AI0-001 Practice Exam 63 — 20 Questions

Question 1

A hospital deploys an LLM assistant that answers clinician questions using a retrieval-augmented generation pipeline over internal patient records. Administrators worry that a malicious document placed in the retrieval index could hijack the assistant's behavior. Which control directly mitigates this indirect prompt injection risk?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Sanitize and validate retrieved content, and isolate it from system instructions in the prompt structure.. Indirect prompt injection exploits the model's tendency to follow instructions embedded in retrieved content. Treating retrieved documents strictly as data, sanitizing them, and separating them from system-level instructions removes the channel the attacker relies on. The other options affect model randomness, domain adaptation, or storage security, none of which prevent injected text from being interpreted as commands.

Answer

Fine-tune the LLM on a corpus of approved clinical question-and-answer pairs.

Answer

Enable encryption of the retrieval index at rest and rotate its access keys.

Answer

Increase the model's temperature setting so responses are less deterministic.

Question 2

A media company is deploying an AI system that generates short news summaries from full articles. Before launch, the responsible AI review board asks the team to define monitoring that will detect harmful or degraded behavior in production. Which TWO monitoring practices should the team implement? (Choose two.)

Accepted Answer

Track hallucination and factual-consistency rates by automatically comparing generated summaries against source articles using a grounded entailment check.. Responsible deployment of a generative summarization system requires monitoring both automated factual grounding and structured human review. Grounding checks quantify hallucinations against source articles, while rubric-based human sampling catches bias, tone, and framing issues that metrics miss. Together they provide the safety and quality signals the review board needs. Infrastructure and length metrics do not measure harmful or degraded content behavior.

Answer

Record the GPU utilization of the inference cluster to confirm the model is running within expected compute bounds.

Answer

Monitor the average token length of generated summaries to ensure outputs stay within the expected range.

Answer

Track the number of API calls per minute to detect traffic spikes that could indicate abuse or system overload.

Question 3

An AI operations team supports a model that scores insurance claims in real time. They need to detect when the live input distribution diverges from the training distribution and alert before claim decisions degrade. Which approach should they implement?

Accepted Answer

Continuously compare live feature distributions against the training baseline using drift metrics such as population stability index, and alert when thresholds are exceeded.. Detecting divergence between live and training input distributions requires continuous statistical comparison of feature distributions, which population stability index and related drift metrics provide. This catches change early and pinpoints the drifting features. Output-mean monitoring lags, quarterly sampling is too slow, and unconditional nightly retraining reacts to noise rather than measured drift.

Answer

Schedule a quarterly manual review of a random sample of claims and adjust the model if reviewers notice problems.

Answer

Monitor only the model's average prediction value and alert when it changes by more than a fixed percentage.

Answer

Retrain the model nightly on the most recent claims regardless of any measured change in the data.

Question 4

A hospital deploys an AI model that summarizes clinical notes for physicians. Before go-live, the AI team must verify that the model does not reproduce patient identifiers in its summaries when they are not clinically necessary. Which activity is the MOST appropriate for this verification?

Accepted Answer

Run a red-team evaluation with prompts designed to elicit protected health information and measure leakage rates.. Verifying that a summarizer suppresses unnecessary identifiers requires exercising the model with prompts that attempt to elicit protected health information and measuring how often leakage occurs. Red-teaming produces that empirical evidence against a threshold. Perplexity, training-data scrubbing, and model cards are useful complements but none demonstrates the deployed model's output behavior under adversarial or realistic clinical prompts.

Answer

Verify that the training dataset was de-identified using an automated named-entity recognition scrubber.

Answer

Confirm the model card documents the intended use and known limitations of the summarization system.

Answer

Compare the model's perplexity on the clinical corpus against its perplexity on public text.

Question 5

A media company uses a generative AI service to draft marketing copy. Legal asks the AI governance team to reduce the risk that outputs reproduce copyrighted passages from the training corpus. Which control most directly addresses that specific risk?

Accepted Answer

Enable content-provenance metadata and output filtering that flags or blocks near-verbatim matches to known copyrighted text.. The risk is that generated marketing copy could contain near-verbatim copyrighted material. A control that inspects output for long matches against reference corpora and blocks or flags them, paired with provenance metadata, directly reduces that exposure and leaves an audit trail. Training, temperature tuning, and logging influence behavior or records but do not detect or stop a reproducing output before publication.

Answer

Log every prompt and response for ninety days and store the logs in the governance archive.

Answer

Require all marketing staff to complete annual training on the organization's acceptable-use policy.

