AI Ethics Comprehensive Quiz & Projects
30 questions on AI Ethics Tutorial.
Question 1: What is the primary source of 'Algorithmic Bias' in modern machine learning systems?
- A. Code syntax errors in neural network backpropagation loops.
- B. Flawed, unrepresentative, or historically prejudiced training data that the model learns to replicate. β (correct answer)
- C. The hardware architecture of graphics processing units (GPUs).
- D. Over-aggressive security guardrails in system prompts.
Explanation: AI models mirror the statistical distributions of their input. If training data reflects historical bias, the model inherits it.
Question 2: What is the difference between SHAP and LIME in Explainable AI (XAI)?
- A. SHAP is used for image classification, while LIME is used for text parsing.
- B. SHAP calculates game-theoretic Shapley values to measure global feature contributions, while LIME builds local surrogate models to explain specific predictions. β (correct answer)
- C. SHAP encrypts model outputs, while LIME compresses neural networks.
- D. SHAP requires model training data, while LIME works only on final code.
Explanation: SHAP provides mathematically consistent, global attribute importances. LIME approximates model behaviors locally around single samples.
Question 3: Under the EU AI Act, which category of AI systems is completely prohibited?
- A. Systems that process clinical trials data.
- B. High-risk systems like automated resume scanners.
- C. Unacceptable-risk systems like social scoring by governments or cognitive behavioral manipulation. β (correct answer)
- D. Generative AI systems like translation chatbots.
Explanation: The EU AI Act takes a risk-based approach. Unacceptable risk systems are banned due to threat to safety and rights.
Question 4: Why is the Human-in-the-Loop (HITL) architectural design critical in high-stakes AI systems (e.g. healthcare, hiring)?
- A. It speeds up system response times on server clusters.
- B. It ensures that a human reviews and authorizes AI recommendations before high-impact decisions are executed. β (correct answer)
- C. It allows humans to manually update model weight matrices.
- D. It registers model code with regulatory authorities automatically.
Explanation: HITL acts as a safety buffer, preventing fully autonomous failures in critical domains where lives or livelihoods are affected.
Question 5: What does the concept of 'Differential Privacy' mathematically guarantee?
- A. That the model will always run in an encrypted sandbox.
- B. That no user data can be shared with third-party APIs.
- C. That the inclusion of an individual's record in a training set will not noticeably alter the output statistics, protecting individual privacy. β (correct answer)
- D. That model accuracy will increase by at least 10%.
Explanation: Differential privacy injects mathematically controlled noise, allowing models to learn aggregate trends without exposing individual rows.
Question 6: What is the alignment problem in AI?
- A. Aligning text columns in database tables.
- B. The challenge of ensuring that AI systems' goals and behaviors are aligned with human values, ethics, and safety expectations. β (correct answer)
- C. Synchronizing server clocks.
- D. Aligning page elements in HTML layouts.
Explanation: Alignment focuses on avoiding systems that optimize objectives in unexpected or harmful ways.
Question 7: What does 'Explainable AI' (XAI) address?
- A. Explaining coding syntax to beginners.
- B. The methods and techniques that make machine learning models' decisions and inner workings understandable to humans. β (correct answer)
- C. Generating documentation templates.
- D. Writing summaries of database schemas.
Explanation: XAI helps dismantle 'black box' issues, crucial for trust in clinical or financial systems.
Question 8: How does 'Data Minimization' promote privacy in AI pipelines?
- A. By compressing database backups.
- B. By collecting and processing only the minimal amount of personal data necessary to train the target model. β (correct answer)
- C. By reducing the number of parameters in neural networks.
- D. By deleting old model version files.
Explanation: Minimization restricts data collection boundaries, lowering exposure risks if a breach happens.
Question 9: What is a 'Deepfake'?
- A. A complex database index query.
- B. Synthetic media in which a person in an existing image or video is replaced with someone else's likeness using deep generative models. β (correct answer)
- C. An encrypted file storage container.
