View Full Amazon AWS Certified AI Practitioner AIF-C01 Exam Dumps and Practice Test Dumps.

Question 341

Which metric measures the proportion of actual negative cases that a classification model correctly identifies?

1. Specificity
2. Precision
3. Recall
4. Accuracy

Correct Answer: 1

Explanation

Specificity measures how effectively a classification model identifies negative cases correctly. It is calculated as true negatives divided by the total number of actual negative cases, including both true negatives and false positives. A high specificity means the model produces relatively few false-positive results. This metric can be particularly important when incorrectly labeling a negative case as positive has meaningful consequences. Specificity is different from precision, which focuses on the correctness of positive predictions, and recall, which measures how many actual positive cases are detected.

Question 342

A team is evaluating a generative AI model that produces summaries of customer-support conversations. Which evaluation approach is most appropriate when exact wording may vary while the meaning should remain consistent?

1. Count only whether every generated word exactly matches the reference summary.
2. Compare the generated summary with reference summaries using semantic similarity.
3. Measure only the number of input tokens processed by the model.
4. Determine whether the model uses the same internal parameters as the reference model.

Correct Answer: 2

Explanation

Semantic similarity is useful when multiple different wordings can correctly express the same meaning. Generative AI systems frequently produce valid outputs that do not exactly match a reference answer word for word. A semantic similarity evaluation can therefore assess whether the generated summary conveys substantially similar information to the expected content. Exact-match evaluation can be too restrictive for such tasks. Token counts and internal model parameters do not directly measure whether the generated summary preserves the intended meaning of the customer-support conversation.

Question 343

Which AWS service can help an organization discover sensitive information stored in Amazon S3 by analyzing the data for potential exposure?

1. AWS Artifact
2. AWS Config
3. Amazon Macie
4. AWS Budgets

Correct Answer: 3

Explanation

Amazon Macie is designed to help organizations discover and protect sensitive data stored in Amazon S3. It can use machine learning and pattern matching to identify potentially sensitive information, including personally identifiable information. This capability can support security and governance activities for AI workloads because AI applications may process large amounts of information from S3. AWS Artifact provides compliance documents, AWS Config evaluates resource configurations, and AWS Budgets helps monitor spending. Macie is therefore the service specifically associated with discovering sensitive data in S3.

Question 344

A company uses an AI model to classify loan applications. The business wants to reduce false-positive classifications even if this causes the model to miss some positive cases. Which approach can directly address this requirement?

1. Increase the number of training epochs without evaluating the model.
2. Remove all negative examples from the evaluation dataset.
3. Optimize only for recall.
4. Adjust the classification decision threshold to favor fewer positive predictions.

Correct Answer: 4

Explanation

The classification threshold determines how readily a model converts a predicted score into a positive classification. Raising the threshold can make the model more conservative about assigning the positive class, which may reduce false positives while potentially increasing false negatives. The appropriate threshold depends on the business requirements and should be selected using representative validation or evaluation data. Simply increasing training epochs does not directly control false positives. Removing negative examples would distort evaluation, while optimizing only for recall could increase the number of positive predictions.

Question 345

What does the area under the ROC curve (AUC) generally indicate?

1. How well a classifier separates positive and negative classes across thresholds
2. The number of parameters contained in the model
3. The amount of storage required for model artifacts
4. The average number of tokens in each prompt

Correct Answer: 1

Explanation

The area under the receiver operating characteristic curve, commonly called ROC AUC, provides a threshold-independent measure of how well a binary classifier distinguishes between positive and negative classes. A higher AUC generally indicates stronger class-separation ability across different classification thresholds. It is different from accuracy, which depends on a particular decision threshold and can be misleading with imbalanced datasets. AUC does not describe model size, storage requirements, or prompt length. It is primarily an evaluation measure for classification discrimination.

Question 346

An organization wants to keep track of changes made to prompts used by several production applications. Which practice provides the clearest way to reproduce a previous prompt configuration?

1. Allow each developer to edit prompts independently without records.
2. Version and document prompts together with their associated configurations.
3. Delete older prompts whenever a new version is created.
4. Change prompts directly in production without testing.

Correct Answer: 2

Explanation

Versioning prompts allows an organization to identify exactly which prompt configuration was used at a particular point in time. When prompt text, model settings, or related configuration changes are documented and versioned, teams can reproduce previous behavior, compare results, and investigate unexpected changes. This is especially useful when several applications share prompt templates. Deleting previous versions removes valuable traceability, while uncontrolled production changes make it difficult to determine why model behavior changed. Prompt versioning therefore supports reproducibility, testing, governance, and controlled improvement.

Question 347

Which capability is associated with Amazon Transcribe?

