View Full Microsoft AI-901 Exam Dumps and Practice Test Dumps

Question 281. Which factor should be evaluated when choosing a generative AI model for an application?

1. Only the model’s name
2. The application’s requirements, performance, and cost
3. The color of the user interface
4. The number of database tables

Correct Answer: 2. The application’s requirements, performance, and cost

Explanation:

Choosing a generative AI model should be based on the requirements of the intended application rather than on the model name alone. Developers should consider factors such as the type of task, expected response quality, supported languages, latency, context requirements, throughput, and cost. Some models may be better suited for complex reasoning, while others may provide faster or more economical responses for simpler tasks. Evaluation should use representative prompts and realistic workloads. Selecting an appropriate model helps ensure that the AI solution provides the required functionality while remaining practical to operate and maintain in its intended environment.

Question 282. What is a benefit of using grounding with a generative AI model?

1. It provides the model with relevant external information
2. It removes the need for evaluation
3. It guarantees every response is correct
4. It disables all user input

Correct Answer: 1. It provides the model with relevant external information

Explanation:

Grounding provides a generative AI model with relevant information from trusted sources so that its responses can be based on specific external context. For example, an organization can retrieve information from internal documents and provide that information to a language model when answering an employee’s question. This can reduce the need for the model to rely only on information contained in its original training. Grounding can improve relevance and factuality, but it does not guarantee that every generated response is correct. Retrieved information should still be accurate, current, appropriately authorized, and relevant to the user’s request.

Question 283. Which practice can help protect sensitive information in an AI application?

1. Sending all available personal data to every model request
2. Removing all authentication
3. Applying data minimization and appropriate access controls
4. Storing passwords in prompts

Correct Answer: 3. Applying data minimization and appropriate access controls

Explanation:

Data minimization means collecting and processing only the information necessary for a specific purpose. In an AI application, this can reduce the amount of sensitive information exposed to models, services, logs, or other components. Access controls further restrict who or what can access data and AI resources. Together, these practices help reduce unnecessary exposure and support privacy and security requirements. Developers should also consider data retention, encryption, identity management, and monitoring. Sensitive information should not be included in prompts simply because it is available. The application should determine what information is actually necessary for each AI operation.

Question 284. What is the purpose of a system message in a generative AI application?

1. To define high-level instructions and behavior for the model
2. To store binary files
3. To create a network connection
4. To calculate database storage

Correct Answer: 1. To define high-level instructions and behavior for the model

Explanation:

A system message provides high-level instructions that establish how a generative AI model should behave within an application. It can define the assistant’s role, response style, rules, constraints, or other instructions that should guide interactions. For example, an application might instruct the model to respond as a technical support assistant and avoid exposing confidential information. System instructions are different from ordinary user prompts because they are typically part of the application’s control structure. Developers should design them carefully and avoid assuming that instructions alone provide complete security. Sensitive actions should also be protected through application-level controls.

Question 285. Which approach can improve the quality of a prompt for a generative AI model?

1. Providing vague instructions with no context
2. Giving clear instructions and relevant context
3. Removing the desired output format
4. Using unrelated information

Correct Answer: 2. Giving clear instructions and relevant context

Explanation:

Clear instructions and relevant context help a generative AI model understand what the application or user expects. A well-designed prompt can specify the task, provide necessary background information, identify constraints, and describe the desired output format. For example, instead of asking a model to summarize a document without additional guidance, a prompt could specify the target audience, required length, and important topics to include. Prompt quality should be evaluated using representative examples because a prompt that performs well for one task may not perform equally well for another. Clear prompts can improve consistency and reduce ambiguity.

Question 286. Which Azure AI Search approach is useful when exact keywords and semantic similarity are both important?

1. Hybrid search
2. Speech recognition
3. Image classification
4. Document translation

Correct Answer: 1. Hybrid search

Explanation:

Hybrid search combines keyword-based retrieval with vector or semantic retrieval. Keyword search is useful when exact words, names, identifiers, or specific phrases are important. Vector-based retrieval can identify content that is semantically related even when the wording differs from the user’s query. Using both approaches can therefore provide broader and more relevant search results. This is particularly useful in enterprise search and retrieval-augmented generation applications, where users may express a concept differently from the wording contained in the source documents. Azure AI Search can support these combined retrieval patterns to help applications find relevant information more effectively.

Question 287. What is an embedding used for in a semantic search solution?

1. Representing content as numerical vectors
2. Encrypting a database
3. Converting text into an audio file
4. Creating a virtual machine

Correct Answer: 1. Representing content as numerical vectors

Explanation:

An embedding represents information such as text as a numerical vector that captures aspects of its semantic meaning. Similar pieces of content can produce vectors that are relatively close to each other in vector space, allowing systems to perform similarity-based retrieval. Embeddings are commonly used in semantic search, recommendation systems, and retrieval-augmented generation. For example, a user’s question can be converted into an embedding and compared with embeddings stored for document chunks. The retrieved content can then be supplied to a generative AI model. The usefulness of embeddings depends on the quality and suitability of the embedding model and data.

