can signal a feedback loop in progress.

• Human review: Having humans periodically override algorithmic decisions and monitoring whether the overrides reveal patterns the system is missing.

🔄 Check Your Understanding: Think about a recommendation system you use daily (social media, music, video). Can you identify a feedback loop? What data does the system generate through its own recommendations, and how might that data reinforce the system's existing patterns?

7.7 Chapter Summary

This chapter explored the three fundamental modes of AI decision-making and the systemic dynamics that make them consequential:

The Three Modes: - Recommendation systems suggest items, content, or actions based on user preferences and behavior. They shape what people see, want, and consume — and they're typically optimized for engagement, not wellbeing. - Classification systems sort inputs into predefined categories. They face inherent trade-offs between false positives and false negatives, and their categories may not fit the complexity of reality. - Prediction systems estimate future outcomes based on historical patterns. They produce probability estimates, not certainties — and they're much better at predicting what will happen than explaining why.

Key Insights: - The accuracy-interpretability trade-off means that the most accurate systems are often the hardest to understand, creating tension in contexts where explanations are needed. - Feedback loops occur when AI decisions influence the data used to train or evaluate the system, potentially amplifying small biases into large ones. - Proxy variables can encode protected characteristics (like race) even when those characteristics aren't directly used. - The distinction between prediction and explanation is crucial: predicting an outcome doesn't tell you why it happens or how to change it.

Recurring Themes in This Chapter: - Tools Built by Humans: AI decision systems embed the assumptions, values, and biases of their designers and their training data. - Who Benefits, Who Is Harmed: The same system can benefit some users while harming others — and the harms often fall on already-marginalized communities. - Human in the Loop: Feedback loops can be interrupted through human oversight, randomization, and external data sources. - AI Literacy as Civic Skill: Understanding how AI decision systems work is essential for evaluating policies, products, and platforms that affect your life.

Spaced Review

These questions revisit concepts from earlier chapters to strengthen your long-term retention:

1. From Chapter 3: We discussed supervised, unsupervised, and reinforcement learning. Which type of learning is most commonly used for classification tasks? Why?

2. From Chapter 4: How does the concept of "training data bias" from Chapter 4 connect to the feedback loop problem described in this chapter? Can you think of a case where biased training data would create a feedback loop?

3. From Chapter 5: In Chapter 5, we explored how LLMs predict the next token. How is this related to the broader concept of "prediction" discussed in this chapter? What are the similarities and differences?

What's Next

In Chapter 8, we'll examine what happens when AI decision systems fail. Not the slow, systemic failures like feedback loops — the acute failures. When AI systems hallucinate, when they encounter data they weren't trained on, when their confidence bears no relationship to their accuracy, and when their errors cascade through interconnected systems. Understanding how AI gets things wrong is as important as understanding how it's supposed to get things right.

📐 Project Checkpoint: For your AI Audit Report, add a section on Decision Types and Processes:

1. Identify the decision mode(s): Does your chosen AI system primarily make recommendations, classifications, predictions, or some combination? Be specific.
2. Trace a decision: Pick one specific decision your system makes and trace it from input to output. What data goes in? What processing occurs? What comes out? What action follows?
3. Map the trade-offs: Where does your system sit on the accuracy-interpretability trade-off? Is it a black box or a transparent system? Does it need to be interpretable for its context?
4. Look for feedback loops: Does the system's output influence its future input data? If so, describe the loop. Is it virtuous or vicious — or could it be either, depending on circumstances?
5. Identify proxy variables: Are there any variables in your system that might serve as proxies for protected characteristics like race, gender, or socioeconomic status?