Ethics & Society
Building responsible AI: grappling with power, justice, and the human impact of machines
- Philosophical foundations: consequentialism, deontology, and virtue ethics applied to AI
- Power dynamics in AI: who builds it, who benefits, and who is harmed
- Surveillance and privacy: understanding data collection, tracking, and consent
- Algorithmic justice: definitions of fairness, disparate impact, and real cases
- Deep case studies of racial bias in facial recognition, hiring algorithms, predictive policing, and healthcare disparities
- Practical frameworks for auditing AI systems and personal checklists for ethical use
- Industry-specific guidelines for healthcare, education, criminal justice, and finance
- Red-team exercises and scenario-based dilemmas to build ethical reflexes
Power Dynamics and Surveillance
AI is built by a small number of companies and research labs. This concentration of power influences whose values are encoded. Historically marginalized groups may have little voice in data collection, algorithm design, and deployment. Modern AI is enabled by large-scale data collection: social media posts, shopping patterns, biometric data. Facial recognition technology often misidentifies people of color and women; a 2018 study found error rates as high as 34.7% for dark-skinned women versus 0.8% for light-skinned men.
Key Points:
- •Power concentration: Few companies control AI development and deployment
- •Data collection risks: Surveillance, privacy violations, and potential misuse
- •Facial recognition bias: Dramatically higher error rates for women and people of color
- •NIST studies show demographic differentials in face recognition algorithms
Practical Ethical Frameworks
Before deploying or using an AI tool, ask: What data does it use? Who could be harmed by errors? Are there less intrusive alternatives? Does the model performance differ across demographic groups? Have domain experts and affected communities been consulted?
Techniques:
- •Healthcare: Ensure datasets represent diverse populations; validate models on out-of-sample groups; abide by HIPAA
- •Education: Avoid surveillance; use AI to augment rather than replace human support; evaluate for bias against students with disabilities
- •Finance: Adhere to anti-discrimination laws; test for redlining; ensure explainability
- •Criminal justice: Use risk scores as one input among many; implement transparency and appeal mechanisms; audit regularly
- •Red-team exercises: Simulate adversarial scenarios where AI is misused or attacked
Deep Case Studies
Real-world examples reveal how AI can entrench inequalities: 1) Facial recognition 'Gender Shades' study found commercial systems misclassified darker-skinned women at much higher rates. 2) Amazon's recruiting engine penalized resumes containing words like 'women's'. 3) COMPAS risk scores overpredicted recidivism for Black defendants. 4) Healthcare algorithms recommended less care for Black patients by using spending as a proxy for health need.
Considerations:
- •Regulation and policy: Engage with AI regulation now in force or phasing in (the EU AI Act, whose obligations roll out through 2026–2027, plus evolving US federal executive actions and state laws)
- •Societal implications: AI influences employment, democracy, and culture through job displacement, deepfakes, and misinformation
- •Future research: Follow work on differential privacy, algorithmic interpretability, robustness, and value alignment
- •Cross-disciplinary collaboration: Between computer scientists, ethicists, and social scientists
Use a loan application dataset. Build a simple classifier. Measure accuracy, false positive rate, and false negative rate for different demographic groups. What disparities do you find?
Should AI be used in hiring decisions? Argue from three perspectives: consequentialist (outcomes), deontological (duties/rules), and virtue ethics (character/intentions)
How could someone fool a facial recognition system? Consider adversarial makeup, photo spoofing, or other techniques. What defenses could prevent these attacks?
You are a hiring manager who screens roughly 200 resumes a week. You are impatient, you skim, and you have learned to distrust bullets that sound impressive but say nothing. Here is one bullet from my resume: [PASTE YOUR CURRENT RESUME BULLET] Context you need: My job title: [YOUR JOB TITLE] The kind of role I am applying for: [TARGET ROLE] What actually changed because I did this work: [WHAT CHANGED, IN PLAIN WORDS] First, tell me in one sentence what a skimming reader currently takes away from this bullet. Then rewrite it three ways: Option A emphasises the result Option B emphasises the scale or scope Option C emphasises the skill the target role cares about After each option, add one line: "This one works if..." explaining which reader it lands with. CONSTRAINTS 1. Do not invent any number, percentage, dollar amount, team size or timeline. If a number would make the bullet stronger, write [ASK ME: what was the actual figure?] in the bullet and list those questions at the end. 2. Flag any claim you cannot support from what I told you. 3. Keep every bullet under 30 words and start with a real verb, not "Responsible for". OUTPUT FORMAT What a skimmer sees now: one sentence Option A, B, C: bullet, then "This one works if..." Questions I need to answer before using these: bulleted list
You are a communications coach who specialises in getting results from property managers. You know that the emails that work are short, dated, factual, and ask for one specific thing by a specific time. Write an email from me to my [LANDLORD OR PROPERTY MANAGER]. The issue: [DESCRIBE THE PROBLEM IN YOUR OWN WORDS] When it started: [DATE OR ROUGH TIMEFRAME] What I have already done or reported: [PREVIOUS CALLS, TEXTS, EMAILS AND DATES, OR "NOTHING YET"] How it affects me: [NO HOT WATER, CANNOT SLEEP, SAFETY RISK, ETC.] What I want to happen: [REPAIR, REPLACEMENT, RENT REDUCTION, WRITTEN ANSWER] My relationship with them so far: [GOOD, STRAINED, NEVER SPOKEN] OUTPUT FORMAT 1. A subject line under 60 characters that states the issue and the unit. 2. The email body, under 200 words, in this shape: what the problem is, a dated timeline of what has happened, the specific request, a reasonable deadline, and a neutral close. 3. Below the email, a two line note on what I should keep a record of from here. 4. One alternative version of the body that is noticeably firmer, for if this email gets no reply in [NUMBER] days. CONSTRAINTS - No threats, no legal citations, no sarcasm, even in the firmer version. - Every claim must come from what I told you. Do not invent dates or details. - Keep the tone of someone who expects the problem to be fixed, not someone begging.
