MIST 5440 Final Exam Review Guide

AI for Good Not Bad: Definitions and Core Concepts

• Defining AI for Not Bad vs. AI for Good     * AI for Not Bad: Defined as the practice of using AI to achieve specific organizational goals while simultaneously mitigating the risks inherent in the technology. It focuses on avoiding ethical pitfalls during the pursuit of objectives.     * AI for Good: Defined as using AI specifically to create a positive social impact as the primary goal.
• Consequences of Ethical Risks: These risks occur at scale and have four primary dimensions: reputational, regulatory, legal, and operational. They are costly in terms of money and resources to address and result in the loss of consumer trust.
• Limitations of Existing Frameworks: Corporate codes of conduct and current regulations are inadequate because they do not specifically account for "AI bad behavior."
• The "Big Three" AI Ethical Challenges: Bias, Lack of Explainability, and Privacy.
• Content vs. Structure in AI Ethics Programs     * Content: Determining exactly what is considered "good" or "bad" within an organization. It identifies the specific risks to avoid and the values the organization upholds.     * Structure: The governance framework, including policies, processes, and formal mechanisms used to operationalize ethical risk mitigation.
• Misperceptions of Ethics as Obstacles     * Subjectivity Myth: The belief that ethics is purely subjective because people disagree, making it impossible to determine what is ethical.     * Science vs. Ethics: The false idea that only science delivers "truth" and since ethics is not a science, it cannot provide truth.     * Authority Dependency: The belief that ethics requires an authority figure to dictate right from wrong, otherwise it remains subjective.     * These misperceptions impede organizational buy-in and stop fruitful discussions.
• The Problem with Consumer Ethical Beliefs: It is unadvisable to base AI ethics programs solely on consumer perceptions because:     * Consumers often have not considered the specific situations where ethics is required.     * Consumer perceptions are often "coarse-grained," while the problems are "fine-grained."

Responsible AI Frameworks

• Definition of Responsible AI: A governance framework that documents how an organization addresses ethical and legal challenges. It involves using AI while proactively mitigating risk and bias, especially as AI scales.
• Essential Characteristics of Responsible AI Frameworks:     * Reliability     * Explainability & Interpretability     * Fairness     * Guidance & Oversight     * Transparency     * Privacy & Security     * Accountability

AI Fairness: Definitions, Methods, and Metrics

• Defining Fairness: Impartial and just treatment or behavior without favoritism or discrimination. Achieving this is difficult because determining what is "fair" is complex.
• Equality vs. Equity     * Equality: Being treated exactly the same.     * Equity: Having equal access to the same opportunities.
• Disparate Treatment vs. Disparate Impact     * Disparate Treatment: Explicitly treating individuals differently or unfairly. Liability is imposed if there is an explicit classification based on a protected attribute or an intent/motive to discriminate.     * Disparate Impact: Using a policy that is neutral on its face but results in a disproportionately adverse impact on minority groups. Liability is still imposed even without discriminatory intent.
• Defining and Quantifying Fairness in ML Systems: Requires three steps:     1. Determining the right definition of a fair outcome for the specific use case.     2. Determining who selects and defines that outcome (and for whom).     3. Selecting the metrics for measurement.
• Individual vs. Group Fairness Tradeoff     * Individual Fairness: Treating similar individuals similarly.     * Group Fairness: Achieving the same outcomes across different demographics by keeping ratios equal.
• Individual Fairness Approaches:     * Aware Approach: The system is aware of protected attributes (race, gender, etc.). Strength: Can define similarity directly. Limitation: Similarity metrics might miss important information, and it is hard to determine the appropriate metric function.     * Unaware Approach (Blindness): The algorithm is blinded to identifiable factors or prohibited attributes. Strength: Reduces unconscious bias in simple settings. Limitation: Success is often tied to historical access to resources, not just merit; legal scholars argue it doesn't promote fair outcomes; other variables can serve as proxies (hidden attributes).
• Group Fairness Metrics:     * Demographic (Statistical) Parity: Population ratios in outcomes should be consistent. Advantage: Ensures representation. Limitation: May lead to underrepresentation of the majority population.     * Equality of False Negatives: Ensures the false negative ratio is consistent across groups. Advantage: Reduces harm from missed diagnoses. Limitation: Does not address false positive errors.     * Equal Opportunity: Ensures the false positive ratio is consistent. Advantage: Fairness for those who truly need treatment. Limitation: Does not address false negative errors.     * Equality as Equalized Odds: Combines same true-positive and false-positive rates. Advantage: Balances fairness better and reduces discrimination from unequal mistakes. Limitation: More complex, can reduce overall accuracy, and is costly.
• The Accuracy-Fairness Tradeoff: Organizations must often trade off model performance (accuracy) to satisfy fairness requirements. Fairness is complex because it depends on the specific situation, and multiple fairness requirements cannot always be satisfied simultaneously.