- D. A server background task logs error.
Explanation: Deepfakes pose ethics risks relating to identity theft, disinformation, and consent violations.
Question 10: What ethical concern is raised by the massive energy consumption of training large foundation models?
- A. High database subscription costs.
- B. Environmental impact and carbon footprint contribution. β (correct answer)
- C. Slow local network routing speeds.
- D. Rapid hardware wear and tear.
Explanation: Training parameters on massive GPU grids requires gigawatt-scale electricity, raising green issues.
Question 11: What does 'Dual-Use' refer to in AI technology policy?
- A. Models running on both CPU and GPU.
- B. AI technologies developed for peaceful civilian purposes that can also be easily modified for harmful military or malicious uses. β (correct answer)
- C. Models that support both SQL and NoSQL.
- D. Writing scripts in both Python and JavaScript.
Explanation: Dual-use tools (like biological synthesis models or network scanners) require strict access controls.
Question 12: What is 'Consent' in the context of training datasets?
- A. Verifying SQL query structures.
- B. Obtaining permission from creators or users to use their data, photos, or text for AI model training. β (correct answer)
- C. Setting administrative user roles in settings.py.
- D. Allowing access to server directories.
Explanation: Training models on scraped web data without explicit consent has triggered copyright disputes.
Question 13: What does the term 'Automation Bias' refer to in human-AI collaboration?
- A. The speed of compiling automated tests.
- B. The human tendency to trust automated decisions blindly, ignoring contradictory evidence or personal judgment. β (correct answer)
- C. Automated scripts throwing syntax errors.
- D. Caching operations running in loops.
Explanation: Automation bias causes supervisors to overlook system errors, accepting false positives.
Question 14: What is 'Model Inversion' in security audits?
- A. Flipping the neural network weight values.
- B. An attack vector where an attacker reconstructs training data features or identifies personal records by querying the model's API. β (correct answer)
- C. Changing database connection ports.
- D. Deleting obsolete model configurations.
Explanation: Model inversion leaks private details, requiring defense vectors like differential privacy.
Question 15: What is the role of an 'AI Ethics Board'?
- A. A developer group for optimizing CPU loops.
- B. A multidisciplinary committee that reviews, evaluates, and establishes guidelines for the ethical development and deployment of AI projects. β (correct answer)
- C. A compliance team checking database schema names.
- D. A group that writes system prompts.
Explanation: Ethics boards evaluate systemic risks, societal impacts, and bias in product pipelines.
Question 16: What does 'Algorithmic Fairness' aim to accomplish?
- A. Ensuring all models compile at identical speeds.
- B. Preventing AI systems from generating discriminatory outcomes or favoring specific groups over others based on protected attributes. β (correct answer)
- C. Restricting database sizes to standard volumes.
- D. Routing requests.
Explanation: Fairness metrics measure equality in predictions across demographics (e.g. gender or race).
Question 17: What is 'Stochastic Parrot' a metaphor for in LLM critiques?
- A. A model repeating database errors.
- B. Large language models generating coherent-sounding text by statistically mapping training sequences without real semantic understanding or context. β (correct answer)
- C. A server loop that hangs.
- D. Models outputting randomized data structures.
Explanation: The term highlights that LLMs match syntax patterns without possessing conceptual understanding.
Question 18: How does 'Transparency' differ from 'Explainability'?
- A. Transparency is encrypting API headers.
- B. Transparency involves documenting data sources, model cards, and training processes, while Explainability focuses on explaining specific model outputs. β (correct answer)
- C. Explainability works only on SQL tables.
- D. There is no difference.
Explanation: Transparency concerns documentation and origin audits. Explainability addresses logic outputs.
Question 19: What is the risk of 'Feedback Loops' in predictive policing AI?
- A. Server cache overflow failures.
- B. The model predicting high crime in areas with more historical arrests, leading to more police dispatch and arrests, reinforcing the initial biased data. β (correct answer)
- C. Modifying model parameters to zero.