Answer

Lower the model's temperature setting so generated text is more deterministic.

Question 6

An ML engineer is tuning a random forest classifier for a medical diagnosis task and observes that training accuracy is 99% while validation accuracy is 78%. She wants to reduce the gap without discarding the ensemble approach. Which change is most likely to reduce the generalization gap?

Accepted Answer

Reduce the maximum tree depth and increase the minimum number of samples required at each leaf.. The large train-validation gap indicates that the individual trees are too complex and memorizing the training set. Constraining tree depth and requiring more samples per leaf regularizes each estimator. Because a random forest averages many decorrelated trees, this regularization typically improves validation accuracy while keeping the ensemble's variance-reduction benefits intact.

Answer

Add more trees to the ensemble until the out-of-bag error stops decreasing.

Answer

Increase the maximum tree depth so each tree can capture more interactions in the training data.

Answer

Disable bootstrap sampling so every tree is trained on the full dataset instead of a random subset.

Question 7

A health-tech firm is preparing a model card for a clinical decision support tool that flags patients at risk of sepsis. The compliance team asks which element of the model card is MOST directly relevant to documenting the system's intended use and out-of-scope applications. Which section should the team prioritize?

Accepted Answer

The 'Intended Use' section, which specifies the clinical populations, care settings, and decision boundaries for which the model was validated.. A model card's intended use section is the formal declaration of the validated operating envelope: the populations, settings, and decision types the model supports, plus explicitly excluded uses. Compliance reviewers rely on it to confirm that a proposed deployment falls inside the validated scope. Training data, performance metrics, and ethical considerations are supporting evidence but do not define the boundary of appropriate use.

Answer

The 'Performance Metrics' section, which reports AUROC, sensitivity, and specificity at the deployed decision threshold.

Answer

The 'Training Data' section, which lists the source datasets, their collection dates, and demographic composition.

Answer

The 'Ethical Considerations' section, which discusses potential harms and mitigation strategies identified during review.

Question 8

A financial services company is deploying a credit-scoring model built with the AI+ toolkit. The model must produce an explanation for each decision that regulators can review, showing which input features most influenced the score. The data science team has already trained a gradient-boosted tree ensemble. Which approach should the team use to satisfy the regulatory requirement?

Accepted Answer

Apply SHAP (SHapley Additive exPlanations) values to the trained ensemble to generate per-instance feature attributions.. Per-decision explanations require local, instance-level attribution methods. SHAP values quantify how much each feature pushed a single prediction above or below the baseline, and TreeSHAP makes this tractable for tree ensembles. Global importance, model substitution, and raw input-output disclosure all fail to explain why a specific applicant received a specific score.

Answer

Publish the model's raw prediction probabilities alongside the input feature values for each applicant.

Answer

Replace the ensemble with a single decision tree limited to a depth of three so that the decision path is human-readable.

Answer

Compute global feature importance by averaging the ensemble's split gains across the entire training set.

Question 9

A financial firm trained a gradient boosting model on two years of loan data. It reported strong AUC during development, but after six months in production, approval rates for a newly launched loan product diverge sharply from expectations. The data science lead suspects the model is stale. Which approach best addresses this deployment issue?

Accepted Answer

Monitor prediction distributions and retrain on recent labeled data when drift is detected. The symptoms describe drift: a new product changes the applicant distribution and recent economic conditions change the relationship between features and default. The durable fix is to monitor for drift and retrain on recent labeled data, then revalidate before deployment. Changing hyperparameters, adding boosting rounds, or swapping metrics leaves the model trained on outdated patterns and cannot restore alignment with current production behavior.

Answer

Retrain the model on the original two-year dataset with a lower learning rate

Answer

Switch the evaluation metric from AUC to accuracy without changing the data

Answer

Increase the number of boosting rounds to improve fit on the existing data

Question 10

A junior data scientist is asked to explain the difference between supervised and unsupervised learning to a product manager. She wants to give a single concrete example that clearly illustrates unsupervised learning. Which example should she choose?

Accepted Answer

Grouping retail customers into segments based on purchasing behavior using k-means clustering.. Unsupervised learning finds structure in data that has no labels. Clustering customers by purchasing behavior is a textbook example because k-means discovers groups from feature similarity alone. The other choices involve a known target (default status, sale price) or a reward signal (game playing), which place them in supervised or reinforcement learning rather than unsupervised learning.

Answer

Predicting whether a loan applicant will default using historical labeled outcomes.

Answer

Training an agent to play a board game by rewarding winning moves over many episodes.

Answer

Estimating a house's sale price from its square footage and location features.