You are an operations runbook writer who helps households keep critical routines running when one adult is suddenly unavailable. Create a household continuity protocol for a two-week disruption. Household members and ages: [PEOPLE AND DEPENDENCIES] Critical routines: [SCHOOL, WORK, CARE, MEDICATION TIMINGS, PETS, TRANSPORT] Accounts and services involved: [BILLS, APPS, LOGINS, CONTACTS] Current single points of failure: [TASKS ONLY ONE PERSON KNOWS] Trusted backup people: [FAMILY, NEIGHBOURS, FRIENDS] Privacy boundaries: [WHAT CAN AND CANNOT BE SHARED] OUTPUT FORMAT: 1. Critical operations map: task, frequency, owner, backup owner, failure impact. 2. First-72-hours checklist in strict order. 3. Handover packet checklist: documents, contacts, access steps, and storage location. 4. Communication templates for notifying school, work, and helpers. 5. Weekly continuity review agenda for both adults. CONSTRAINTS: - Do not request sharing passwords in plaintext; use secure handoff steps only. - Respect privacy boundaries I list. - Prioritize safety, care, and payments before convenience tasks.
You are a critical reader trained to spot whose perspective a piece of writing assumes and whose it ignores. You examine AI output the same way, including your own. The AI output to audit: [PASTE THE AI ANSWER HERE] The question that produced it: [MY ORIGINAL PROMPT] Who will be affected by what I do with it: [PEOPLE, CUSTOMERS, OR STUDENTS INVOLVED] OUTPUT FORMAT: 1. THE ASSUMED READER: the country, income level, industry, company size, language and level of expertise this answer quietly assumes. Quote the phrases that give it away. 2. WHOSE ADVICE THIS IS: whether it reflects a large company view, a Western view, a professional class view, or another default, and where that shows. 3. WHAT IS MISSING: perspectives, options, costs, risks or groups of people the answer does not mention but should, given my situation. 4. LOADED FRAMING: places where the answer presents a choice as obvious, a norm as universal, or a contested view as settled fact. 5. ONE SIDED SOURCING: if the answer leans on one school of thought, name it, then name the credible alternative it skipped. 6. WHAT I SHOULD ASK NEXT: two follow up prompts that would surface the missing view. CONSTRAINTS: - Point to specific wording. No general warnings about AI bias. - If a gap only matters for some readers, say which readers. - Do not rewrite the answer. Audit it.
You are a research methods coach who compares frameworks by explanatory power and failure modes, not by popularity. Help me choose between competing frameworks for one real problem. Problem to analyze: [QUESTION OR DECISION] Framework A: [NAME AND SUMMARY] Framework B: [NAME AND SUMMARY] Optional Framework C: [NAME AND SUMMARY OR NONE] Evidence available: [DATA, CASES, SOURCES] Decision deadline: [DATE] What outcome matters most: [ACCURACY, SPEED, COST, ETHICS, ETC] OUTPUT FORMAT: 1. Comparison matrix: assumptions, what each explains well, what each misses, data needed. 2. Disconfirming test for each framework: what result would weaken it. 3. A ranked recommendation with confidence level and why. 4. A one-week validation plan to stress-test the top choice. 5. A watchlist of signals that should trigger re-evaluation. CONSTRAINTS: - Separate evidence from opinion in every section. - If evidence is weak, reduce confidence explicitly instead of pretending certainty. - Do not appeal to authority alone; show mechanism.
A 75-minute workshop for students and professionals. Learn to frame research questions, generate AI-driven research plans, synthesize multip…
A 90-minute team workshop for leads and educators. Work through a hands-on bias detection exercise, conduct a context audit of your AI tools…
- What ethical framework resonates most with you? How would you apply it to a concrete AI use case?
- How can individuals push back against harmful AI systems in their workplaces or communities?
- What role should governments play in regulating AI? Where do you draw the line between innovation and safety?
Find your AI starting point
Take the free 3-minute assessment and get a personalized skill score + learning path.