A century ago, a human read every telegram. Half a century ago, a human routed every phone call. Twenty years ago, a human selected every recommended article. Today, AI is the silent layer underneath your news feed, your credit score, your commute, your apartment application, your doctor's diagnostic support, and the traffic lights on your way home.
This is what infrastructure looks like: technology that used to be remarkable becomes unremarkable, then becomes invisible, then becomes load-bearing. The electrical grid, the highway system, and the internet all went through this. AI is completing it in a decade instead of a century.
The problem with invisible infrastructure is that its failures and biases become invisible too. This course is about making the invisible visible — where AI shows up in everyday life, what it decides, who gets helped and hurt, what it costs in energy and water, and what questions to ask when the next system quietly slides into place.
If you finish every module, here's who you become:
Mapping the systems shaping your day before you notice them
Maria's alarm went off at 6:47 — not 6:45, as she had set it, but at the moment her sleep tracker's algorithm determined she was in a light sleep phase. Her phone's face unlock recognized her in the dark. Her news app had already curated twelve stories it predicted she would read.
She hadn't made any of those decisions. Neither had a human at any of those companies. Three AI systems had shaped her morning before she got out of bed — and she hadn't noticed any of them.
AI is not a future technology. It is the operating infrastructure of daily life — embedded in systems so familiar that most people don't recognize them as AI at all. The first step in understanding AI in society is making this invisible infrastructure visible.
The most common AI systems people encounter fall into four categories: Recommendation systems shape what content, products, and information you see — on streaming platforms, social media, search engines, and e-commerce. Prediction systems anticipate behavior — from predictive text to fraud detection. Recognition systems identify faces, voices, and patterns. Generative systems produce text, images, and audio on demand.
The Invisibility Effect
AI that works well tends to become invisible. A recommendation algorithm surfacing content you enjoy gets experienced as "the app knows me" — not as "an algorithm is deciding what I see." This invisibility is by design. But it makes these systems much harder to question, evaluate, or challenge.
A human editor who occasionally makes a biased selection is a problem for the people they affect directly. An algorithm embedding the same bias, applied to two billion users, is a structural social problem. Scale transforms individual errors into systemic effects.
When a single recommendation algorithm shapes what billions of people see about politics, health, and each other, its design choices carry effects no individual human could produce. This is why AI in society is a different kind of question than AI as a personal tool — the stakes are collective, not just individual.
Think through a typical day from waking to sleeping. For each AI system you encounter, identify what type it is (recommendation / prediction / recognition / generative), what decision it makes, the stakes, and whether you noticed it at the time.
Try starting with: "In a typical day I interact with these AI systems: [list them morning to night]"
Recommendation algorithms, filter bubbles, and the politics of what you see
Two neighbors, same street, same city. One opened their social media feed and saw outrage about a local political issue — inflammatory posts, partisan commentary, rising anger. The other saw recipes, family updates, sports highlights.
They used the same app. They lived in the same neighborhood. But they inhabited completely different information environments — both constructed by the same algorithm, optimizing for the same thing: engagement.
Recommendation algorithms predict what content a user will engage with — clicking, watching, liking, sharing — and serve that content preferentially. The optimization target is engagement, not information quality or accuracy.
This creates a structural dynamic: content triggering strong emotional responses generates more engagement than neutral information. Outrage, fear, and excitement drive clicks more reliably than nuance. Algorithms optimizing for engagement therefore systematically amplify emotionally activating content — not because anyone designed them to promote outrage, but because outrage is engaging.
Engagement vs. Wellbeing
Optimizing for engagement is not the same as optimizing for user wellbeing. Research consistently finds that passive consumption of algorithmically curated content is associated with lower wellbeing than active, intentional content use. Engagement optimization is a business objective — not a design goal for human flourishing.
The filter bubble concept describes the personalized information environment created when algorithms serve content aligned with a user's existing preferences. The filter is invisible — constructed by the algorithm's prediction of what the user will engage with, which tends to reinforce existing interests and beliefs.
Research finds that both algorithmic personalization and user self-selection contribute to filter bubbles — algorithms amplify pre-existing tendencies rather than creating them from nothing. Their relative contribution is actively debated. What is less debated: people in different filter bubbles experience different information realities, making shared public conversation harder.
