GCP Professional Standards & Ethics 4 — Questions and Answers
Question 1: What is 'algorithmic bias' in the context of machine learning systems?
- Hardware acceleration techniques that favor certain ML architectures
- Systematic unfair outcomes resulting from prejudiced training data or flawed model design (Correct answer)
- A preference for deploying models in specific geographic cloud regions
- Optimization methods that reduce model size and inference latency
Correct answer: Systematic unfair outcomes resulting from prejudiced training data or flawed model design
Algorithmic bias occurs when a model produces systematically unfair results due to biased training data, flawed assumptions, or discriminatory design choices.
Question 2: Which practice most directly supports transparency in AI systems used for high-stakes decisions such as loan approvals?
- Using exclusively black-box models to maximize predictive accuracy
- Applying model explainability and interpretability techniques (Correct answer)
- Concealing model architecture from regulators to protect IP
- Encrypting model weights to prevent reverse engineering
Correct answer: Applying model explainability and interpretability techniques
Explainability techniques allow stakeholders and regulators to understand why a model made a particular decision, supporting accountability.
Question 3: When deploying a facial recognition system for law enforcement use, what is the primary ethical concern that professionals must address?
- The cost of model training on cloud infrastructure
- API response latency under peak load conditions
- Potential for demographic bias causing disproportionate impact on certain groups (Correct answer)
- Long-term storage costs for the image dataset
Correct answer: Potential for demographic bias causing disproportionate impact on certain groups
Facial recognition systems have documented higher error rates for certain demographic groups, raising serious civil rights and due process concerns.
Question 4: What does the principle of 'human oversight' require in automated AI decision-making pipelines?
- That human engineers manually build every AI model component
- That consequential automated decisions include mechanisms for human review and intervention (Correct answer)
- That AI systems are never deployed in production without constant monitoring
- That all AI decisions are final and cannot be appealed by affected individuals
Correct answer: That consequential automated decisions include mechanisms for human review and intervention
Human oversight requires that humans can review, override, and correct consequential AI decisions rather than treating them as infallible.
Question 5: A company wants to use ML to rank employees for promotions. What ethical safeguard is most critical?
- Maximizing the size of the training dataset regardless of data quality
- Auditing the model for discriminatory bias against protected classes before deployment (Correct answer)
- Achieving the highest possible accuracy metric on the validation set
- Keeping the model proprietary to prevent gaming by employees
Correct answer: Auditing the model for discriminatory bias against protected classes before deployment
Employment decisions affect people's livelihoods, so auditing for bias against legally protected classes is essential before deployment.
Question 6: What is 'data poisoning' as an attack against ML systems?
- Deliberately corrupting training data to cause a model to learn incorrect or harmful behaviors (Correct answer)
- Encrypting training datasets to prevent unauthorized access
- Removing personally identifiable information from datasets before training
- Compressing training data to reduce cloud storage costs
Correct answer: Deliberately corrupting training data to cause a model to learn incorrect or harmful behaviors
Data poisoning involves injecting malicious or misleading examples into training data to manipulate the model's learned behavior.
Question 7: Under responsible AI principles, what should an organization do when an AI system begins producing harmful or discriminatory outputs in production?
- Continue operating until the next quarterly model review cycle
- Immediately investigate, pause the system if necessary, and implement corrective measures (Correct answer)
- Attribute the harm to end users who submitted the inputs that triggered the outputs
- Delete all training data and retrain the model from scratch before investigating
Correct answer: Immediately investigate, pause the system if necessary, and implement corrective measures
Responsible AI practice requires halting harm promptly, investigating root causes, and implementing fixes rather than waiting for scheduled reviews.
GCP Professional Standards & Ethics 4
What is 'algorithmic bias' in the context of machine learning systems?