1. Generating synthetic images from text
2. Converting written text into speech
3. Converting speech into text
4. Creating vector embeddings from documents

Correct Answer: 3

Explanation

Amazon Transcribe is an AWS speech-to-text service that converts spoken audio into written text. It can be used for applications such as meeting transcription, call-center analysis, captions, and searchable audio content. This capability can also provide text that is subsequently processed by other AI services. Amazon Polly performs text-to-speech conversion, while image-generation services address visual content. Vector embeddings are produced through embedding models rather than being the primary purpose of Amazon Transcribe. Therefore, converting spoken language into text is the relevant capability.

Question 348

An AI application uses an external tool to send emails automatically. What is an important security measure before allowing the model to invoke that tool?

1. Give the model unrestricted access to all available tools.
2. Allow every generated request to execute without inspection.
3. Remove authentication requirements from the email system.
4. Restrict the model to an explicit allowlist of permitted actions and validate tool inputs.

Correct Answer: 4

Explanation

AI agents that can invoke external tools should operate with tightly controlled permissions. An explicit allowlist can limit the actions the agent is authorized to perform, while input validation can prevent malformed or unsafe parameters from reaching the external system. This reduces the potential impact of unexpected model behavior, prompt manipulation, or malicious input. Giving an agent unrestricted tool access increases risk, and removing authentication weakens security. Tool execution should therefore be constrained by application-level authorization, validation, and appropriate access controls.

Question 349

What is label noise?

1. Incorrect or inconsistent labels assigned to training examples
2. Excessive network latency during inference
3. Duplicate model versions stored in Amazon S3
4. A high number of parameters in a foundation model

Correct Answer: 1

Explanation

Label noise occurs when training examples contain incorrect, inconsistent, ambiguous, or otherwise unreliable labels. For supervised learning, models use these labels as targets, so inaccurate labels can cause the model to learn incorrect relationships. For example, if some images of cats are incorrectly labeled as dogs, the model may receive conflicting signals during training. Identifying and correcting noisy labels can improve training-data quality and model performance. Label noise is a data-quality issue rather than a problem involving network latency, model size, or storage.

Question 350

A company evaluates a text-generation model using two different prompts and two different model versions. To determine which change actually improved performance, what should the company do?

1. Change the dataset, prompt, and model simultaneously.
2. Compare results using a consistent evaluation dataset and controlled evaluation procedure.
3. Evaluate each version using randomly selected metrics.
4. Select the version that produces the longest responses.

Correct Answer: 2

Explanation

A controlled evaluation procedure makes it easier to attribute observed performance differences to a specific change. Using the same representative evaluation dataset, evaluation criteria, and measurement process reduces unrelated variation. If the prompt, dataset, and model are all changed simultaneously, it becomes difficult to determine which factor caused an improvement or decline. Response length alone is not a reliable indicator of quality. Consistent testing therefore provides a stronger basis for comparing model versions or prompt configurations.

Question 351

Which AWS service provides compliance reports and agreements that can help customers understand AWS compliance information?

1. Amazon Inspector
2. AWS Security Hub
3. AWS Artifact
4. Amazon GuardDuty

Correct Answer: 3

Explanation

AWS Artifact provides on-demand access to AWS security and compliance documents, including reports and agreements relevant to AWS services. Organizations can use these resources when assessing their compliance requirements and understanding the controls and certifications associated with AWS. It is particularly useful during governance, auditing, and compliance activities. AWS Security Hub aggregates security findings, Amazon Inspector assesses workloads for vulnerabilities, and Amazon GuardDuty provides threat detection. AWS Artifact is therefore the service specifically focused on providing AWS compliance documentation and related agreements.

Question 352

A retrieval-augmented generation application retrieves 20 documents for every question, but most are unrelated to the user’s request. Which adjustment can reduce unnecessary context while preserving the most relevant information?

1. Increase the number of retrieved documents to 100.
2. Remove all metadata from indexed documents.
3. Disable retrieval and rely entirely on the foundation model.
4. Tune the retrieval top-k value to return a smaller relevant set.

Correct Answer: 4

Explanation

The top-k retrieval setting controls how many of the highest-ranked results are returned for a query. If an application retrieves too many unrelated documents, reducing k can limit the amount of irrelevant information passed into the generation stage. The appropriate value should be determined through evaluation because setting it too low could exclude useful evidence. Increasing the number of documents would generally add more irrelevant content, while disabling retrieval removes the grounding mechanism entirely. Top-k tuning can therefore improve retrieval efficiency and contextual relevance.

Question 353

Which measure is commonly used to evaluate the quality of a regression model by calculating the square root of the average squared prediction errors?

1. RMSE
2. Precision
3. Specificity
4. Recall

Correct Answer: 1

Explanation

Root Mean Squared Error, or RMSE, measures the square root of the average squared differences between predicted and actual values. Because errors are squared before averaging, larger errors receive greater weight, making RMSE useful when large prediction errors are particularly important. The resulting value is expressed in the same units as the target variable, which can make it easier to interpret than mean squared error. Precision, specificity, and recall are classification metrics and do not directly measure continuous regression prediction error.