Question 11

A logistics company runs a route-optimization model on a Kubernetes cluster. During peak hours the inference pods are frequently evicted because the nodes run out of memory, even though average GPU utilization stays below 40 percent. The team wants to reduce evictions without changing the model or adding nodes. Which action best addresses the root cause?

Accepted Answer

Set explicit CPU and memory requests and limits on the inference containers so the scheduler can place and protect them accurately.. Memory-pressure eviction happens when the kubelet must reclaim node memory, and pods without declared memory requests are the first to be sacrificed because the scheduler never reserved capacity for them. Setting realistic requests and limits makes placement accurate and bounds each pod's consumption, so the node stays below its eviction threshold without changing the model or adding hardware.

Answer

Lower the container image size by switching to a distroless base image for the inference service.

Answer

Enable horizontal pod autoscaling on GPU utilization so additional replicas start during peak hours.

Answer

Increase the GPU memory allocation per pod by requesting additional nvidia.com/gpu resources.

Question 12

A hospital's AI triage model was trained on five years of historical admissions. A governance review finds that patients over 75 are systematically assigned lower acuity scores than clinically equivalent younger patients, even though age is not an input feature. Which governance control most directly addresses this finding?

Accepted Answer

Run a proxy-variable and disparate-impact analysis on the model's outputs before each deployment.. The disparity appears despite age being excluded, which is the classic signature of a proxy variable carrying protected-class information into the model. Testing outcomes for disparate impact across age cohorts, and tracing which features correlate with age, is the control that both detects and documents the harm. Data volume, cryptography, and shadow comparison each address different concerns and leave the indirect bias unmeasured.

Answer

Deploy the model in shadow mode and compare its predictions against clinician judgment for one week.

Answer

Increase the volume of historical training data by adding ten more years of admissions records.

Answer

Encrypt the training dataset at rest and restrict access to the model registry.

Question 13

A computer vision team trains a convolutional neural network for manufacturing defect detection on a workstation with an NVIDIA RTX A6000 GPU. They want to reduce training time by increasing throughput without changing model architecture or batch size. Which action should they take?

Accepted Answer

Enable mixed-precision training using NVIDIA Tensor Cores with FP16 compute and FP32 accumulation.. Mixed-precision training exploits Tensor Cores to compute in FP16 while accumulating in FP32, delivering substantial throughput gains on modern NVIDIA GPUs with minimal accuracy impact. Because the architecture and batch size remain unchanged, it fits the constraint precisely. Other listed actions either alter training semantics or move work to slower hardware, so they do not satisfy the goal.

Answer

Increase the batch size to the maximum that fits in GPU memory and keep the learning rate unchanged.

Answer

Shard the dataset across multiple CPU cores using NumPy and disable GPU acceleration for the convolution layers.

Answer

Convert the trained model to TensorFlow Lite and redeploy it for training on the GPU.

Question 14

An operations team runs a computer-vision model that flags manufacturing defects on an assembly line. Auditors require evidence that any single prediction can be reconstructed and explained months later. The team already logs model version, input image hash, and prediction score. Which additional logging practice best satisfies the audit requirement?

Accepted Answer

Log the preprocessed feature vector, the model version identifier, and the explanation artifacts such as saliency or SHAP values for each inference.. Auditability of an individual prediction requires the exact inputs the model saw, the version of the model that scored them, and the explanation artifacts generated at inference time. Capturing the preprocessed feature vector, model version identifier, and attributions such as SHAP or saliency values lets auditors replay and justify any single decision. Aggregates, raw frames, or reviewer names cannot reconstruct the original reasoning.

Answer

Log the prediction score together with the operator who reviewed the flagged unit.

Answer

Log the raw camera frames indefinitely and rely on the current model to regenerate explanations on demand.

Answer

Log only the aggregate daily defect rate and the model's overall precision.

Question 15

A financial services firm runs an AI model that scores loan applications. Regulators require the firm to explain any adverse decision to an applicant. The model is a gradient-boosted tree with hundreds of features. Which implementation approach best satisfies the explainability requirement without replacing the model?

Accepted Answer

Generate per-applicant explanations using SHAP values that attribute the decision to the most influential features.. Adverse action explanations must be specific to the individual applicant, and SHAP values provide locally accurate attributions for each prediction. This preserves the gradient-boosted tree's performance while producing reasons that can be communicated and audited. Global importance describes the population, logistic regression replaces the model, and raw logs record inputs without explaining the decision.

Answer

Replace the gradient-boosted tree with a logistic regression so coefficients are directly interpretable.