AI Bias: Sources and Mitigation

• Definition of AI Bias: Occurs when an algorithm produces results systematically prejudiced against specific groups. It affects opportunities, health, and quality of life.
• Sources of Bias: Rooted in problems with training data or flaws in testing and use-case conceptualization.
• Identifying and Measuring Bias:     * Needs computational measures based on definitions of fairness.     * Considerations: Timing of identification, legal issues, and selection of mitigation strategies.     * Challenges: Lack of demographic data to compare against outputs.
• Representational vs. Allocative Harms     * Representational Harms: Discriminatory depictions or algorithmically filtered depictions. These are easier to identify through computational fairness measures.     * Allocative Harms: Unfair distribution of goods, services, resources, or opportunities. These are captured by computational measures focused on distribution.
• Bias Mitigation Strategies:     * Input Data: Collect more/better data, find better proxies, adjust weights, or use bias-corrected synthetic data.     * Decisions/Outputs: Adjusting thresholds for outputs or using binary outcomes.

Explainability in AI

• Definition: The ability to understand and explain machine learning outputs in human terms. It is critical for monitoring, accountability, and ensuring fairness/debiasing.
• Objectives of Explanation:     * To Justify: Why the AI is needed and why an outcome occurred.     * To Control: Preventing errors and identifying model drift.     * To Discover: Helping humans learn from the data.     * To Improve: Continuous model refinement.
• Stakeholders: End users, developers, system builders, decision-makers, and regulatory bodies.
• People vs. Machine Explanations:     * People Explanations: Reasoning behind human-made decisions.     * Machine Explanations: How the model arrived at outputs given specific inputs.
• Types of Machine Explanations:     * Global vs. Local: Global explains the whole model logic; Local explains a specific prediction.     * Intrinsic vs. Post-hoc: Intrinsic models are inherently simple; Post-hoc applies methods to "black boxes" (like neural networks) after training.     * Model Specific vs. Model Agnostic: Specific works for certain models; Agnostic works for any model by looking at input-output pairs.
• The Accuracy-Explainability Tradeoff: Generally, as the ability to explain a model increases, its accuracy decreases.
• Conditions for Explanations:     * Not Needed: When a model doesn't deal with human treatment decisions or when informed consent for a black box is provided.     * Needed: To show respect, to inform users on how to improve results, when outputs are strange, or when justifying treatment is ethically required.
• Calibrating Trust: Explanations help avoid "overreliance" and "algorithmic aversion." Trust should be calibrated based on situational stakes and model confidence.
• Example-based Explanations: Using similar past cases (e.g., in loans or hiring) to explain a current prediction.
• Confidence Displays: Showing the AI's level of certainty. Useful for prioritizing cases for "human in the loop" intervention, but can be confusing if misrepresented.