- D. Deleting database log lists.
Explanation: Feedback loops amplify historical biases, creating self-fulfilling prediction metrics.
Question 20: What is 'Data Provenance'?
- A. Compressing database tables.
- B. Documenting the origin, history, ownership, and collection methods of training data. β (correct answer)
- C. Setting environment variables.
- D. The structure of JSON payloads.
Explanation: Data provenance ensures training datasets are audited for legal compliance and consent.
Question 21: What is the ethical concern with 'Emotional Recognition' AI?
- A. It consumes high network bandwidth.
- B. It lacks scientific validity, can misinterpret cultural expressions, and can be used for invasive surveillance. β (correct answer)
- C. It requires expensive GPU configurations.
- D. It does not support JSON responses.
Explanation: Classifying emotional states from facial features carries high privacy and bias risks.
Question 22: Under the EU AI Act, what classification is given to automated CV parsing tools used in hiring?
- A. Minimal Risk
- B. High Risk β (correct answer)
- C. Unacceptable Risk
- D. Medium Risk
Explanation: Employment screening tools are High Risk, requiring audits, logging, and human oversight.
Question 23: What is 'Bias Mitigation'?
- A. Deleting the model training scripts.
- B. The process of detecting and reducing systematic bias in training data, algorithms, or model outputs. β (correct answer)
- C. Restricting port access limits.
- D. Caching operations optimization.
Explanation: Mitigation includes dataset balancing, pre-processing, and adversarial debiasing.
Question 24: What does the term 'Anthropomorphism' mean regarding AI?
- A. The speed of the event loop.
- B. The human tendency to attribute human traits, emotions, and consciousness to AI assistants. β (correct answer)
- C. A database indexing design.
- D. Compiling model weights.
Explanation: Anthropomorphism can trick users into trusting systems, sharing confidential variables.
Question 25: What is the concern of 'Data Poisoning' in collaborative AI training?
- A. Deleting database records via SQL.
- B. Attackers injecting malicious, misleading samples into training sets to compromise model integrity or create backdoors. β (correct answer)
- C. Disabling the local firewall.
- D. Compressing files into zip format.
Explanation: Data poisoning ruins classification boundaries, introducing security vulnerabilities.
Question 26: What does 'AI Safety' address?
- A. Setting passwords on servers.
- B. Preventing accidental, unintended, or harmful behaviors from AI systems during operation. β (correct answer)
- C. Restricting access to repository directories.
- D. Backing up code files.
Explanation: Safety research includes alignment, robust boundaries, and fail-safe parameters.
Question 27: What is a 'Model Card'?
- A. A hardware acceleration chip.
- B. A short document providing structured metadata about a model's performance, limitations, intended use, and bias evaluations. β (correct answer)
- C. A template for styling web pages.
- D. A database schema diagram.
Explanation: Model cards promote transparency, helping developers select appropriate models for use cases.
Question 28: What is the ethical challenge of 'Gig Work' labeling for AI datasets?
- A. High data transmission latencies.
- B. Underpaid, precariously employed crowd workers processing toxic content without mental health support. β (correct answer)
- C. Writing scripts in JavaScript.
- D. The structure of JSON vectors.
Explanation: Behind clean models are humans labeling toxic text and graphic images for low wages.
Question 29: How does an 'Auditing API' help verify model bias?
- A. It speeds up query runs.
- B. It exposes model inferences to third-party reviewers to check error distributions across demographics. β (correct answer)
- C. It encrypts communication channels.
- D. It deletes session caches.
Explanation: Audits check outcomes, checking systems for disparate impact issues.
Question 30: What does 'Responsible AI' encompass?
- A. Writing clean code syntax.
- B. Synthesizing ethics, safety, fairness, privacy, and accountability principles into AI product lifecycles. β (correct answer)
- C. Migrating database tables.
- D. Launching applications to cloud environments.
Explanation: Responsible AI converts ethical values into technical checkpoints during design.