Answer

Publish the model's global feature importance ranking so applicants can see which factors matter most overall.

Answer

Log the model's raw input features and output score for each application so auditors can review the data.

Question 16

A hospital's clinical decision support model was validated at 94 percent accuracy on a held-out set. After go-live, clinicians report that the model's suggestions are frequently irrelevant for elderly patients, even though overall accuracy in the monitoring dashboard has barely moved. Which monitoring practice would have surfaced this problem?

Accepted Answer

Tracking performance metrics disaggregated by patient demographic and clinical subgroups, with alerts when any subgroup falls below its own threshold.. The clue is that aggregate accuracy held steady while a specific patient group received poor suggestions, which is the classic signature of a subgroup masked by population averaging. Disaggregated performance monitoring with per-segment thresholds is the practice that reveals it. Aggregate accuracy alerts, latency metrics, and input drift detection each observe a different dimension and cannot confirm that elderly patients specifically are being served badly.

Answer

Measuring inference latency and throughput to confirm the model responds within clinical workflow time limits.

Answer

Monitoring overall model accuracy and alerting when it drops more than two percentage points from the validation baseline.

Answer

Comparing the live distribution of input features against the training distribution to detect covariate shift.

Question 17

A retail company runs an AI-powered demand forecasting service in a Kubernetes cluster. The inference pods scale based on CPU utilization, but during flash sales the request queue grows rapidly and p99 latency spikes before new pods become ready. The operations team needs to reduce latency during these spikes without changing the model itself. Which action should the team take?

Accepted Answer

Configure a Horizontal Pod Autoscaler (HPA) based on a custom external metric that reflects request queue depth, and lower the scale-up stabilization window.. The root cause is that CPU-based autoscaling reacts too slowly to a sudden queue buildup. Using a custom metric tied to queue depth or concurrent requests, combined with a shorter scale-up stabilization window, lets the HPA add capacity before latency degrades. The other options either add latency, waste resources, or do not improve reaction speed.

Answer

Enable a Kubernetes PodDisruptionBudget for the inference deployment and increase the replica count permanently.

Answer

Set the Kubernetes resource requests and limits for the inference pods to the maximum available node capacity.

Answer

Increase the model's batch size and enable request batching to improve throughput per pod.

Question 18

An MLOps engineer is deploying a scikit-learn random forest model to a Kubernetes cluster for a low-traffic internal API. The team wants to avoid maintaining a custom Flask wrapper and prefers a standard serving solution that supports REST and gRPC. Which serving component should they choose?

Accepted Answer

KServe with a scikit-learn model server runtime.. KServe offers a purpose-built scikit-learn runtime that serves pickled models over REST and gRPC on Kubernetes, with autoscaling and rollout features included. It eliminates the need for a bespoke Flask wrapper, which the team explicitly wants to avoid. The other options either require custom code, target GPU deep learning workloads, or demand unsupported format conversion.

Answer

NVIDIA Triton Inference Server with a Python backend script that loads the pickle file.

Answer

TensorFlow Serving configured with a SavedModel export of the random forest.

Answer

A Kubernetes Deployment running a Flask app that loads the model and exposes only a REST endpoint.

Question 19

A natural language processing team is building a system to classify support tickets into categories. They have a large corpus of unlabeled ticket text and a small set of manually labeled tickets. They want to leverage both to improve classification performance. Which approach is MOST suitable?

Accepted Answer

Use semi-supervised learning that combines the unlabeled corpus with the labeled set. Semi-supervised learning is the best fit because it explicitly uses both the large unlabeled corpus and the small labeled set. Methods like self-training or consistency regularization can improve classification by learning from unlabeled tickets while still respecting the predefined categories. The other options either ignore one of the data sources or fail to produce a classifier aligned with the desired categories.

Answer

Apply unsupervised clustering to the unlabeled tickets and treat clusters as categories

Answer

Train a supervised classifier only on the small labeled set

Answer

Perform principal component analysis on the ticket text and train a classifier on the components

Question 20

A startup is developing an AI chatbot and wants to use a pre-trained language model to generate responses. They need to integrate the model into their application with minimal latency and cost. Which approach should they take?

Accepted Answer

Use a pre-trained model via a managed API service. The startup needs to integrate a pre-trained language model with minimal latency and cost. Using a managed API service provides immediate access to state-of-the-art models without infrastructure management, and the pay-as-you-go model keeps costs low. This is the most efficient approach for a startup with limited resources.

Answer

Use a rule-based chatbot system

Answer

Train a custom transformer model from scratch on their own data

Answer

Deploy the model on a local server with a single GPU