Free AAIA practice test: 20 questions
Twenty questions across the three AAIA domains, weighted roughly as the real exam is. AAIA is an audit credential: the right answer is usually the one an auditor could evidence, not the one that sounds most thorough.
AI Governance and Risk
Question 1. An auditor is asked to assess whether AI is governed adequately. What should be examined first?
- A. The accuracy of the highest-profile model
- B. The GPU procurement contracts
- C. The AI system inventory, with owners and risk classification
- D. The data science team’s training certificates
Show answer
Answer: C
Without a complete inventory of AI systems with owners and risk ratings, no assurance can be given about coverage. Everything else rests on knowing what exists.
Question 2. Management states that an AI policy exists. What is the strongest evidence that it is operating?
- A. A signed policy document
- B. Completed risk assessments and approvals for systems actually in production
- C. An email announcing the policy
- D. A slide deck presented to the board
Show answer
Answer: B
Evidence of the policy being applied, such as completed assessments and approvals for actual systems, demonstrates operating effectiveness. A signed document alone shows design only.
Question 3. Who should approve deployment of a high-risk AI system?
- A. The data scientist who built it
- B. The cloud provider
- C. The external auditor
- D. The accountable business owner and the designated governance body
Show answer
Answer: D
Accountability sits with business ownership and a governance body with authority to accept the risk. The building team cannot approve its own risk acceptance.
Question 4. An organisation claims its AI risk appetite is defined. What would evidence that claim?
- A. Documented thresholds and cases where systems were escalated or rejected against them
- B. A statement that the organisation is risk-averse
- C. The number of AI projects approved
- D. The size of the AI budget
Show answer
Answer: A
Thresholds that translate appetite into decisions, with documented cases where systems were escalated or rejected, show a real appetite. A general statement does not.
Question 5. Which finding indicates the weakest governance regardless of model quality?
- A. Model documentation is incomplete
- B. Performance metrics are reported quarterly rather than monthly
- C. Production AI systems exist with no identified owner
- D. The training data is stored in two regions
Show answer
Answer: C
If nobody is accountable for a production system, no control can be enforced or remediated. Documentation and metric gaps are serious but recoverable.
Question 6. Third-party AI is embedded in a critical business process. What should governance require?
- A. Acceptance of the vendor’s marketing claims
- B. Due diligence, contractual controls and ongoing monitoring of the provider
- C. A copy of the vendor’s logo usage guidelines
- D. Nothing, as the vendor is responsible
Show answer
Answer: B
Due diligence, contractual assurance and ongoing monitoring extend governance across the supply chain. Accepting vendor claims without evidence is not assurance.
Question 7. Regulation requires human oversight of consequential automated decisions. What evidences compliance?
- A. Records of human review and override for individual decisions
- B. A design document describing intended oversight
- C. The model’s accuracy score
- D. A statement in the privacy notice
Show answer
Answer: A
Records showing a human reviewed and could override specific decisions demonstrate the oversight actually happened. A design document only shows intent.
AI Operations
Question 8. A model in production has not been re-evaluated since deployment 18 months ago. What is the principal concern?
- A. The documentation may be outdated
- B. Training costs may have risen
- C. The model may be using an older framework version
- D. Performance may have degraded undetected through drift
Show answer
Answer: D
Without periodic evaluation, drift and degradation go undetected, so current performance is unknown. The original validation says nothing about today.
Question 9. Which control best supports reproducing a past model decision during an investigation?
- A. Monthly aggregate accuracy reporting
- B. Versioned model artefacts, inputs and parameters with retained decision logs
- C. A backup of the production database only
- D. The developer’s recollection
Show answer
Answer: B
Versioned models, inputs and parameters with retained logs allow a decision to be reconstructed. Aggregate metrics cannot explain an individual outcome.
Question 10. An auditor finds that anyone in the data team can deploy a model to production. Which principle is violated?
- A. Segregation of duties
- B. Data minimisation
- C. Least privilege on storage
- D. Encryption at rest
Show answer
Answer: A
Segregation of duties requires that the person building a model is not the one who approves and deploys it unchecked. This is a classic control weakness.
