AI0-001 Practice Test 25 — 15 Questions

Question 1

A company is implementing an AI ethics board. Which TWO responsibilities should the board typically have?

Accepted Answer

Approving or rejecting high-risk AI initiatives. An AI ethics board should oversee AI projects for ethical compliance and approve high-risk AI initiatives. Implementing technical models and auditing data privacy are operational tasks not typically under the board's direct purview.

Answer

Writing code for fairness algorithms

Answer

Developing marketing strategies for AI products

Answer

Conducting daily data privacy audits

Question 2

Which NIST AI RMF function involves identifying the context, risks, and potential impacts of an AI system, including mapping the AI lifecycle and stakeholders?

Accepted Answer

Map. The AI RMF's four functions are: Govern, Map, Measure, Manage. Map focuses on context and risk identification. Govern sets policies. Measure evaluates metrics. Manage addresses risks through controls.

Answer

Manage

Answer

Measure

Answer

Govern

Question 3

A company develops an AI model that recommends job candidates. The model inadvertently discriminates against a protected group. Which approach is most effective for mitigating this bias?

Accepted Answer

Use a fairness-aware machine learning algorithm. Fairness-aware machine learning algorithms explicitly incorporate fairness constraints or objectives during model training, directly addressing and mitigating bias against protected groups. Unlike simple removal of protected attributes, these algorithms can detect and correct for proxy discrimination and disparate impact, ensuring the model's recommendations are equitable by design.

Answer

Remove the protected attribute from the training data

Answer

Analyze model predictions after deployment

Answer

Collect more training data from the protected group

Question 4

A company is deploying an AI model that processes financial transactions. They want to implement privacy-preserving machine learning. Which THREE techniques achieve this goal? (Select three.)

Accepted Answer

Differential privacy. Differential privacy (B) is correct because it adds calibrated noise (e.g., via the Laplace or Gaussian mechanism) to computations or gradients so that any single individual's transaction data has a bounded influence on the model output, providing a formal privacy guarantee. Homomorphic encryption (D) is correct because it allows computations to be performed directly on encrypted financial data (e.g., using schemes like Paillier, BFV, or CKKS), so the model can train or infer without ever decrypting sensitive values. Federated learning (E) is correct because it keeps raw transaction data on local devices or silos and only shares model updates (often combined with secure aggregation or differential privacy), minimizing centralized exposure of private records. Model pruning (A) merely removes redundant weights to reduce model size and compute, and data augmentation (C) synthetically expands training data for robustness; neither provides a privacy guarantee, so they do not belong.

Answer

Model pruning

Answer

Data augmentation

Question 5

A healthcare AI system uses patient data to predict disease risk. To comply with privacy regulations, the organization wants to ensure that the model cannot reveal whether a specific patient's data was used in training. Which technique should they implement?

Accepted Answer

Differential privacy. Differential privacy is the correct technique because it adds calibrated statistical noise (e.g., via the Laplace or Gaussian mechanism) to query results or gradients so that the inclusion or exclusion of any single patient's record produces a nearly indistinguishable output. This provides a formal, mathematically provable guarantee against membership inference, which is exactly the requirement stated. Homomorphic encryption, federated learning, and model validation address different concerns (computation on encrypted data, decentralized training, and general model quality) and do not by themselves prevent membership disclosure.

Answer

Homomorphic encryption

Answer

Federated learning

Answer

Model validation

Question 6

A company is deploying a code generation AI assistant for internal developers. They want to ensure the assistant does not generate code with security vulnerabilities. Which testing approach is MOST critical?

Accepted Answer

Integration tests that send security-focused prompts and validate the generated code against a static analysis tool. The goal is to ensure the assistant does not generate vulnerable code, so the most critical testing is security-focused: send adversarial security prompts and validate the generated code with a static analysis tool (SAST). This directly measures whether the model produces exploitable code and provides actionable feedback, unlike generic quality metrics.

