Ethics in AI: Bias, Fairness, and Responsibility
After this lesson you can identify the sources of AI bias (data, algorithmic, human), distinguish notions of fairness, and apply responsibility principles like transparency, accountability, and explainability to mitigate harm.
{"contentFormat":"slides.v1","completion":{"requireAllSlides":true,"requireQuiz":true},"slides":[{"kind":"title","eyebrow":"Module 1: AI Foundations and Mindset","title":"Ethics in AI: Bias, Fairness, and Responsibility","body":"Explore the ethical dimensions of AI—understanding bias, fairness, and responsibility—to build systems that are trustworthy and equitable. Updated with 2025-2026 developments including the EU AI Act and generative AI challenges.","outcomes":["Define bias in AI and identify its sources (data, algorithm, human).","Compare different notions of fairness (equality of opportunity, equal accuracy, proportionality).","Explain responsibility in AI including transparency, accountability, and explainability.","Analyze real-world cases from 2016-2026 (COMPAS, Amazon, healthcare, deepfakes).","Describe global regulatory frameworks: EU AI Act, US Executive Orders, China's approach.","Apply a step-by-step bias mitigation process to a practical scenario."],"narration":"Welcome to this lesson on ethics in AI. We'll cover bias, fairness, and responsibility, using up-to-date examples and global perspectives. By the end, you'll be equipped to identify ethical pitfalls and contribute to responsible AI development."},{"kind":"content","heading":"Key Concepts: Bias, Fairness, Responsibility","body":"Bias in AI refers to systematic errors that reflect societal prejudices. It can stem from:\n- Data bias: historical inequalities or underrepresentation (e.g., facial recognition trained mostly on light skin).\n- Algorithmic bias: design choices that favor certain groups.\n- Human bias: developers' unconscious assumptions.\n\nFairness aims to prevent discrimination. Common notions:\n- Equality of opportunity: equal chance for qualified individuals.\n- Equal accuracy: consistent performance across groups.\n- Proportionality (equity): adjusting for existing disparities.\n\nResponsibility includes:\n- Transparency: open about how systems work.\n- Accountability: clear lines of responsibility.\n- Explainability (XAI): decisions understandable to humans.","callout":{"variant":"insight","title":"Why This Matters Now","text":"Generative AI (e.g., GPT-4o, DALL-E 3) introduces new ethical challenges: hallucinations, misinformation, and deepfakes. The EU AI Act (passed 2024) classifies AI by risk. US federal AI policy shifted in 2025 and is still evolving — treat it as partly voluntary and verify current status."},"narration":"Let's start with the core concepts. Bias can enter through data, algorithms, or human decisions. Fairness has multiple definitions, and responsibility requires transparency and accountability. Recent regulations like the EU AI Act are shaping how we address these issues."},{"kind":"content","heading":"Real-World Examples: Then and Now","body":"Classic cases remain instructive, but newer incidents highlight evolving challenges.","table":{"headers":["Example","Year","Issue","Outcome"],"rows":[["COMPAS recidivism tool","2016","Racial bias: Black defendants flagged as high-risk more often than white with similar records.","Widely criticized; led to research on fairness metrics."],["Amazon hiring tool","2018","Gender bias: penalized resumes with 'women's' keywords.","Scrapped; highlighted data bias."],["Healthcare algorithm (Optum)","2019","Racial bias: under-referred Black patients to care programs.","Retrained after study; showed bias in commercial algorithms."],["Deepfake scandals (2024-2025)","2024-2025","Misinformation: AI-generated videos of politicians and celebrities.","Prompted US executive orders and EU Digital Services Act enforcement."],["Generative AI hallucinations","2024-2026","False information presented as fact (e.g., legal citations).","Ongoing; companies add disclaimers and retrieval-augmented generation."]]},"narration":"Here are key examples. COMPAS and Amazon are still taught, but newer cases like biased healthcare algorithms and deepfakes show the problem persists. Generative AI adds risks like hallucinations and misinformation."