Question 11. Model changes are pushed without recorded approval. What is the audit finding?
- A. Inadequate model accuracy
- B. Excessive cloud spend
- C. Change management over AI models is not operating
- D. Insufficient training data
Show answer
Answer: C
An absent change control process over models means changes are unauthorised and untraceable. This is a process finding, not a technical performance one.
Question 12. What should an auditor request to assess data quality controls over a training pipeline?
- A. A copy of the raw dataset
- B. Validation rules, and evidence of exceptions detected and resolved
- C. The names of the data engineers
- D. The storage account cost report
Show answer
Answer: B
Documented validation rules with evidence of exceptions raised and resolved shows the control operating. A sample of the data alone shows content, not control.
Question 13. An incident occurred where a model produced discriminatory output for two weeks before detection. Which control most directly failed?
- A. Access control on the model endpoint
- B. Encryption of training data
- C. Vendor due diligence
- D. Ongoing fairness monitoring with alert thresholds
Show answer
Answer: D
Ongoing fairness monitoring with alerting would have detected it far sooner. Pre-deployment testing is necessary but cannot cover post-deployment behaviour.
Question 14. Which measure best limits the impact of an AI agent misbehaving in production?
- A. Least-privilege scoping of the agent’s permitted actions and data
- B. A more detailed system prompt
- C. A larger model
- D. More frequent status reporting
Show answer
Answer: A
Least-privilege scoping of the agent’s actions and data caps what harm is possible. Longer instructions and larger models do not constrain capability.
Question 15. An organisation cannot say which personal data fields feed a production model. What is the consequence for the audit?
- A. No consequence if the model performs well
- B. A minor observation about documentation
- C. A significant finding, as lawful basis and data subject rights cannot be evidenced
- D. A recommendation to increase the budget
Show answer
Answer: C
Without data lineage, neither lawful basis nor data subject rights can be demonstrated, so a significant finding is warranted.
Question 16. Which is the strongest evidence that an AI system's outputs are monitored for quality in production?
- A. A dashboard that exists in the tooling
- B. Dated monitoring records showing thresholds, breaches and remediation actions
- C. An engineer stating that they check regularly
- D. The absence of user complaints
Show answer
Answer: B
Dated monitoring records showing thresholds, breaches and the actions taken demonstrate an operating control across the period.
AI Auditing Tools and Techniques
Question 17. An auditor wants assurance over a population of 400,000 model decisions. Which technique is appropriate?
- A. Statistical sampling with a documented method and confidence level
- B. Reviewing the five decisions management selects
- C. Reviewing whichever decisions are easiest to access
- D. Reviewing none, relying on management representation
Show answer
Answer: A
Statistical sampling with a defined method and confidence level supports a conclusion about the population. Judgemental selection of a handful does not.
Question 18. Management asserts that a control operates continuously. Which test provides the strongest evidence?
- A. Inquiry of the control owner
- B. Observation on a single day
- C. Inspection of the control’s design document
- D. Re-performance or automated testing across the full period
Show answer
Answer: D
Re-performance over a period, or automated testing of the full population, evidences continuous operation. Inquiry and observation on one date are weaker.
Question 19. An auditor uses an AI tool to summarise evidence. What is the key professional requirement?
- A. Disclosing the tool’s brand in the report
- B. Using the tool only outside working hours
- C. Verifying AI-generated summaries against source evidence and retaining responsibility for conclusions
- D. Avoiding AI tools entirely
Show answer
Answer: C
The auditor remains responsible for conclusions and must verify AI-produced summaries against source evidence. Delegating judgement to the tool is not acceptable.
Question 20. Which finding should be rated highest severity in an AI audit report?
- A. Inconsistent naming conventions for model artefacts
- B. A high-risk model in production with no approval or oversight
- C. A backlog of low-priority documentation updates
- D. Quarterly rather than monthly reporting
Show answer
Answer: B
A high-risk system operating without approval or oversight represents an unmanaged risk to the organisation. Naming conventions and backlog size are administrative.
How did you do?
Sixteen or more correct suggests you are close. Below fourteen, the domain guides in this section are the quickest fix — AAIA rewards thinking like an auditor, where evidence beats intention every time.