Privacy in the Era of AI

• Definition: The claim of individuals/groups to determine when, how, and to what extent information about them is communicated. It involves ethics, regulatory compliance, and cybersecurity.
• Privacy Act of 1974: Centers on "notice and consent," which is becoming ineffective because privacy policies are too long to read, especially with IoT and chatbots.
• AI-Unique Privacy Issues:     * Data Hunger: AI requires massive datasets.     * Analysis Power: AI magnifies the speed and power of personal info analysis.     * Data Persistence: Data exists longer than the humans who created it.     * Data Repurposing: Using data for reasons other than the original intent.     * Data Spillovers: Collecting data on people who are not the intended targets.
• Four Elements of Privacy Level:     1. Transparency: Knowing what is collected and with whom it is shared.     2. Data Control: Ability to edit/delete info and opt in/out.     3. Opt out by default: Whether the default state is non-collection.     4. Full services: Whether an organization adjusts service levels based on shared data.
• The Five Levels of Organizational Privacy:     1. Blindfolded and Handcuffed: Not knowledgeable and not in control (Passive).     2. Handcuffed: Knowledgeable but not in control (Passive).     3. Pressured: Knowledgeable with some degree of control.     4. Slightly Curtailed: Knowledgeable; data is not used without consent.     5. Free and Grateful: Fully independent; users decide exactly what data to provide.
• The Data Economy:     * Surveillance Capitalism: Monitoring behavior to predict/influence future actions.     * Attention Economy: Treating human attention as a scarce, valuable resource.     * Intention Economy: Using AI to predict and influence user desires.
• Privacy Tradeoffs: Privacy often conflicts with the greater good (safety/security), innovation, and prediction accuracy.

AI Governance and Ethics Statements

• Main Components of Structure: Roles and responsibilities, policies, and processes/procedures.
• AI Ethics Statements: The 4-Step Actionable Approach:     1. Identify Ethical Nightmares: State values by imagining worst-case scenarios.     2. Connect to Mission: Explain why values matter relative to the organization's purpose.     3. Define the Impermissible: Explicitly state what is ethically off-limits.     4. Articulate Realization: Describe how to specifically avoid nightmares or realize goals.
• 7-Step Approach to AI Governance Structure:     1. Articulate clear AI ethical standards.     2. Create organizational awareness.     3. Provide teams with tools and processes for mitigation.     4. Establish expert oversight (beyond just data scientists).     5. Assign role-specific responsibilities.     6. Establish an AI ethical risk program with KPIs.     7. Ensure executive ownership (provides authority and resources).
• Ethical Case Law: Creating a repository of prior ethical dilemmas and their resolutions (real or fictional) to develop skills in tackling tough questions a priori.
• AI Ethics Committee (AIEC):     * Function: Oversight of in-house and third-party AI products.     * Membership: Data scientists, ethical experts, subject matter experts, and at least one unaffiliated member.     * Jurisdiction: Reviewing applications across all functions (HR, Marketing, etc.).
• Accountability Structure: Includes role-specific responsibilities, incentives/disincentives, and performance evaluations.

AI Ethics for Developers

• Why Developer Tools Fail: Lack of buy-in, tools that don't fit the workflow, lack of training, or no organizational incentive.
• Harming vs. Wronging:     * Harming: Affecting psychological or physical states.     * Wronging: Defaulting on obligations, violating rights, or preventing someone from receiving what they deserve. Focusing on "wronging" is more productive for risk mitigation.
• Five Categories for Identifying Risk:     1. What you create: Should we even do this?     2. How you create it: Does the creation process itself involve ethical risk?     3. What people do with it: Risks from ignorant or malicious users.     4. What impacts it has: Behavior of AI "in the wild."     5. What you do about impacts: Accounting for unforeseen consequences after deployment.
• Ethics by Design: The intentional embedding of ethical principles into the entire software development lifecycle.

AI Regulation and Intellectual Property

• The Balancing Act: Regulating to safeguard against negative consequences without stifling innovation.
• Spectrum of Regulation: Ranges from no regulation and self-regulation to regulation redesign, new regulation, and total moratoriums (Hard law vs. Soft law).
• The EU AI Act (Risk Tiers):     * Unacceptable Risk: Banned systems.     * High Risk: Require assessment and registration in an EU database.     * Generative AI: Transparency requirements regarding content and copyrighted training data.     * Minimal Risk: Basic transparency requirements.
• The AI Bill of Rights (US): Identifies principles for safe systems, algorithmic discrimination protection, notice/explanation, and human alternatives. Applies to systems impacting rights and access to resources.
• Legal and IP Questions:     * Liability: Re-examining law to balance liability between users, organizations, and developers.     * Personhood: Awarding personhood to AI creates problems regarding lack of consciousness, intent, and accountability.     * IP Ownership: Determining who owns AI-generated creations and who is responsible for them.