Artificial Intelligence AI Ethics and Bias 2 — Questions and Answers
Question 1: What is 'differential privacy' designed to protect?
- Model accuracy differences between groups
- The privacy of individuals in datasets by adding calibrated noise to outputs (Correct answer)
- Differential model architectures
- Separate encryption keys per user
Correct answer: The privacy of individuals in datasets by adding calibrated noise to outputs
Differential privacy adds carefully calibrated noise to data or outputs so that individual records cannot be inferred from results.
Question 2: What is an 'adversarial example' in the context of AI safety?
- A training sample with a wrong label
- An input crafted to fool a model into making an incorrect prediction (Correct answer)
- A sample from an adversarial distribution
- A competing model's training data
Correct answer: An input crafted to fool a model into making an incorrect prediction
Adversarial examples are inputs with small, often imperceptible perturbations that cause AI models to make confident but wrong predictions.
Question 3: What is 'AI alignment' research primarily concerned with?
- Aligning model architecture layers properly
- Ensuring AI systems behave in accordance with human values and intentions (Correct answer)
- Synchronizing distributed training across GPUs
- Matching AI predictions to database schemas
Correct answer: Ensuring AI systems behave in accordance with human values and intentions
AI alignment studies how to ensure advanced AI systems pursue goals that are beneficial and consistent with human values.
Question 4: Which US agency is primarily responsible for regulating AI in consumer financial products?
- FDA
- CFPB (Consumer Financial Protection Bureau) (Correct answer)
- FTC (Federal Trade Commission)
- NIST
Correct answer: CFPB (Consumer Financial Protection Bureau)
The CFPB oversees AI-driven credit scoring and lending decisions to ensure they comply with fair lending laws.
Question 5: What does 'debiasing' a dataset mean?
- Removing all variance from training samples
- Correcting imbalances or prejudicial patterns in training data to reduce unfair outcomes (Correct answer)
- Reducing the dataset size to speed up training
- Applying data augmentation equally across classes
Correct answer: Correcting imbalances or prejudicial patterns in training data to reduce unfair outcomes
Debiasing involves identifying and reducing skewed representations or prejudicial labels in training data that lead to biased models.
Question 6: What is 'human-in-the-loop' AI?
- An AI that trains only on human-generated data
- A system where human judgment is incorporated at critical decision points (Correct answer)
- A chatbot that uses human agents to respond
- A model fine-tuned with human feedback only once
Correct answer: A system where human judgment is incorporated at critical decision points
Human-in-the-loop systems keep a human reviewer involved at key stages to validate, correct, or approve AI decisions.
Artificial Intelligence AI Ethics and Bias 2
What is 'differential privacy' designed to protect?