Answer

Unit tests for the data pipeline that preprocesses prompts

Answer

Evaluation framework that measures BLEU score on a held-out set of code samples

Answer

Regression tests that compare outputs of new model versions against a golden dataset

Question 7

A developer is building a RAG system and needs to choose a similarity metric for retrieving document chunks. The embedding model they use produces normalized vectors (unit vectors). Which similarity metric is equivalent to cosine similarity in this case?

Accepted Answer

Dot product. For normalized (unit-length) vectors, the dot product equals the cosine of the angle between them, because the magnitudes are both 1. Cosine similarity is defined as the dot product divided by the product of magnitudes; when magnitudes are 1, the denominator is 1, so dot product and cosine similarity are mathematically identical. This makes dot product the correct equivalent metric.

Answer

Jaccard similarity

Answer

Euclidean distance

Answer

Manhattan distance

Question 8

A financial institution uses a machine learning model to approve loans. They want to protect against membership inference attacks. Which THREE techniques are effective?

Accepted Answer

Applying model truncation or output perturbation. Option A (model truncation or output perturbation) is correct because reducing the precision or adding calibrated noise to the model's outputs limits the information an attacker can extract about whether a specific record was in the training set, directly mitigating membership inference. Option B (training with differential privacy) is correct because DP-SGD and related mechanisms provide a formal guarantee that the inclusion or exclusion of any single training record has a bounded effect on the model's behavior, which is the canonical defense against membership inference. Option C (limiting the granularity of model outputs, e.g., returning scores instead of probabilities) is correct because coarse, bucketed outputs reduce the signal an attacker can use to distinguish members from non-members, lowering attack success rates. Option D (federated learning) is not inherently a membership-inference defense: it keeps raw data local but the shared model updates can still leak membership information, so it does not by itself provide the required protection. Option E (shadow models to distract attackers) is not a recognized defense; shadow models are an attacker technique used to train attack classifiers, not a mitigation, so it does not belong here.

Answer

Implementing federated learning

Answer

Using shadow models to distract attackers

Question 9

A data scientist is training a resume screening model to rank job applicants. The training data includes historical hiring decisions from the past 10 years. The company wants to avoid unfair bias against underrepresented groups. Which type of bias is most likely present in the training data?

Accepted Answer

Historical bias. Historical bias occurs when training data reflects past human decisions or societal patterns that were themselves biased, causing the model to learn and perpetuate those patterns. A resume screening model trained on 10 years of hiring decisions will inherit whatever demographic skews existed in those decisions.

Answer

Algorithmic bias

Answer

Selection bias

Answer

Confirmation bias

Question 10

A financial institution is developing a fraud detection model using historical transaction data. The dataset contains over 10 million records, but only 0.01% of transactions are fraudulent. The current model uses a neural network trained with standard cross-entropy loss, and the team applies random undersampling of the majority class to create a balanced training set. However, the model still produces a high number of false positives (legitimate transactions flagged as fraud) and misses approximately 30% of actual fraud cases. The business requires that at least 95% of frauds be caught, and the false positive rate must be below 1% to avoid overwhelming fraud analysts. The team has limited resources to collect additional data and cannot change the model architecture significantly. Which approach should the team take to best meet the business requirements?

Accepted Answer

Use cost-sensitive learning by assigning a higher misclassification cost to the fraud class.. Cost-sensitive learning directly addresses the business requirement by penalizing fraud misclassifications more heavily, which shifts the decision threshold to favor recall on the fraud class while still allowing the team to tune the trade-off between false positives and false negatives. Because the team cannot collect more data or change the architecture significantly, adjusting the loss function's class weights is the most practical lever to hit the 95% recall and <1% FPR targets. It also avoids the information loss caused by random undersampling, which discards 99.99% of legitimate transactions and distorts the true class distribution.

Answer

Apply feature selection to remove noisy predictors and then retrain the current model.

Answer

Switch to an anomaly detection algorithm such as Isolation Forest or One-Class SVM.

Answer

Collect more transaction data, especially fraudulent examples, to naturally balance the classes.

