Module 5 (V 2.0): Governing the release, monitoring and maintenance of the AI model

Question 1

Name some AI-specific risks that organizations should monitor for after deployment?

Answer 1

• Model inversion
• Extraction
• Poisoning
• Evasion

Question 2

Name two predictable AI-specific risks post-deployment and mitigations for these risks.

Answer 2

1) Risk: Using the model for a different purpose than originally intended. Mitigation: Documentation (via model card or your org’s standard documentation).

2) Risk:Introducing new data into an algorithm. Mitigation: Store snapshots of the algorithm and its output.

Question 3

List some considerations for monitoring AI systems.

Answer 3

• Inventory all AI systems and attach a risk score to each.
• Create a challenger model to assess drift and unexpected results.
• Implement tests to assess performance, reliability, and safety.
• Monitor changes to the model.
• Have a procedure in place to deactivate a model or localize it as needed.
• Follow current industry best practices.

Question 4

What are some current industry best practices for managing and monitoring external and internal risks?

Answer 4

• Define a baseline to measure future iterations of the model.
• Retrain with new data as needed and with human input and feedback to improve performance as the data changes.
• Determine and prioritize the level of the risks and the appropriate responses.
• Conduct internal or external red teaming exercises for AI systems (should also be done pre-deployment).
• Consider bug bashing/bug bounties to generate user engagement and extensive feedback.

Question 5

What are some causes of AI incidents?

Answer 5

• Brittleness
• Lack of robustness
• Lack of quality data
• Insufficient testing
• Model or data drift

Question 6

Name two AI governance automation tools.

Answer 6

1) AI Verify: Launched by the Singapore government, it is an AI governance testing framework and toolkit to help systems meet performance benchmarks. It validates AI system performance against 11 ethics principles
2) The Model Card Regulatory Check app - It automates regulatory compliance of AI systems based on accepted AI documentation tools like model cards.