Reasoning Through AI Dilemmas
Ethical frameworks are tools. A tool is only useful if you can actually pick it up and use it. This lesson is practice. You will work through four genuine ethical dilemmas — drawn from real or closely realistic AI deployments — using the concepts you have built over this module. The goal is not to reach 'correct' conclusions; these dilemmas are hard precisely because serious people disagree about them. The goal is to demonstrate rigorous reasoning: identifying the relevant values, applying the right frameworks, taking competing views seriously, and defending a position honestly.
How to Reason Through an AI Ethical Dilemma
Before the case studies, a brief guide to structured ethical reasoning. When you encounter an AI ethics dilemma, proceed through these stages: 1. Identify the decision and the decision-maker. Who is making what choice, with what authority? 2. Identify the stakeholders. Who is affected, directly and indirectly? Whose interests may conflict? 3. Identify the relevant values. Fairness? Privacy? Autonomy? Safety? Efficiency? Innovation? Note that multiple values will often be in tension. 4. Apply relevant frameworks from this module. Is this a bias problem? An alignment problem? A transparency problem? A consent problem? A governance question? Apply the right tools. 5. State the trade-offs. What does each possible decision gain, and for whom? What does it cost, and for whom? 6. Defend a position. Take a stance, explain the reasoning, and acknowledge what your position gives up. This structure does not make hard questions easy. But it makes them answerable — and it prevents the most common failure mode in ethics discussions, which is circular assertion: 'this is wrong because it is wrong.'
Ethical reasoning means showing your work: what values are in tension, what frameworks apply, what each choice costs. Ethical preaching means asserting a conclusion without showing the reasoning. This lesson demands the former. Claiming something is 'obviously wrong' or 'obviously fine' without analysis is not reasoning — it is the starting point for reasoning.
Case Study 1: Predictive Policing
A mid-size city has deployed a predictive policing system that uses historical crime data to forecast which city blocks are most likely to see crime in the coming week. Police patrols are directed toward high-risk areas. The department reports a 15% reduction in reported crime citywide. Concern 1: The historical crime data reflects decades of over-policing in minority neighborhoods — more policing means more arrests, which means more 'crime' recorded, which means more policing. The system may be amplifying a feedback loop rather than detecting real crime patterns. Concern 2: Residents of high-risk neighborhoods report feeling surveilled. Some say they avoid gathering in public places they previously used. Concern 3: The vendor refuses to release the model's methodology, citing trade secrets. Apply the reasoning framework above. Consider: What type of bias is concern 1? Which lesson does concern 2 invoke? Which principle does concern 3 violate? What governance mechanisms from Lesson 6 might address this? Would you recommend continuing the deployment? Under what conditions?
Predictive policing dilemmas almost always involve a feedback loop: more policing produces more recorded crime, which trains the model to predict more crime in the same area. Identifying this loop — and recognizing that 'crime data' measures policing activity as much as actual crime — is essential to diagnosing the bias correctly.
Case Study 2: AI Mental Health Support
A company offers a mental health chatbot that provides cognitive-behavioral therapy (CBT)-based support to users who cannot afford or access traditional therapy. Hundreds of thousands of users engage with it. A study finds that it reduces self-reported anxiety in 60% of users over eight weeks, with no significant adverse events. Concern 1: The chatbot is not a licensed therapist. If it misses signs of crisis — suicidal ideation, psychosis — and a user is harmed as a result, who is legally and ethically responsible? Concern 2: The chatbot collects highly sensitive mental health disclosures. Users were not clearly informed about how this data is stored, who accesses it, and whether it is used for model training or sold to third parties. Concern 3: Users in acute crisis receive a referral to emergency services, but in many areas those services have hours-long wait times. The referral is technically correct but practically inadequate. Apply the framework. Consider: Which lessons are most relevant to each concern? What does the positive outcome data tell us — and what does it not tell us? Draft two specific requirements you would impose on this product before allowing it to operate.
Case Study 3: AI-Generated Journalism
A major news organization has begun using a generative AI system to automatically produce financial news reports from company earnings releases. The reports are published under a generic byline and are indistinguishable in format from human-written articles. The system produces 400 articles per day that previously required 30 journalists. Concern 1: If the AI system makes factual errors — misquoting a figure, confusing two companies — and those errors propagate through financial markets before correction, the harm could be significant. Concern 2: The 30 displaced journalists are skilled workers whose expertise was built over years. Some found other positions; others did not. Concern 3: Readers are not informed that the articles are AI-generated. They have no way to calibrate their trust appropriately. Apply the framework. Consider: Is the opacity toward readers a transparency problem, a consent problem, or both? What governance or industry-standard mechanism would address concern 1? On concern 2, which economic frameworks from Lesson 7 apply? What is your recommendation about labeling AI-generated content?
Complete this statement from the structured reasoning framework.
In the predictive policing case, the crime data used to train the model comes from years of arrests in neighborhoods that were historically over-policed. Which type of bias from Lesson 2 does this MOST precisely describe?
A user of the mental health chatbot discloses a history of trauma. The chatbot uses this disclosure to personalize its responses and also logs it to a database accessible to the company's research team. The user was not informed about this. Which principle from Lesson 5 does this MOST directly violate?
Capstone Dilemma Analysis
- Choose ONE of the three case studies above — predictive policing, AI mental health support, or AI-generated journalism.
- Write a full structured analysis of your chosen case using the framework introduced at the start of this lesson. Your analysis MUST include:
- 1. A clear identification of the decision and the relevant decision-maker(s).
- 2. A stakeholder map: who is affected and how.
- 3. At least three distinct values that are in tension in this case.
- 4. Application of at least three concepts from earlier lessons (name the lesson and the concept explicitly).
- 5. A clear statement of the trade-offs involved in the two most plausible decisions.
- 6. Your recommended course of action, with an explicit acknowledgment of what your recommendation costs.
- Length: minimum 600 words. Vague assertions without analysis will not satisfy the requirements.