Domain 4: deployment and use of AI

Question 1

Variance

Answer 1

High variance: data spread, overfitting, low bias (high complexity)

Low variance: data tight, increase bias (decrease complexity)

Think all one race in a data set as low variance

Question 2

AI Dev Lifecycle - Deploy

Answer 2

Assess readiness
Continuous monitoring
Deployment and implementation
Adapt and govern

Question 3

Two categories of AI products

Answer 3

1. Integrated into business operations
   
   • hiring, screening, chatbots
2. COTS (commercial off the shelf)
   
   • chatGPT, Grammerly

Question 4

Evaluating third-party agreements (6)

Answer 4

Product category
Data
Tech specs
Security and safety
Bias and fairness
Monitoring and maintenance

Question 5

Types of disclosures (5)

Answer 5

End user agreements
Sector specific
Jurisdiction specifics
System specific
Rights specific

Question 6

Accountability mechanism

Answer 6

measures implemented to foster responsible AI and trustworthy dev life cycle

Includes audits, assessments, and human oversight

User and regulator assurance

Question 7

Automation tools for AI (2)

Answer 7

AI Verify (Singapore)
OECD Model Card Regulatory Check

Question 8

Incident response documentation

Answer 8

Incident
Model version
Dataset
Inputs and outputs
Cause (if known)
Status of response
Mitigation
Stakeholder comms
Lessons learned, after action report

Question 9

Failure causes for AI

Answer 9

Model decay
Model complexity
Cybersecurity attacks
Other
- Brittleness
- Lack of robustness
- Poor quality data
- Insufficient testing
- Model or data drift

Question 10

Incident response - 6 stages

Answer 10

Preparation
Identification
Containment
Eradication
Recovery
Lessons learned