Search engines operate as information gatekeepers at enormous scale. Their ranking decisions — which sources appear first, which are buried — shape what people find and trust. These are consequential public information choices made by private companies with limited external accountability.
Choose a platform you use regularly — social media, a news app, a streaming service, or a search engine. Spend 10 minutes scrolling or searching normally, then analyze what the algorithm is showing you.
Start with: "I analyzed [platform]. Here's what I observed about what the algorithm surfaced: [your findings]"
When algorithms influence jobs, loans, healthcare, and criminal justice
James applied for a job online. His resume was screened by an algorithm before any human saw it. The system had been trained on the company's past successful hires — predominantly from certain universities, with certain keyword patterns, from certain zip codes.
James didn't go to those universities. He never heard back. He never knew why. He never knew it was an algorithm.
AI is now embedded in consequential decisions across major life domains. Employment: resume screening, candidate ranking, interview analysis, performance monitoring. Credit and lending: loan approval, interest rate setting, credit scoring. Healthcare: diagnostic AI, treatment recommendation, risk stratification. Criminal justice: pretrial risk assessment, recidivism prediction, sentencing guidance.
In each domain, the stakes differ fundamentally from a movie recommendation. Errors, biases, and opacity in high-stakes AI affect economic opportunity, health outcomes, and liberty — not just what someone watches tonight.
COMPAS and the Fairness Problem
The COMPAS recidivism algorithm controversy revealed that different mathematical definitions of "fair" are incompatible — you cannot satisfy all fairness criteria simultaneously. Choosing which definition to apply is a values question, not a technical one. This is not a bug that better engineering can fix. It is a fundamental property of fairness itself.
In high-stakes AI contexts, opacity has profound consequences. When James doesn't hear back after an algorithmic screen, he has no way to understand why — was his resume weak? Was the algorithm biased? Was there an error? Opacity denies affected people the information they need to contest decisions, improve outcomes, or identify discrimination.
A second problem: even when humans are formally "in the loop," automation bias means they systematically over-rely on AI recommendations — even when they have reason to question them. Nominally advisory AI becomes effectively determinative in practice. The human provides accountability theater without genuine oversight.
Choose a real documented case of AI in hiring, lending, healthcare, or criminal justice. Analyze: What decision did the AI make or influence? Who was affected and how? What bias, opacity, or accountability problem was identified? Was it resolved — and how?
Try: "I want to analyze [case name] — [brief description of what happened]"
The political economy of AI deployment
The logistics company replaced its scheduling department with an algorithm. Efficiency improved 23%. The CEO praised the technology at an investor call. The twelve people who had done the scheduling eventually found other work.
The efficiency gains went to shareholders. The transition costs were borne by workers. Who benefits and who pays are almost never the same people — and AI accelerates the gap.
Across AI deployments, a consistent pattern emerges: efficiency gains flow upward — to shareholders and employers — while costs are borne by workers and communities with less power to negotiate or opt out. This reflects the broader political economy of technology in market economies; AI accelerates and scales the dynamic.
AI's most immediate labor market effect is task automation — specific tasks within jobs get automated, changing job content rather than eliminating jobs entirely. Structural displacement (whether AI-created jobs absorb displaced workers) is more uncertain. What is consistent: transition periods are real and painful, and the shift in labor market power trends toward employers.
AI systems require data — and the people whose data trains AI systems are rarely compensated. Users generate behavioral data that trains recommendation algorithms. Patients' records train diagnostic AI. Workers' performance data trains productivity systems. In each case, data is extracted from people with little choice, then used to create value for others.
AI and Power
Ultimately, AI in society is partly a story about power — who has resources to build and deploy AI, who captures its benefits, who can contest its decisions, and who bears its costs. These asymmetries are not technical facts. They are political and economic choices that can be made differently.
Choose a specific AI deployment — Amazon warehouse AI, Uber algorithmic pricing, hospital diagnostic AI, Netflix recommendations, or bank fraud detection. Map who captures the efficiency gains, who bears the costs, who had no choice about being affected, and whether any governance mechanisms mediate the distribution.
Start with: "I want to analyze [company/deployment] — here's who I think benefits and who pays: [your initial analysis]"
15 questions. Complete all to finish the module.