AI-901 Practice Test

Sample AI-901 Practice Questions

10 real questions from the free AI-901bank. Pick an answer in your head, then open “Show answer” for the explanation.

1. 1. An organization's AI Governance Committee is reviewing a proposed AI system for automated fraud detection. Which component of responsible AI governance should they require before approving the system for production deployment?

   • A. An AI impact assessment that documents the intended use, potential harms, mitigations, monitoring plan, and accountable stakeholders
   • B. A business case document demonstrating the expected return on investment and projected cost savings from automation, with comparable benchmarks from industry peers
   • C. A technical architecture review confirming that the model achieves a minimum accuracy threshold and that the training infrastructure meets enterprise security standards
   • D. A user acceptance testing report confirming that the fraud detection team finds the system's interface intuitive and that the false positive rate is acceptable for operational workflows
   Show answer

   Correct answer: A

   An AI impact assessment is the foundational accountability artifact before production deployment. It documents the system's intended use, who could be affected, potential harms and their mitigations, monitoring plans, and which stakeholders are accountable for what. This ensures responsible AI considerations are reviewed before launch. B is wrong because a business case captures financial value but does not cover responsible-AI risk. C is wrong because a technical architecture review is useful but does not by itself address impact on people. D is wrong because user acceptance testing measures usability and false positive rates but does not cover the broader responsible AI considerations the committee is reviewing.

2. 2. A hiring algorithm achieves equalized odds - equal true positive rates and equal false positive rates across male and female applicants. Despite this, women are hired at lower rates because fewer women are in the applicant pool. An advocate argues the system is still discriminatory. What tension does this scenario most accurately illustrate?

   • A. The advocate is incorrect because equalized odds mathematically guarantees equal treatment of equally qualified individuals, and differences in pool composition reflect societal factors beyond the algorithm's scope
   • B. Different fairness criteria can conflict: a system can be fair across individuals while still showing unequal group outcomes when the underlying groups are unequally represented
   • C. Equalized odds is a mathematically incomplete fairness metric that fails to detect real discrimination when group representation in the applicant pool reflects historical exclusion rather than current merit differences
   • D. The disparity in hiring rates reflects a societal problem entirely outside the scope of algorithmic fairness - the AI system is operating correctly according to established fairness metrics
   Show answer

   Correct answer: B

   This scenario illustrates that different conceptions of fairness can conflict. Individual fairness (such as equalized odds) ensures equally qualified individuals are treated the same regardless of group membership, while group fairness focuses on equal outcomes across groups. When the groups are unequally represented in the applicant pool, both can be reasonable concerns and can pull in different directions. A is wrong because dismissing the disparity as purely societal ignores the legitimate group-fairness concern. C is wrong because equalized odds is a real fairness metric; the issue is that it does not capture every conception of fairness, not that it is mathematically flawed. D is wrong because saying the algorithm is operating 'correctly' under one metric does not by itself answer the harder question of which fairness conception to prioritize.

3. 3. A data scientist uses the Model interpretability component of Azure Machine Learning's Responsible AI dashboard to investigate why a loan-approval model rejects certain applicants. According to Microsoft's documentation, what is this component specifically designed to help uncover?

   • A. Only aggregate accuracy and error-rate statistics for the whole dataset, with no per-feature detail
   • B. Which rows in the training dataset are duplicates, so they can be removed before retraining
   • C. Potential sources of unfairness, by revealing whether the model bases predictions on sensitive features or on features that are highly correlated with them (proxies)
   • D. The exact monetary cost of each misclassification, based on a configurable business rule
   Show answer

   Correct answer: C

   Microsoft's documentation states that model interpretability should be used to 'uncover potential sources of unfairness by understanding whether the model is basing predictions on sensitive features or on features that are highly correlated with them' (option C) - directly addressing proxy-variable concerns like a feature that stands in for a protected or sensitive attribute. A understates the component's per-feature global/local explanation capability. B and D describe capabilities the component doesn't provide.

4. 4. What is algorithmic bias in the context of AI systems?

   • A. Systematic, unfair discrimination by an AI system against individuals or groups based on characteristics such as race, gender, or age, often rooted in biased training data or design
   • B. A mathematical error in the model's optimization function that causes the model to reach statistically incorrect conclusions during inference
   • C. A hardware limitation in GPU processing that causes AI models to perform differently across different computational environments
   • D. A security vulnerability that allows malicious actors to inject false data into AI training pipelines, altering model behavior
   Show answer

   Correct answer: A

   Algorithmic bias is when an AI system produces systematically unfair results that favor or disadvantage certain groups, typically because of biased training data, feature choices, or feedback loops that amplify existing inequalities. Because automated systems can scale decisions, algorithmic bias is a core fairness concern. B is wrong because a simple math error in the optimizer is not the standard meaning of algorithmic bias. C is wrong because hardware variation across environments is not what 'bias' refers to in this responsible AI context. D is wrong because that describes a security/data-poisoning concern, not algorithmic bias.

5. 5. A healthcare organization wants to train a machine learning model using patient data from multiple hospitals without patient data ever leaving each hospital's premises, due to strict privacy regulations. Which privacy-preserving technique best addresses this constraint?

   • A. Differential privacy, where statistical noise is added to each hospital's data so that individual patient records cannot be reconstructed from the aggregated training dataset
   • B. Federated learning, where each hospital trains a local model on its own data and only shares model updates - not raw patient data - with a central aggregator
   • C. Data anonymization, where patient identifiers are removed and records are generalized before being combined in a central repository for joint model training
   • D. Synthetic data generation, where each hospital creates AI-generated fake patient records that statistically mirror real patient distributions without containing actual patient information
   Show answer

   Correct answer: B

   Federated learning directly addresses the constraint that raw patient data cannot leave each hospital. Each hospital trains a local model on its own data and only shares model updates (such as gradients or weights) with a central aggregator. This allows organizations to collaboratively improve a shared model while each retains custody of its sensitive data. A is wrong because differential privacy adds noise to data or queries but does not by itself keep raw data inside each hospital. C is wrong because anonymization combined with central pooling still requires moving the data to a central repository, which violates the constraint. D is wrong because synthetic data still requires the synthetic records to be centralized and may not preserve fine-grained collaborative training signal.