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

 

Question 261

A company is preparing a dataset for an AI application and discovers that some numerical records contain impossible values, such as a customer’s age being 350 years. What should the company perform?

1. Data validation and anomaly checking
2. Increase the model temperature
3. Expand the context window
4. Change the model’s output format

Correct Answer: 1

Explanation

Data validation and anomaly checking can identify values that fall outside reasonable or permitted ranges. Impossible values can result from data-entry mistakes, system errors, corrupted records, or incorrect transformations. If such records are used without investigation, they may negatively affect model training or evaluation. Organizations should define acceptable ranges based on the meaning of each field and investigate suspicious records before deciding whether to correct, remove, or retain them. These checks are part of broader data-quality practices and should be repeated as new data enters the AI system.

Question 262

Which statement best describes pseudonymization in the context of protecting personal information?

1. Permanently deleting all personal information
2. Replacing identifying information with substitute identifiers while retaining the ability to associate it under controlled conditions
3. Encrypting every model parameter
4. Converting text into numerical embeddings

Correct Answer: 2

Explanation

Pseudonymization replaces directly identifying information with substitute values, such as randomly generated identifiers. The original identity may still be recoverable when additional information or a controlled mapping is available. This can reduce unnecessary exposure of direct identifiers during data processing, although pseudonymized information may still be considered personal data under applicable requirements. Organizations should protect the mapping information and restrict access appropriately. Pseudonymization is different from anonymization, where information is processed so that identifying an individual is intended to no longer be reasonably possible.

Question 263

A development team wants to compare two foundation models using the same prompts, evaluation examples, and scoring criteria. Why is this approach useful?

1. It guarantees that the more expensive model will perform better
2. It eliminates the need for human review
3. It provides a more consistent basis for comparing model performance
4. It prevents all hallucinations

Correct Answer: 3

Explanation

Using the same prompts, evaluation examples, and scoring criteria creates a more consistent basis for comparing foundation models. Without a consistent methodology, differences in prompts or test data could influence the results and make comparisons less meaningful. Evaluation can consider factors such as response quality, factuality, safety, latency, and cost depending on the application’s requirements. A controlled comparison does not guarantee that one model is universally better than another. Instead, it helps organizations understand how each model performs against the requirements and scenarios that matter for their particular workload.

Question 264

An application allows a generative AI model to call external tools. Which practice can reduce the risk of the model repeatedly calling a tool because of an unexpected loop?

1. Increasing the number of available tools
2. Removing all execution logs
3. Allowing unlimited recursive calls
4. Setting execution limits and loop-detection controls

Correct Answer: 4

Explanation

Execution limits and loop-detection controls can help prevent an AI agent from repeatedly invoking tools because of unexpected reasoning or workflow behavior. Applications can establish limits on the number of actions, execution time, retries, or repeated calls within a single task. Tool results should also be validated before additional actions are authorized. These controls are particularly important when tools can modify data, incur costs, or interact with external systems. Agent safety should not depend entirely on the model behaving as expected because application-level controls provide an additional protection layer.

Question 265

A company wants to reduce unnecessary storage of customer information that is not required for its AI use case. Which responsible data practice should it follow?

1. Data minimization
2. Data duplication
3. Unlimited data retention
4. Maximum feature collection

Correct Answer: 1

Explanation

Data minimization means collecting, processing, and retaining only the information necessary for a specific purpose. Limiting unnecessary customer information can reduce privacy exposure, storage requirements, and the potential impact of a data breach. Organizations should first determine what information is actually required for the AI application and establish appropriate retention rules. Data minimization does not mean removing information blindly; required information should remain available when justified by the use case. The practice should be incorporated into data collection, preprocessing, model development, deployment, and ongoing data-management procedures.

Question 266

A company wants an AI application to stop retaining customer prompts after the information is no longer required for the application’s stated purpose. What should the organization establish?

1. Higher model temperature
2. A data-retention and deletion policy
3. More training epochs
4. A larger vector dimension

Correct Answer: 2

Explanation

A data-retention and deletion policy defines how long information should be retained and when it should be securely deleted or otherwise disposed of. For AI applications, such policies can cover user prompts, generated outputs, logs, uploaded documents, and other stored information. Retention periods should reflect business requirements, legal obligations, privacy considerations, and operational needs. Organizations should also identify where copies of information exist because deleting one copy may not remove related records elsewhere. Clearly defined retention procedures help reduce unnecessary exposure and support consistent information lifecycle management.

Question 267

An organization wants to document important information about an AI model, including its intended uses, limitations, evaluation results, and known risks. What artifact is most appropriate?

1. A network diagram
2. A database index
3. A model card
4. A DNS record

Correct Answer: 3

Explanation

A model card is a documentation artifact that can describe important characteristics of a model, including intended uses, limitations, evaluation information, and known risks. Model cards can improve transparency by giving developers and users a structured reference for understanding how a model should and should not be used. The exact contents can vary depending on the model and organization. Maintaining accurate documentation is particularly useful when models are reused across applications or updated over time. Model documentation should complement technical testing, governance procedures, and operational controls rather than replace them.

