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 1: AI Governance and Program Management
Q641 A recruitment model trained on historical hiring data favors one gender because past decisions were skewed. Under the NIST SP 1270 categories, this is MOST likely which type of bias?
- A. Human bias that arises from the way one user phrases a single prompt to a generative tool during testing
- B. Hardware bias caused by the processors used in the training environment
- C. Statistical bias caused only by a coding error in a database query
- D. Systemic bias, where institutional practices advantage some groups over others
Answer: D
The manual cites a recruitment tool that favored male candidates as an example of systemic bias, in which institutional procedures advantage some groups. Fairness constraints and bias audits before deployment could have helped.
Domain 1: AI Governance and Program Management
Q642 A facial recognition system has far higher error rates for some demographic groups because its training images were not representative. Which action would MOST directly reduce the disparity?
- A. Increase the screen resolution of the cameras used by the vendor
- B. Remove the human reviewers who currently check uncertain matches, which saves cost and speeds up decisions
- C. Publish a statement that the system is accurate on average
- D. Improve the diversity and representativeness of the training dataset
Answer: D
Statistical or computational bias comes from samples that are not representative, and the manual says improving dataset diversity helps. Accuracy on average hides group gaps, and removing reviewers weakens safeguards.
Domain 1: AI Governance and Program Management
Q643 Two analysts use the same generative AI tool but reach different conclusions because they phrase their prompts differently and trust their first result. Which control is MOST appropriate for this human bias?
- A. Giving each analyst a separate model with no shared review
- B. User awareness training and requiring cited, explainable output
- C. Switching off all logging so analysts cannot compare their sessions
- D. Restricting the tool to the most senior analyst only
Answer: B
Human biases such as anchoring and confirmation can be reduced through user awareness and training and by ensuring output is cited and explainable. Removing logs or review hides the problem.
Domain 1: AI Governance and Program Management
Q644 Which set of ethical goals is MOST central to how the manual says AI solutions must be designed?
- A. Guarantee perfect accuracy, remove human review and automate approvals
- B. Minimize bias and ensure fairness, be transparent and explainable, and promote trust and safety
- C. Maximize processing speed, reduce licensing cost and shorten release cycles
- D. Increase data volume, extend retention periods and broaden sharing with commercial partners and data brokers
Answer: B
The three goals are minimizing bias and ensuring fairness, ensuring transparency and explainability, and promoting trust and safety to humans and data. Speed, cost and automation do not capture ethical design.
Domain 1: AI Governance and Program Management
Q645 A team argues that an AI feature that harms vulnerable groups is acceptable if overall accuracy is high. Which response is MOST consistent with the manual?
- A. Harm to small groups is acceptable if the project meets its schedule
- B. Fairness means equitable and accurate results that do not harm vulnerable or marginalized groups
- C. Fairness only applies to systems that make hiring decisions
- D. Accuracy across the whole population is the only fairness measure that matters for regulators and customers
Answer: B
Fairness in AI means promoting equitable, accurate results that do not harm vulnerable groups. Average accuracy can hide unfair outcomes, and fairness is not limited to one use case.
Domain 2: AI Risk Management
Q646 A company starts its AI risk program and wants the MOST useful first step in risk identification. What should it do?
- A. Purchase a monitoring product and wait for it to report which AI risks exist across the organization's network
- B. Rank vendors by price and select the cheapest for review
- C. Ask developers to describe risks they think are unlikely
- D. Inventory AI use across the organization, including AI built into existing tools
Answer: D
The manual says identification starts with accurately inventorying AI use, since employees may not know AI is built into tools and devices they already use. Buying tools or ranking vendors skips the inventory.
Domain 2: AI Risk Management
Q647 An HR chatbot and a public customer-support chatbot share the same backend without segmentation. Which risk is MOST likely?
- A. A remote attacker uses prompt injection on the support bot to reach staff personal data
- B. The HR system automatically gains stronger encryption from the support system
- C. The two chatbots become more accurate because they share training data
- D. Customer queries are answered faster because the backend is shared, so overall risk is reduced and costs fall
Answer: A
The manual gives this example: without segmentation, prompt injection on one AI system can expose staff PII held by another. Sharing a backend widens the attack surface.
Domain 2: AI Risk Management
Q648 Which scenario BEST illustrates the OWASP LLM risk called excessive agency?
- A. An LLM produces a confident but false statement about a historical date that a user then repeats in a report
- B. A prompt in the system configuration contains an internal password
- C. An LLM connected to email and payment functions takes actions beyond what the user intended
- D. An attacker floods an LLM with requests until the service slows down
Answer: C
Excessive agency is the ability of an LLM to make decisions or call functions in other systems via extensions based on prompts. False statements are misinformation, floods are unbounded consumption, and leaked prompts are system prompt leakage.
Domain 2: AI Risk Management
Q649 A user receives a fluent answer from an LLM that cites a policy that does not exist. Which OWASP LLM risk category does this BEST match?
- A. Supply chain, where third-party packages are tampered with
- B. Misinformation, where false or misleading output appears credible
- C. Vector and embedding weaknesses, where stored data is exposed
- D. Unbounded consumption, where excessive requests cause outages
Answer: B
Misinformation occurs when LLMs produce false or misleading information that appears credible, often because of hallucinations. The other categories involve resource use, tampering or retrieval data exposure.
Domain 2: AI Risk Management
Q650 An attacker submits a huge volume of inference requests, driving up costs and degrading service for other users. Which OWASP LLM risk is this MOST likely?
- A. System prompt leakage
- B. Unbounded consumption
- C. Improper output handling
- D. Sensitive information disclosure
Answer: B
Unbounded consumption occurs when end users can draw an excessive number of inferences, leading to denial of service, financial loss and even model theft. The other categories concern data exposure, unsafe output use or leaked instructions.