Background
Week 6's ethical scenarios dealt with the consequences of design choices: reward signals that optimized for the wrong thing, decision trees trained on biased historical data, transparency claims that did not withstand scrutiny. Those were foundational issues — the kinds of problems that emerge from the basic mechanics of how AI systems are built.
Week 7's scenarios deal with something harder. The systems covered this week — reinforcement learning, unsupervised clustering, and large language models — are already deployed at scale in contexts that directly affect your students: in the tools they use for schoolwork, in the platforms that shape what they see and believe, and increasingly in the question of what kinds of work will exist for them when they finish school. The ethical questions here are not hypothetical. They are happening now, to real students, in real schools.
The three scenarios below are grounded in Week 7 content. In each case, the question is not just "what went wrong?" but "given what you now know about how these systems work, what would have needed to be different?"
How to Use These Scenarios
Use the same approach established in earlier weeks: read each scenario carefully, write down your initial thoughts using the five guiding questions, and come prepared to discuss in your small group. A full description of the process is available on the SEC scenario discussion guide.
Scenarios
Scenario 1 — The Reward That Wasn't What We Wanted
A large social media company builds a content recommendation system using reinforcement learning. The reward signal is straightforward: maximize the amount of time users spend on the platform. The system learns quickly and performs well by its own measure — average session length increases by 23% within three months of deployment.
Researchers studying the platform begin to notice a pattern: the recommendation system has learned that emotionally provocative content — outrage, fear, and conflict — keeps users engaged longer than neutral content. It has not been told to promote this content. It has simply discovered, through millions of trials, that it is an effective strategy for maximizing the reward it was given. The system is doing exactly what it was designed to do. The problem is that what it was designed to do and what users — or society — actually benefit from are not the same thing.
- The engineers who built this system did not intend to promote harmful content. Does intent matter when evaluating responsibility?
- "Maximize time on platform" is a clear, measurable, computable reward signal. What would a better reward signal look like — and what makes it harder to define?
- This system learned a strategy no human programmed. Who is responsible for the consequences of a strategy that emerged from the learning process itself?
- Your students use recommendation systems every day. What would you want them to understand about how these systems work and why they behave the way they do?
Scenario 2 — The Confident Answer That Was Wrong
A middle school science teacher in a district that has adopted AI-assisted grading and feedback tools asks the system to generate formative feedback for a set of student responses to a question about the water cycle. The system produces specific, detailed, well-structured feedback for each response. The feedback reads fluently and confidently.
The teacher reviews a sample and notices that several responses have received feedback that is subtly but clearly incorrect — the system has identified a misconception the student does not actually hold, or praised an answer for a feature it does not contain. When the teacher queries the system about one of these responses, the system defends its feedback with additional explanation that is also incorrect but equally fluent and confident.
The district's technology coordinator, when informed, says this is a known limitation and that teachers should "review all AI-generated feedback before sharing with students." The teacher's class has thirty-two students. The system was adopted specifically to reduce her workload.
- You studied hallucination in Topic 7c: LLMs generate text by predicting likely continuations, not by retrieving verified facts. Given this, why does the system produce confident incorrect feedback rather than flagging uncertainty?
- The technology coordinator's guidance — "review all AI-generated feedback before sharing" — is technically sound advice. Is it a workable solution? What does it assume about teacher capacity?
- The tool was adopted to reduce workload. The review requirement may increase it. Who bears the cost of this gap between the tool's promise and its actual behavior?
- This scenario involves a specific classroom use case. What general principles for evaluating AI tools in educational contexts does it suggest?
- As a teacher, you will encounter students and colleagues who trust AI-generated content without scrutiny, and others who reject it categorically. How do you model and teach calibrated judgment?
Scenario 3 — The Question of What Work Is For
This scenario is different from the others. It does not describe a specific system that failed or a specific harm that occurred. It describes a set of questions that are not yet settled — and that your students will live through.
AI systems are automating an expanding range of tasks that were previously done by humans: diagnosing medical images, drafting legal documents, writing code, generating marketing copy, answering customer service questions, tutoring students, screening job applicants. Some of this automation creates new jobs. Some of it eliminates jobs with no clear replacement. Economists disagree about the net effect. What they broadly agree on is that the pace of change is faster than at any previous point in the industrial era, and that the jobs most at risk are not only manual labor — they include many white-collar and professional roles that were considered secure a decade ago.
Some technologists and economists have proposed universal basic income (UBI) as a response: a guaranteed income floor that allows people to meet their basic needs regardless of employment status, funded by the productivity gains from automation. Proponents argue this would free people to pursue meaningful work, caregiving, creative pursuits, and education without the coercive pressure of economic necessity. Critics argue that work provides not just income but structure, identity, community, and purpose — and that a society in which large numbers of people are economically supported but occupationally idle is not a society anyone has chosen or tested.
The teaching profession sits in an interesting position in this conversation. Teaching involves judgment, relationships, responsiveness, and care in ways that are difficult to automate — but AI tutoring systems are improving rapidly, and school districts facing budget pressure are already asking whether AI can replace some of what teachers do.
- Which aspects of your work as a teacher do you believe are genuinely difficult to automate? What makes those aspects hard for an AI system to replicate?
- If an AI system could handle the grading, lesson planning, and content delivery aspects of teaching, and you were paid a livable wage to focus entirely on mentorship, relationships, and the unmeasurable work of caring for students — would that be a better or worse version of the profession? Who gets to decide?
- The question "what jobs are AI-proof?" is probably the wrong question. A better question might be: what human capacities remain valuable in a world where AI handles most cognitive tasks? How does your answer shape what you think schools should be developing in students?
- If a guaranteed basic income removed the coercive pressure to work, would most people find meaningful ways to spend their time? What does your answer reveal about what you believe work is for?
- You are preparing students for a labor market that will look substantially different in ten to twenty years, in ways that are genuinely hard to predict. How do you think about that responsibility? What, concretely, does it change about what you teach?
These scenarios are intended as starting points for discussion, not definitive case studies. You don't need to cover all three in depth — two discussed well is better than three skimmed. Bring your reactions — including disagreements — to your small group session.