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
Q501 Which statement about AI and human rights is MOST accurate?
- A. Human rights concerns arise only after a regulator opens a formal investigation
- B. Human rights concerns apply only to governments and public bodies, so private enterprises are not required to assess them in AI projects
- C. Human rights are protected automatically whenever a model reaches high accuracy
- D. AI systems should be assessed for impacts on rights such as privacy, equality and freedom from discrimination
Answer: D
Faults such as bias and discrimination can infringe human rights, so organizations should assess these impacts proactively. Accuracy alone does not protect rights, and enterprises are accountable too. Waiting for regulators is reactive.
Domain 1: AI Governance and Program Management
Q502 A sustainability officer questions the energy use of training large models. Which governance response is MOST appropriate?
- A. Move training to a cheaper data center to reduce cost, and defer any measurement of energy use until the program is mature
- B. Ignore energy use because it is an infrastructure matter only
- C. Include environmental impact in the AI cost-benefit analysis and set sustainability measures
- D. Prohibit all AI training regardless of business value
Answer: C
The manual treats environmental impact as an ethical and societal consideration, so it belongs in the cost-benefit analysis and KPIs. Ignoring it, banning training, or relocating without measurement do not manage the impact.
Domain 1: AI Governance and Program Management
Q503 A bank finds its credit model approves fewer applicants from one postcode, although the model never uses ethnicity. Which type of bias should the bank suspect FIRST?
- A. Network bias from the latency of the application servers, which delays decisions for applicants in some regions more than others
- B. Hardware bias from the processors used in training
- C. Statistical or computational bias from unrepresentative data or proxy variables
- D. Licence bias from the vendor's pricing structure
Answer: C
Proxy variables such as postcode and unrepresentative samples cause statistical or computational bias. Hardware, network and licence factors do not shape model fairness. NIST SP 1270 describes systemic, statistical and human bias.
Domain 1: AI Governance and Program Management
Q504 Prompts written by different analysts lead to very different conclusions from the same GenAI tool. Which bias category does this BEST represent?
- A. Human bias, such as anchoring or confirmation bias in prompt design
- B. Hardware bias from the GPU architecture
- C. Systemic bias embedded in institutional processes, which is mainly produced by the way the organization is structured
- D. Statistical bias from a skewed sample
Answer: A
Different users' prompts can introduce human biases such as anchoring or confirmation bias. Systemic bias stems from institutional practices, and statistical bias comes from unrepresentative data. User training and cited outputs help reduce human bias.
Domain 1: AI Governance and Program Management
Q505 Which tool is MOST appropriate to test a model's behavior in hypothetical situations and visualize performance across data subsets?
- A. A disk defragmenter
- B. A password manager
- C. A network packet analyzer
- D. The What-If Tool
Answer: D
The What-If Tool lets developers probe performance in hypothetical situations and across subsets of data. AI Fairness 360 is another open-source toolkit for bias. Packet analyzers, defragmenters and password managers do not analyze model fairness.
Domain 1: AI Governance and Program Management
Q506 Before deploying a recruitment model, which action would have BEST reduced the risk of gender bias amplification?
- A. Switching to a faster algorithm to shorten the screening process
- B. Increasing the number of input features without a review of their fairness impact
- C. Deploying first and monitoring complaints from candidates after launch
- D. Implementing fairness constraints and running a bias audit before deployment
Answer: D
Fairness constraints and bias audits before deployment help detect and mitigate bias. Adding features without review or deploying first leaves bias unchecked, and speed does not address fairness.
Domain 1: AI Governance and Program Management
Q507 Developers hesitate to publish details of their model because attackers might exploit the information. How should the organization BEST balance transparency?
- A. Publish all source code and internal configuration so independent researchers can review the model and its security controls in full
- B. Provide clear documentation of purpose, risk level and training data while protecting sensitive security details
- C. Publish nothing about the model to avoid any possible exposure
- D. Share model details only after a security incident has occurred
Answer: B
Transparency can be achieved through clear documentation about the model, data, bias and metrics, while security-sensitive details remain protected. Publishing everything or nothing, or waiting for an incident, does not balance trust and security.
Domain 1: AI Governance and Program Management
Q508 Which approach BEST makes an AI output explainable to a user?
- A. Hiding the sources used so that the vendor's data remains confidential
- B. Citing the sources used so users can verify how the result was reached
- C. Stating that the model is advanced and has been trained at scale
- D. Shortening each answer so users have less to question or verify
Answer: B
Citing sources lets users answer how the AI arrived at its result. Hiding sources, appealing to sophistication, or shortening output does not provide explainability.
Domain 1: AI Governance and Program Management
Q509 A CISO is asked how the AI security program should relate to existing security work. Which approach is MOST appropriate?
- A. Create a stand-alone AI security program that is unrelated to existing security controls, with its own tools, owners and reporting lines
- B. Adapt existing cybersecurity programs and add AI-centric practices such as security by design before investing in AI
- C. Limit AI security responsibility to the data science team only
- D. Defer AI security work until after the first major incident occurs
Answer: B
Adapting current programs and creating AI-centric practices should occur before investing in AI strategies. A disconnected program duplicates effort, waiting leaves the organization unprepared, and limiting scope to one team ignores enterprise-wide use.
Domain 1: AI Governance and Program Management
Q510 An AI lead is documenting how the enterprise has used AI tools over time. What is the PRIMARY value of this documentation?
- A. It is only needed to support marketing claims about responsible AI
- B. It substitutes for employee training on acceptable AI use
- C. It informs future decisions and lessons learned for later AI implementations
- D. It replaces the need for a formal AI risk assessment on future projects, because past results will be reused
Answer: C
Documenting the history of AI use informs future decisions and lessons learned. It does not replace risk assessment or training and has value beyond marketing.