Question 288. Which scenario is an example of anomaly detection?

1. Predicting the exact price of a house
2. Identifying unusual transactions in a normal transaction stream
3. Translating a document
4. Converting speech into text

Correct Answer: 2. Identifying unusual transactions in a normal transaction stream

Explanation:

Anomaly detection focuses on identifying observations or patterns that differ significantly from expected behavior. In a transaction system, for example, most transactions may follow normal patterns while a small number could have unusual amounts, timing, locations, or other characteristics. An anomaly detection solution can flag these unusual observations for additional investigation. Unlike ordinary classification, anomaly detection may be useful when examples of every possible abnormal condition are not available. Developers should carefully define what constitutes unusual behavior for the specific scenario and evaluate detection results because false positives and false negatives can both affect business processes.

Question 289. What is feature engineering in a machine learning workflow?

1. Creating or transforming input variables to improve model learning
2. Deleting every training example
3. Deploying a website
4. Encrypting model names

Correct Answer: 1. Creating or transforming input variables to improve model learning

Explanation:

Feature engineering involves creating, transforming, selecting, or otherwise preparing input variables so that they provide useful information for a machine learning model. For example, raw transaction timestamps could be transformed into features such as day of week or hour of day, depending on the problem. Effective features can help a model identify meaningful patterns more easily. Feature engineering should be based on knowledge of the data and the business problem, and transformations must be applied consistently to training and production data. Poorly designed features can introduce noise or leakage, so they should be evaluated carefully during model development.

Question 290. Which dataset is normally used to measure final model performance on data not used for training or tuning?

1. Training dataset
2. Validation dataset
3. Test dataset
4. Configuration dataset

Correct Answer: 3. Test dataset

Explanation:

A test dataset is used to evaluate a trained model against data that was kept separate from the training and model-tuning process. This provides an estimate of how well the model may perform on previously unseen data. The training dataset is used to learn model parameters, while the validation dataset can be used during development to compare approaches or tune settings. Keeping a separate test dataset helps reduce the risk of reporting overly optimistic performance caused by repeatedly evaluating the model on data used during development. The test set should be representative of the intended real-world scenario.

Question 291. What does model generalization describe?

1. The ability of a model to perform well on unseen but relevant data
2. The number of servers used for training
3. The amount of storage in a subscription
4. The number of prompts in a database

Correct Answer: 1. The ability of a model to perform well on unseen but relevant data

Explanation:

Generalization refers to how well a machine learning model performs on new data that was not used during training but comes from a relevant distribution. A model that memorizes training examples may achieve excellent training performance but perform poorly on new examples. Good generalization indicates that the model has learned useful patterns rather than simply memorizing the training data. Developers can evaluate generalization using validation and test datasets that represent the intended use case. Data quality, model complexity, feature selection, and appropriate training practices can all influence how effectively a model generalizes.

Question 292. Which situation is a common sign of overfitting?

1. Similar performance on training and unseen data
2. Excellent training performance but significantly worse validation performance
3. No training data is available
4. The model contains no parameters

Correct Answer: 2. Excellent training performance but significantly worse validation performance

Explanation:

Overfitting occurs when a model learns the training data too closely, including patterns or noise that do not generalize to new examples. A common indication is very strong performance on the training dataset combined with noticeably weaker performance on validation or test data. Developers can address overfitting through approaches such as obtaining more representative training data, reducing model complexity, regularization, feature selection, or other appropriate techniques. The correct approach depends on the model and problem. Evaluating the model on data that was not used for training is essential for identifying whether the model generalizes adequately.

Question 293. What is one reason representative training data is important for an AI model?

1. It helps the model learn patterns relevant to the intended users and scenarios
2. It guarantees zero errors
3. It eliminates the need for testing
4. It prevents all model updates

Correct Answer: 1. It helps the model learn patterns relevant to the intended users and scenarios

Explanation:

Representative training data should reflect the important characteristics, populations, conditions, and scenarios in which the model is expected to operate. If important groups or situations are missing or poorly represented, the model may perform differently across those groups or conditions. For example, a vision model trained mostly on one type of lighting may perform poorly under different lighting conditions. Representative data does not guarantee perfect performance, but it provides a stronger foundation for learning relevant patterns. Developers should also evaluate model performance across important groups and scenarios to identify potential limitations and improve the reliability of the system.

Question 294. Which metric is particularly useful for evaluating a classifier when false negatives are costly?