Question 354

A developer wants to compare two foundation models for a customer-support application. One model produces shorter answers while the other provides longer answers. Which criterion is most useful for making the comparison?

1. Choose the model with the largest parameter count automatically.
2. Evaluate both models against the application’s defined quality requirements using the same test cases.
3. Choose the model with the longest responses.
4. Choose the model with the highest token usage.

Correct Answer: 2

Explanation

Foundation models should be compared according to the requirements of the intended application rather than a single characteristic such as response length or parameter count. Using the same representative test cases and predefined evaluation criteria allows the organization to compare factors such as correctness, relevance, safety, latency, and cost as appropriate. A longer response is not necessarily a better response, and higher token usage can increase cost without improving usefulness. Consistent application-specific evaluation provides evidence for selecting a suitable model.

Question 355

Which practice helps reduce the risk that an AI application’s logs expose sensitive information?

1. Store every prompt and response indefinitely.
2. Disable all security monitoring.
3. Redact or filter sensitive information before retaining logs.
4. Give every administrator unrestricted access to logs.

Correct Answer: 3

Explanation

Logs can contain prompts, model responses, identifiers, or other information that may be sensitive. Redacting or filtering sensitive content before logs are retained reduces the amount of confidential information stored in monitoring systems. Organizations should also apply appropriate access controls and retention policies. Keeping every prompt and response indefinitely increases exposure, while disabling monitoring can remove useful operational and security visibility. Unrestricted administrator access also increases risk. Secure logging should balance observability with privacy and data-protection requirements.

Question 356

An AI-powered application must continue serving users when its primary model temporarily becomes unavailable. Which architecture pattern can provide this behavior?

1. Configure a secondary model or fallback path for supported requests.
2. Increase the prompt temperature whenever the primary model fails.
3. Delete the application’s evaluation dataset.
4. Disable error handling to allow requests to retry indefinitely.

Correct Answer: 1

Explanation

A fallback architecture can improve application resilience by routing supported requests to an alternative model or processing path when the primary model is unavailable. The fallback should be tested to ensure that it meets minimum requirements for quality, safety, cost, and latency. Simply changing generation settings does not resolve model availability problems. Unlimited retries can create additional failures or unnecessary load. A planned fallback strategy allows the application to degrade gracefully while maintaining service availability during temporary dependency failures.

Question 357

Which concept describes the ability of a model to perform well on previously unseen examples rather than only memorized training examples?

1. Regularization
2. Generalization
3. Tokenization
4. Quantization

Correct Answer: 2

Explanation

Generalization is the ability of a machine learning model to perform effectively on new, unseen data. A model that generalizes well has learned useful patterns rather than simply memorizing the training examples. Generalization is commonly assessed using data that was not used to train the model. Regularization can help improve generalization by discouraging overly complex models, but it is a technique rather than the outcome itself. Tokenization converts text into model-readable units, while quantization reduces numerical precision to improve efficiency.

Question 358

A company wants to process a large collection of independent documents with an AI model without requiring an immediate response for each document. Which inference approach is appropriate?

1. Interactive real-time inference only
2. Streaming inference for every document
3. Batch inference
4. Manual inference by an administrator

Correct Answer: 3

Explanation

Batch inference is appropriate when many independent inputs can be processed together without requiring an immediate response to each individual request. This approach can be useful for large-scale document processing, periodic analysis, classification jobs, or generating results from an existing dataset. Real-time inference is more appropriate when users require immediate responses. Streaming can be useful for continuously delivered results but is not necessary for every offline processing workload. Batch processing can therefore simplify large workloads where immediate interactive responses are not required.

Question 359

Why might an organization use Amazon Polly in an AI-powered application?

1. To identify malicious network traffic
2. To extract tables from scanned documents
3. To convert text into natural-sounding speech
4. To search vector embeddings

Correct Answer: 3

Explanation

Amazon Polly is an AWS service that converts text into lifelike spoken audio. It can be used in applications such as voice assistants, accessibility solutions, notifications, educational applications, and other systems that need to present text audibly. It is different from Amazon Transcribe, which converts speech into text. Document extraction is associated with Amazon Textract, while vector search is performed through services designed for retrieval workloads. Therefore, converting written text into synthesized speech is the primary capability represented in this question.

Question 360

An organization wants to determine whether its AI model is safe enough for production by testing harmful requests, unexpected inputs, and attempts to bypass safety controls. What activity best describes this process?

1. Data normalization
2. Model compression
3. Feature scaling
4. Red-team testing

Correct Answer: 4

Explanation

Red-team testing deliberately challenges an AI system with harmful, adversarial, unexpected, or policy-violating inputs to identify weaknesses before deployment. Testers may attempt to bypass safeguards, expose sensitive information, manipulate model behavior, or trigger unsafe outputs. Findings can then be used to improve prompts, guardrails, application controls, and monitoring. This activity differs from data normalization, feature scaling, and model compression, which address data preparation or model efficiency. Red-team testing is therefore an important part of evaluating AI safety and security before production use.