Question 268

A company deploys an AI application in several AWS Regions and wants users to receive service access only where the required AI capability is available. What should the company verify before deployment?

1. The color of the application’s interface
2. Regional service and model availability
3. The number of keyboard shortcuts
4. The document font size

Correct Answer: 2

Explanation

AWS services and individual AI models or capabilities may have different availability across AWS Regions. Before deploying an AI application, organizations should verify that the required service, model, feature, and related dependencies are available in the intended Regions. Regional availability can affect architecture, latency, data-residency requirements, and operational planning. Teams should also verify quotas and service-specific restrictions. Designing around assumed availability can lead to deployment problems if a required capability cannot be accessed in the selected Region, so availability should be checked as part of architecture planning.

Question 269

A generative AI application needs to prevent users from requesting content about certain prohibited subjects defined by the organization. Which control is most directly relevant?

1. Topic-based content filtering
2. Increasing token limits
3. Feature scaling
4. Data compression

Correct Answer: 1

Explanation

Topic-based content filtering can restrict or identify requests involving subjects that an organization has determined should not be handled by its AI application. Such controls can be useful for implementing safety requirements and business-specific restrictions. The organization should define prohibited topics clearly and test the control with different phrasings because users may express the same intent in many ways. Filtering should be combined with other safeguards, such as input validation, output controls, monitoring, and human review where appropriate. No single safety mechanism should be treated as a complete protection against every possible misuse.

Question 270

An AI application experiences temporary service throttling when many requests arrive simultaneously. Which design approach can help the application handle temporary throttling more gracefully?

1. Disable all error handling
2. Send unlimited requests immediately
3. Retry with appropriate backoff and rate control
4. Remove authentication

Correct Answer: 3

Explanation

Retry strategies with appropriate backoff and rate control can help applications handle temporary throttling more gracefully. Instead of immediately sending repeated requests, the application can wait before retrying and gradually increase the delay when appropriate. Rate control can also prevent the application from generating excessive request volume in the first place. Retry logic should distinguish temporary conditions from permanent errors and should include reasonable limits. Proper handling improves application resilience while reducing the risk that repeated retries will increase service pressure or unnecessarily increase operational costs.

Question 271

Which AWS capability can provide a private network path from a VPC to supported AWS services without requiring traffic to traverse the public internet?

1. AWS PrivateLink or VPC endpoints
2. Amazon Route 53 public hosted zones
3. Amazon CloudFront distributions
4. Internet-facing load balancers

Correct Answer: 1

Explanation

AWS PrivateLink and supported VPC endpoint configurations can provide private connectivity between a VPC and supported AWS services or endpoints. This can help organizations design workloads where service traffic does not need to use a public internet path. Private connectivity can be especially relevant for applications handling sensitive information or operating under strict network requirements. The exact endpoint type and supported service depend on the AWS architecture. Network controls, security groups, routing, and service availability should also be reviewed because private connectivity does not automatically make an application secure.

Question 272

An organization wants to assign different AI-related costs to departments according to the resources they use. Which practice can help organize this information?

1. Increasing model parameters
2. Applying resource tags for cost allocation
3. Changing prompt wording
4. Increasing token randomness

Correct Answer: 2

Explanation

Resource tagging can help organizations categorize AWS resources according to departments, applications, projects, or other business dimensions. When supported by the relevant AWS billing and cost-management capabilities, tags can contribute to cost allocation and reporting. This can help teams understand which workloads generate AI-related expenses and support internal budgeting. Tags should be applied consistently and according to an agreed naming convention. Cost allocation practices should also account for services or charges that cannot be directly attributed through resource tags, so tagging should be viewed as one part of broader cloud cost management.

Question 273

A team wants to protect a sensitive encryption key used by an AI application and ensure that only authorized identities can use it for permitted cryptographic operations. What should the team focus on?

1. Model temperature
2. Prompt length
3. Key access policies and permissions
4. Image resolution

Correct Answer: 3

Explanation

Key access policies and permissions determine which identities or services can use an encryption key and what operations they are allowed to perform. Restricting access to authorized identities helps reduce the risk of unauthorized decryption or cryptographic operations. Organizations should follow least-privilege principles when configuring permissions and should monitor key usage where appropriate. Key management should also consider rotation, auditing, and lifecycle requirements. Encryption alone does not guarantee protection if unauthorized principals can access the keys, so effective key access controls are an important part of securing AI data and application resources.

Question 274

A company evaluates a generative AI system using both automated metrics and reviewers who judge whether responses are useful, accurate, and appropriate. What type of evaluation is this?

