This is not a replacement for the official study guide. Please read the official study guide first — use this exam prep only to reinforce or practice your understanding.
Free AAISM exam practice questions with answers and explanations, organized by domain and part.
Domain 3: AI Technologies and Controls
Q551 Which control would MOST reduce the privacy risk of a model that retains personal details from training data?
- A. Add more personal data to improve model accuracy on rare cases
- B. Publish training samples so researchers can detect memorization
- C. Remove or pseudonymize identifiers before training and test for memorization
- D. Turn off privacy reviews to shorten the development schedule
Answer: C
Removing or pseudonymizing identifiers and testing for memorization reduce leakage. Adding data, publishing samples and removing reviews raise risk.
Domain 3: AI Technologies and Controls
Q552 A data subject asks an organization what automated decisions were made about them. Which AI governance capability is MOST needed to respond?
- A. A new marketing campaign that explains the benefits of the system
- B. A faster GPU cluster to reduce training time for the model
- C. Explainability and records of the decision logic and data used
- D. A larger vector database to store more document embeddings
Answer: C
Responding requires explainability and decision records. Hardware, storage capacity and marketing do not provide the explanation.
Domain 3: AI Technologies and Controls
Q553 An AI customer chatbot may produce harmful advice. Which safety control is MOST effective?
- A. Allowing unrestricted outputs to keep the conversation natural
- B. Removing all monitoring to avoid incorrect alerts on the chatbot
- C. Guardrails such as content filters, escalation to humans and continuous testing
- D. Training only on unreviewed social media posts for realism
Answer: C
Guardrails with human escalation and continuous testing reduce harmful outputs. Removing monitoring and unreviewed data raise risk.
Domain 3: AI Technologies and Controls
Q554 Which approach is MOST appropriate for a safety-critical AI function where errors could cause harm?
- A. Full autonomy with no human override to make decisions faster
- B. Fail-safe design, rigorous testing and defined human override
- C. Deployment without testing to reduce cost and time to market
- D. Reliance on vendor claims alone without independent verification
Answer: B
Safety-critical functions need fail-safe design, rigorous testing and human override. Full autonomy, no testing or vendor claims alone are inadequate.
Domain 3: AI Technologies and Controls
Q555 A security team feeds AI logs into a SIEM. What is the MOST important benefit?
- A. Replacing the incident response team with automated dashboards
- B. Reducing the number of logs collected to zero to save storage
- C. Eliminating the need for access controls on the AI platform
- D. Correlating AI events with other security data for earlier detection and investigation
Answer: D
Correlating AI events with broader security data speeds detection and investigation. A SIEM does not replace access controls or incident response.
Domain 3: AI Technologies and Controls
Q556 An AI audit finds that no one can say who approved a model change. Which control gap is MOST apparent?
- A. Strong role separation between developers and approvers
- B. Excessive encryption of the model files stored in the repository
- C. Weak governance and traceability of approvals
- D. Too many backups of the model artifacts held by the team
Answer: C
Missing approval records indicate weak governance and traceability. The other options describe unrelated or positive conditions.
Domain 3: AI Technologies and Controls
Q557 Which control MOST reduces the risk that an LLM agent performs unauthorized actions through connected tools?
- A. Granting administrator rights to the agent so it can complete tasks
- B. Allowing unrestricted API keys that are shared across all agents
- C. Disabling logs of agent actions to protect user privacy
- D. Least-privilege permissions for tools with approval for sensitive actions
Answer: D
Agents should have the minimum permissions necessary, with human approval for sensitive actions and logging. Administrator rights and unrestricted keys enable misuse.
Domain 3: AI Technologies and Controls
Q558 Which measure is MOST appropriate to protect the confidentiality of model weights?
- A. Store the weights on personal drives that engineers manage themselves
- B. Publish the weights openly so the community can review them
- C. Share the weights through unmanaged chat tools for convenience
- D. Restrict access, encrypt at rest and monitor downloads and exports
Answer: D
Model weights are valuable assets, so access restrictions, encryption and monitoring of exports protect them. Publication or unmanaged sharing exposes them.
Domain 3: AI Technologies and Controls
Q559 A team wants early warning of data drift on an input feature. Which monitoring method is MOST appropriate?
- A. Disable monitoring after launch to reduce operating cost
- B. Compare live input distributions with training baselines using thresholds and alerts
- C. Check model accuracy once a year during the annual review
- D. Rely on customer complaints to indicate when inputs have changed
Answer: B
Comparing live distributions with training baselines and alerting on thresholds detects drift early. Annual checks and complaints are late, and disabling monitoring removes visibility.
Domain 3: AI Technologies and Controls
Q560 After a model update, which monitoring step is MOST important to confirm that it behaves as expected?
- A. Assume the updated model is better because it is newer
- B. Compare performance and fairness metrics against the prior version and baselines
- C. Check only the file size of the new model artifact
- D. Wait for the quarterly report to review changes in behavior
Answer: B
Comparing performance and fairness with previous versions and baselines confirms improvements and detects regressions. Newer does not mean better.