},{"kind":"content","heading":"Global Regulatory Frameworks","body":"Different regions take distinct approaches to AI governance. This picture is as of 2026 and moves quickly — always verify the current status before relying on it.","cards":[{"title":"EU AI Act (2024)","text":"Risk-based classification: unacceptable, high, limited, minimal. High-risk AI (e.g., hiring, credit) must meet transparency, accuracy, and human-oversight rules. Fines up to 7% of global turnover for the most serious violations."},{"title":"United States (evolving)","text":"The 2023 EO 14110 set federal AI safety, testing, and reporting directives, but was rescinded in January 2025. EO 14179 now directs an AI Action Plan. Many US commitments remain voluntary or sector-specific."},{"title":"China's AI Governance (2023-2025)","text":"Focus on social stability and state control. Regulations require algorithmic transparency, ban certain deepfakes, and mandate content moderation. Emphasis on 'core socialist values'."}],"narration":"Globally, regulations are evolving. The EU AI Act is the most comprehensive, the US uses executive orders and industry self-regulation, and China emphasizes state control. Understanding these helps you design compliant systems."},{"kind":"content","heading":"Step-by-Step: Mitigating Bias in Practice","body":"A practical approach to reduce bias in your AI projects.","callout":{"variant":"exercise","title":"Try It: Audit a Dataset","text":"Take a sample dataset (e.g., UCI Adult Income). Check for imbalances: gender, race, age. Use Python libraries like Aequitas or Fairlearn to measure disparities. Document findings and propose mitigation."},"table":{"headers":["Step","Action","Tools/Techniques"],"rows":[["1. Data Audit","Examine training data for representation and historical biases.","Pandas profiling, Aequitas, Fairlearn."],["2. Bias Detection","Measure performance disparities across groups.","Disparate impact ratio, equal opportunity difference."],["3. Data Augmentation","Balance dataset via oversampling, synthetic data.","SMOTE, GANs (for images)."],["4. Fairness Constraints","Incorporate fairness during model training.","Adversarial debiasing, reweighting."],["5. Monitoring","Continuously track model outputs for drift.","MLflow, WhyLogs, custom dashboards."],["6. Transparency","Document model decisions and limitations.","Model cards, Datasheets for Datasets."]]},"narration":"Here's a six-step process to mitigate bias. Start with auditing your data, detect disparities, then use techniques like data augmentation and fairness constraints. Always monitor and document. Try the exercise with a real dataset."},{"kind":"quiz","heading":"Check Your Understanding","questions":[{"question":"Which of the following is an example of data bias?","options":["An algorithm that prioritizes features correlated with race.","A facial recognition system trained mostly on light-skinned faces.","A developer assuming all users have high-speed internet.","A model that uses a fairness constraint."],"questionId":"cmrf73k3g000cpd27olw02fzh"},{"question":"What is the primary goal of the EU AI Act?","options":["To ban all AI applications.","To classify AI by risk and impose requirements on high-risk systems.","To promote AI innovation without regulation.","To create a global AI ethics standard."],"questionId":"cmrf73k3g000dpd27xwcpm8k1"},{"question":"Which fairness notion means that equally qualified individuals from different groups have the same chance of a positive outcome?","options":["Equal accuracy","Proportionality","Equality of opportunity","Demographic parity"],"questionId":"cmrf73k3g000epd27hio3ekez"}],"narration":"Now let's test your knowledge with three questions. Choose the best answer for each.","quizId":"qz_cmk5q7r6x000cg4eo3nltf88q"},{"kind":"summary","heading":"Key Takeaways","takeaways":["Bias in AI can arise from data, algorithms, or human decisions; it must be actively identified and mitigated.","Fairness has multiple definitions; choose the one appropriate for your context and stakeholders.","Responsibility requires transparency, accountability, and explainability throughout the AI lifecycle.","Real-world examples from 2016-2026 show persistent issues; generative AI adds new challenges like hallucinations and deepfakes.","Global regulations (EU AI Act, US Executive Orders, China's laws) are shaping ethical requirements; stay informed.","Use a systematic process: audit, detect, mitigate, monitor, and document."],"narration":"To summarize: bias is pervasive but manageable. Fairness is context-dependent. Responsibility is non-negotiable. Keep up with regulations and always test your systems. Thank you for learning about AI ethics."}]}