1. Recall
2. Storage capacity
3. Latency only
4. Number of model parameters

Correct Answer: 1. Recall

Explanation:

Recall measures the proportion of actual positive cases that a classification model correctly identifies. It is especially important when missing a positive case, known as a false negative, has significant consequences. For example, in a scenario where failing to detect a critical event is more costly than investigating additional false alarms, recall may be an important evaluation metric. Precision may also matter because it measures the proportion of predicted positives that are actually positive. The appropriate metric depends on the business problem and the consequences of different errors. Developers should consider multiple metrics rather than relying on one measurement alone.

Question 295. What does model monitoring help an organization detect after deployment?

1. Changes in model behavior or input data that may affect performance
2. The physical location of every user
3. The color of model output
4. The number of programming languages installed

Correct Answer: 1. Changes in model behavior or input data that may affect performance

Explanation:

Model monitoring helps organizations observe AI systems after deployment and identify conditions that may affect their expected performance. Monitoring can include prediction quality, input data characteristics, data drift, operational metrics, latency, errors, and other indicators relevant to the application. Continuous monitoring is important because real-world conditions can change after a model is deployed. For example, user behavior or business processes may change, causing production data to differ from the original training data. Monitoring does not automatically solve these issues, but it provides information that teams can use to investigate, update, retrain, or otherwise manage the deployed model.

Question 296. Which Azure Machine Learning deployment approach is designed for processing large amounts of data asynchronously?

1. Batch endpoint
2. Online endpoint
3. Chat prompt
4. Speech endpoint

Correct Answer: 1. Batch endpoint

Explanation:

An Azure Machine Learning batch endpoint is designed for asynchronous processing of larger volumes of data. Instead of requiring an immediate prediction for every request, an application can submit a batch workload and allow the endpoint to process the data over time. This approach is useful for scenarios such as scoring large datasets, generating predictions for scheduled reports, or processing periodic data pipelines. Online endpoints are better suited to real-time inference where applications need immediate responses. Choosing between batch and online deployment depends on latency requirements, workload size, processing patterns, and the business needs of the application.

Question 297. Which practice helps ensure that an AI model can be rolled back when a new deployment performs poorly?

1. Maintaining previous model versions
2. Deleting all earlier models
3. Disabling model evaluation
4. Removing deployment records

Correct Answer: 1. Maintaining previous model versions

Explanation:

Keeping previous model versions allows a team to identify which version is currently deployed and retain an earlier version that can be restored if a new deployment causes unexpected problems. A controlled model lifecycle can include version tracking, evaluation, deployment stages, monitoring, and rollback procedures. Before replacing an existing model, teams should compare the new version against appropriate performance and business requirements. If production monitoring shows that the new version does not meet expectations, having a known previous version available can simplify recovery. Version management therefore supports safer and more controlled AI deployment practices.

Question 298. Which responsible AI principle focuses on making AI systems understandable to relevant stakeholders?

1. Transparency
2. Storage optimization
3. Network routing
4. Data compression

Correct Answer: 1. Transparency

Explanation:

Transparency involves providing appropriate information about how an AI system works, what it is intended to do, what data or processes may influence its behavior, and what limitations users should understand. The amount and type of information required can vary depending on the application and its risks. Transparency can help users and stakeholders make informed decisions about when and how to rely on AI-generated results. It can also support accountability by making system behavior and limitations easier to examine. Transparency does not mean that every technical implementation detail must be exposed to every user; information should be communicated appropriately for the intended audience.

Question 299. Why should AI model outputs be evaluated before being used in a production application?

1. To determine whether the model meets the application’s quality and safety requirements
2. To guarantee that the model never changes
3. To eliminate all user feedback
4. To increase the number of database tables

Correct Answer: 1. To determine whether the model meets the application’s quality and safety requirements

Explanation:

AI model evaluation helps determine whether a model performs adequately for its intended purpose before it is released into production. Evaluation can include accuracy or task-specific metrics, representative test cases, safety testing, robustness checks, fairness considerations, latency, and other requirements. Generative AI applications may also require evaluation of factuality, relevance, harmful content, and instruction-following behavior. Testing should reflect realistic inputs and important edge cases rather than relying only on ideal examples. Evaluation does not guarantee perfect behavior, but it provides evidence that the system has been tested against defined requirements and helps identify areas requiring improvement or additional safeguards.

Question 300. Which practice can help improve the reliability of an AI application over time?

1. Monitoring performance and updating the system when evidence shows it is needed
2. Never reviewing production results
3. Removing all evaluation datasets
4. Ignoring changes in real-world data

Correct Answer: 1. Monitoring performance and updating the system when evidence shows it is needed

Explanation:

AI applications operate in changing environments, so ongoing monitoring is important after deployment. User behavior, input data, business processes, and external conditions can change over time, potentially affecting model performance or relevance. Monitoring can identify changes in data distributions, prediction quality, latency, error rates, or safety-related behavior. When evidence indicates that performance has degraded, teams can investigate the cause and determine whether retraining, updating, reconfiguration, or another corrective action is appropriate. Continuous improvement should be based on measured evidence and defined requirements rather than automatically changing a model without evaluation.