AAIA Sample Questions & Answers

This statement is false. While statistical sampling allows for quantifiable measurement of sampling risk, it does not eliminate all bias. Furthermore, in AI auditing, judgmental (or risk-based) sampling is often necessary to target high-risk areas, such as testing the model's behavior on known edge cases or transactions involving sensitive data, which might be missed by a purely random statistical sample.

If the model predicts a failure but cannot explain why (e.g., which sensor reading or combination of factors led to the prediction), it is extremely difficult for maintenance engineers to perform targeted diagnostics and root cause analysis. They know the machine is likely to fail, but they don't know which component is the problem. This lack of insight hinders proactive repair and learning from potential failures.

A key risk indicator must be a measurable metric that provides an early warning of increasing risk. Measuring the statistical variance in pricing across defined demographic segments (e.g., based on geography, inferred income, etc.) directly targets the risk of price discrimination. A rising variance would trigger an investigation, making it the most effective KRI for this specific risk.

The 'high-risk' classification under the EU AI Act mandates a comprehensive set of obligations. The audit scope must therefore be significantly expanded to include conformity assessments against these legal requirements, which cover the entire lifecycle of the AI system, including the quality of training data, technical documentation, transparency for users, human oversight mechanisms, and a robust risk management system.

The Model Registry is the central component for ensuring integrity and reproducibility. It should version and store not just the trained model artifact, but also the code used for training, the version of the data it was trained on, its evaluation metrics, and its dependencies. This allows an auditor or data scientist to precisely recreate the model and its training environment, which is the essence of reproducibility.

The core issue is the uncontrolled and untested changes being pushed directly to production. This creates significant operational risk, as evidenced by the system's periodic failures. Implementing a formal change management process with a pre-production testing (staging) environment is the most critical control to ensure that updates are vetted for performance and stability before they can impact live operations and cause financial loss.

Substantive testing is designed to detect material misstatements in financial data. By independently recalculating the forecast (an analytical procedure), the auditor is gathering evidence about the validity and accuracy of the financial numbers produced by the AI model. This is a classic example of a substantive test.

The primary purpose of a Model Card or AI Bill of Materials is to provide clear, standardized documentation about an AI model's purpose, performance, limitations, and ethical considerations. This includes details about the training data, evaluation metrics, intended use cases, and potential biases. This practice directly supports the principles of transparency and accountability in AI governance.

Under regulations like GDPR, using personal data for a new purpose (like training an AI model) requires a valid legal basis, often specific and informed consent from the data subjects (the students).

The principle of data minimization requires that only data essential for the specific, stated purpose should be collected and processed. The university should not use extraneous student data to train the model.

Good data governance requires maintaining data lineage—a record of the data's origin, transformations, and usage. This is crucial for auditing the model's training process, ensuring data integrity, and investigating issues like bias.

A model inversion attack specifically aims to reverse-engineer the training data. By having access to the model and its outputs (e.g., class labels or confidence scores), an attacker can infer sensitive features or even reconstruct representative samples of the data used to train it. This is a severe privacy breach, especially for models trained on PII or medical images.