Question 11

A security analyst discovers that an attacker has been querying a production LLM API with thousands of carefully crafted prompts and using the responses to build a local copy of the model. Which attack is occurring?

Accepted Answer

Model extraction. Model extraction (also called model stealing) occurs when an attacker queries a deployed model API repeatedly with crafted inputs and uses the input-output pairs to train a surrogate model that approximates the target's behavior. The scenario—thousands of queries used to build a local copy—is the textbook definition. The goal is to replicate the model's functionality, often to avoid API costs or to enable further attacks like adversarial example transfer.

Answer

Prompt injection

Answer

Data poisoning

Answer

Membership inference

Question 12

An AI engineer is fine-tuning a transformer-based language model for a domain-specific task. They want to improve the model's factual accuracy and reduce hallucinations. Which THREE strategies should they consider? (Select THREE)

Accepted Answer

Fine-tune the model on a curated domain-specific corpus. Option B is correct because fine-tuning on a curated, high-quality domain-specific corpus injects accurate, task-relevant knowledge into the model's weights, which directly improves factual grounding and reduces the likelihood of fabricating domain facts. Option D is correct because chain-of-thought prompting encourages the model to decompose complex queries into intermediate reasoning steps, which empirically reduces errors and unsupported claims on multi-step or reasoning-heavy tasks. Option E is correct because Retrieval-Augmented Generation grounds generation in externally retrieved, up-to-date documents at inference time, so the model conditions its output on verifiable evidence rather than relying solely on parametric memory, which is the standard technique for reducing hallucinations. Option A is not appropriate because simply increasing the context window beyond the training limit is not a supported or effective strategy—models cannot attend beyond their trained positional range without architectural changes or extrapolation techniques, and a larger window alone does not improve factual accuracy. Option C is incorrect because raising the temperature increases sampling randomness and diversity, which typically increases hallucination rather than reducing it; lower temperature is preferred for factual accuracy.

Answer

Increase the model's context window size beyond the training limit

Answer

Use a higher temperature setting during generation

Question 13

A healthcare startup needs to deploy an AI model for real-time patient monitoring on IoT devices with limited battery and compute. The model must run locally with minimal latency. Which TWO strategies are most appropriate?

Accepted Answer

Use TensorFlow Lite to convert and run the model on the device. TensorFlow Lite is specifically designed to run TensorFlow models on resource-constrained edge devices like IoT sensors. It optimizes the model for low latency inference by using a specialized interpreter and hardware acceleration delegates (e.g., NNAPI, GPU), enabling real-time patient monitoring without cloud dependency.

Answer

Apply model distillation to create a smaller student model

Answer

Deploy the model on a cloud server and stream data

Answer

Use ONNX Runtime with a GPU backend

Question 14

A company is deploying an AI system that screens job applications. According to the EU AI Act, this system is likely classified as high-risk because it affects employment opportunities. Which requirement must the company implement for high-risk AI systems?

Accepted Answer

A human-in-the-loop mechanism that enables override of the AI's decisions. Under the EU AI Act, high-risk AI systems must implement appropriate human oversight measures, including human-in-the-loop mechanisms that allow humans to override or intervene in the AI's decisions. This is a key requirement to ensure accountability and prevent harm in critical areas like employment.

Answer

Full transparency by publishing the model's source code and training data

Answer

Annual third-party audits of the model's energy consumption

Answer

Obtaining explicit consent from each applicant to process their data

Question 15

An organisation is deploying an AI system for credit scoring, which is considered high-risk under the EU AI Act. Which requirement is NOT typically mandated for high-risk systems?

Accepted Answer

Publish the complete source code of the AI system. The EU AI Act mandates several requirements for high-risk AI systems, including risk management, data governance (relevant and representative training data), technical documentation, transparency, human oversight, and accuracy/robustness. Publishing the complete source code is not a typical requirement; the Act focuses on transparency and documentation, not open-sourcing proprietary code. Therefore, option B is the requirement that is NOT typically mandated.

Answer

Ensure training data is relevant and representative

Answer

Establish a risk management system

Answer

Provide human oversight mechanisms