1. Network evaluation
2. Storage evaluation
3. Human and automated evaluation
4. DNS evaluation

Correct Answer: 3

Explanation

Combining automated metrics with human evaluation provides multiple perspectives on generative AI quality. Automated methods can efficiently evaluate large numbers of responses according to defined criteria, while human reviewers can assess qualities that may be difficult to capture with a simple numerical metric, such as usefulness, relevance, tone, or appropriateness. Human evaluation should use clear guidelines and representative examples to improve consistency between reviewers. The combination of evaluation methods can provide a more comprehensive assessment, particularly when generated responses are open-ended and cannot be judged reliably by a single automated measurement.

Question 275

A company is considering a smaller foundation model for a simple classification task because the larger model provides little additional value. What is a potential benefit of selecting the smaller model?

1. Potentially lower cost and latency
2. Guaranteed higher accuracy
3. Elimination of security requirements
4. Unlimited context capacity

Correct Answer: 1

Explanation

A smaller foundation model may provide lower inference cost and latency when it is capable of meeting the application’s quality requirements. Model selection should consider the actual workload rather than assuming that a larger model is always necessary. Teams can compare models using representative prompts, expected workloads, quality criteria, safety requirements, and performance targets. A smaller model is not automatically better for every task, and it may have limitations in reasoning, context handling, or output quality. The appropriate choice depends on the application’s requirements and measured evaluation results.

Question 276

Which practice can help reduce the risk that confidential information is accidentally included in prompts sent to a generative AI application?

1. Increasing output randomness
2. Expanding the model context window
3. Input screening and sensitive-data redaction
4. Increasing the number of model parameters

Correct Answer: 3

Explanation

Input screening and sensitive-data redaction can help prevent confidential information from being unnecessarily included in prompts. An application can inspect incoming content for defined sensitive information and remove, mask, or otherwise handle it according to organizational requirements before sending the request to the model. The appropriate approach depends on the data and use case. Screening should be tested carefully because excessive filtering may remove information required for legitimate tasks. Organizations should also establish clear policies for prompt handling, logging, retention, and access because data protection extends beyond the initial request.

Question 277

An organization wants to compare the effect of two prompt versions on a generative AI application’s response quality. Which approach provides the clearest comparison?

1. Change the prompt, model, and evaluation data simultaneously
2. Use controlled A/B testing with the same evaluation criteria
3. Evaluate each version using unrelated datasets
4. Judge only the longest generated response

Correct Answer: 2

Explanation

Controlled A/B testing can provide a clearer comparison between two prompt versions by changing the intended variable while keeping other evaluation conditions consistent. The same or equivalent representative test cases, model configuration, and scoring criteria should be used where appropriate. This helps isolate the effect of the prompt change. Evaluation should consider the requirements of the application, such as factuality, relevance, safety, latency, and cost. A controlled comparison does not guarantee that one prompt will be better for every scenario, so results should be interpreted against defined acceptance criteria.

Question 278

A security team is concerned that an attacker might attempt to discover information about a model by repeatedly querying it and analyzing its responses. Which threat does this describe?

1. Data normalization
2. Model extraction or model stealing
3. Data labeling
4. Feature scaling

Correct Answer: 2

Explanation

Model extraction, sometimes called model stealing, involves attempting to reproduce or approximate the behavior of a target model by making queries and analyzing its outputs. An attacker may use collected input-output examples to build another model that imitates the target. Organizations can reduce exposure by applying appropriate access controls, authentication, rate limits, monitoring, and usage restrictions. Sensitive model interfaces should not necessarily be exposed without safeguards. The specific risk depends on the model, API design, information revealed through responses, and the attacker’s ability to generate a large number of queries.

Question 279

An AI application produces responses that are properly formatted but sometimes include unsafe instructions. Which evaluation area should receive additional attention?

1. Safety evaluation
2. Storage capacity
3. Network addressing
4. Database indexing

Correct Answer: 1

Explanation

Safety evaluation examines whether an AI system produces harmful, inappropriate, or otherwise unacceptable outputs under relevant conditions. A model can produce technically well-formatted responses while still failing safety requirements, so output structure alone is not sufficient. Testing should include representative benign requests as well as challenging or adversarial scenarios relevant to the application’s purpose. Organizations can define safety criteria, apply appropriate filtering or guardrails, and include human review for higher-risk workflows. Safety evaluation should be repeated after significant changes to the model, prompts, data, or application controls.

Question 280

A team wants to reduce the amount of computation and storage required by a deployed machine learning model while preserving acceptable performance. Which technique may help?

1. Increasing the model size
2. Increasing training epochs indefinitely
3. Model compression
4. Duplicating model artifacts

Correct Answer: 3

Explanation

Model compression refers to techniques that reduce the computational or storage requirements of a model while attempting to preserve acceptable performance. Depending on the approach, compression may involve reducing numerical precision, removing unnecessary components, or using other optimization techniques. A compressed model should be evaluated carefully because reducing size or computation can affect accuracy, latency, or other quality characteristics. The appropriate technique depends on the model and deployment environment. Organizations should compare the compressed model against the original using representative workloads and predefined performance and quality requirements.