#### 01. A company in the finance sector is using machine learning to predict customer churn and detect fraudulent transactions. The data science team needs to select appropriate machine learning models that use labeled data to make predictions based on past examples. Which of the following methods are examples of supervised learning? (Select TWO.) - Decision tree - Association rule mining - Principal Component Analysis (PCA) - Neural networks - K-means clustering **CORRECT:** "Neural networks" is a correct answer. Neural networks are a type of supervised learning algorithm when they are trained using labeled data. This means the model learns from input-output pairs—for example, customer behavior data (input) and whether they churned or not (output). Once trained, the neural network can predict future outcomes based on new inputs. Neural networks are commonly used in tasks such as image recognition, fraud detection, and churn prediction. They work well for complex relationships between features and target variables. **CORRECT:** "Decision tree" is also a correct answer. Decision trees are also supervised learning algorithms. They work by splitting data based on features to create a tree-like structure that leads to decisions. For instance, a decision tree can learn patterns from customer data to predict if someone is likely to leave a service (churn). Since decision trees require labeled training data to learn the rules and outcomes, they fall under supervised learning. They are easy to interpret and useful for classification and regression tasks. **INCORRECT:** "K-means clustering" is incorrect. K-means is an unsupervised learning method. It groups data into clusters based on similarity, but it does not use labeled data. It's useful when you want to explore patterns in the data without predefined categories, such as customer segmentation. **INCORRECT:** "Principal Component Analysis (PCA)" is incorrect. PCA is not a predictive model—it's a dimensionality reduction technique. It's used to reduce the number of input variables while keeping the essential patterns. Since it doesn't use labels or make predictions, it's considered unsupervised learning. **INCORRECT:** "Association rule mining" is incorrect. This is also unsupervised learning. It's used to find interesting relationships or patterns in data, such as which products are bought together. It does not use labeled data or make predictions like supervised learning does. **References:** https://aws.amazon.com/compare/the-difference-between-machine-learning-supervised-and-unsupervised Domain: Fundamentals of AI and ML --- #### 02. A global marketing team is leveraging Amazon Bedrock to generate high-quality promotional images using foundation models. The creative director wants to ensure that the model produces visually appropriate content by avoiding offensive, violent, or explicit imagery, especially when interpreting vague or open-ended prompts. Which prompt engineering strategy should be applied to define these content boundaries directly within the prompt? - Retrieval-augmented prompting with a curated set of safe image references retrieved at runtime. - Negative prompting that includes explicit instructions to avoid generating harmful, violent, or explicit visuals. - Few-shot prompting with visual description examples that include both suitable and unsuitable image generations. - Chain-of-thought prompting that guides the model through the reasoning steps for composing safe and compliant visuals. **CORRECT:** "Negative prompting that includes explicit instructions to avoid generating harmful, violent, or explicit visuals" is the correct answer. Negative prompting is a prompt engineering strategy where the prompt includes clear instructions on what the model should avoid generating. In the case of image generation, especially when using tools like Amazon Bedrock with foundation models, it's important to ensure that the outputs are aligned with brand safety and compliance policies. Negative prompting helps guide the model by explicitly stating what not to include — such as violent, offensive, or explicit content — which is especially useful when prompts are vague or open to interpretation. By setting these boundaries directly in the prompt, the model is more likely to produce visuals that are appropriate, safe, and aligned with organizational goals. For example: "Generate a promotional image of a happy family in a park. Avoid violent, explicit, or disturbing elements." This form of instruction helps the model filter its generations accordingly. **INCORRECT:** "Few-shot prompting with visual description examples that include both suitable and unsuitable image generations" is incorrect. Few-shot prompting provides examples of desired outputs, but including examples of unsuitable generations might confuse the model. While few-shot prompting can help set style or tone, it's not the most direct or effective way to enforce content safety boundaries in generative outputs. **INCORRECT:** "Retrieval-augmented prompting with a curated set of safe image references retrieved at runtime" is incorrect. Retrieval-augmented generation (RAG) is a method where external knowledge or references are pulled into the prompt dynamically. While RAG helps in grounding responses with factual or reference data, it's not designed for enforcing safety instructions or content moderation during image generation. **INCORRECT:** "Chain-of-thought prompting that guides the model through the reasoning steps for composing safe and compliant visuals" is incorrect. Chain-of-thought prompting is useful for improving logical reasoning in language generation tasks. However, it's not typically applied to content safety. It adds reasoning steps but doesn't provide strict boundaries like negative prompting does. **References:** https://docs.aws.amazon.com/nova/latest/userguide/prompting-image-negative.html Domain: Applications of Foundation Models --- #### 03. A data science team at a global retail analytics firm is responsible for aggregating and analyzing market data from a range of third-party vendors. These datasets include industry reports, financial indicators, and consumer trends, all sourced externally. The team wants a scalable solution within the AWS ecosystem that enables them to subscribe to licensed data, manage data usage permissions, and automatically receive updates when vendors refresh datasets. Which AWS approach is the most suitable for meeting these needs? - Subscribe to third-party datasets directly through AWS Data Exchange, and use Amazon S3 and AWS Glue to manage, transform, and integrate the data into the organization's data lake. - Use AWS Lake Formation to create secure data lakes and manage third-party data access using fine-grained permissions and data sharing features across accounts. - Use Amazon QuickSight to connect to external datasets and visualize third-party data insights with dashboards and embedded analytics. - Deploy Amazon Redshift Data Sharing to exchange datasets with external providers and automate data refresh through cross-account Redshift federated queries. **CORRECT:** "Subscribe to third-party datasets directly through AWS Data Exchange, and use Amazon S3 and AWS Glue to manage, transform, and integrate the data into the organization's data lake" is the correct answer. AWS Data Exchange is a fully managed service that enables users to discover, subscribe to, and use third-party data in the AWS ecosystem. It provides access to thousands of datasets from a wide range of providers, including financial, market, and consumer data sources. When a team subscribes to a dataset, AWS Data Exchange automatically delivers it to Amazon S3 and notifies subscribers when updates are published. These updates can be integrated into data pipelines using AWS Glue for cataloging and transformation. This makes it the best choice for teams looking to automate access, manage permissions, and handle updates of externally licensed data for analytics and AI workloads—all while staying within the AWS ecosystem. **INCORRECT:** "Use AWS Lake Formation to create secure data lakes and manage third-party data access using fine-grained permissions and data sharing features across accounts" is incorrect. Lake Formation is designed for organizing, securing, and managing access to internally managed data lakes. It doesn't provide a mechanism to subscribe to or automatically receive updates from external data vendors. It is more suited for sharing data between internal AWS accounts. **INCORRECT:** "Deploy Amazon Redshift Data Sharing to exchange datasets with external providers and automate data refresh through cross-account Redshift federated queries" is incorrect. Redshift Data Sharing is ideal for near real-time access to data between Redshift clusters, primarily within internal AWS accounts or business units. It doesn't support third-party data subscriptions or automated external vendor updates. It's not intended for managing marketplace or vendor-published datasets. **INCORRECT:** "Use Amazon QuickSight to connect to external datasets and visualize third-party data insights with dashboards and embedded analytics" is incorrect. Amazon QuickSight is a business intelligence tool used to visualize data and build dashboards. It doesn't handle data ingestion, transformation, or subscriptions to external datasets. While useful for visual analytics, it does not meet the requirement to automate data updates from third-party vendors. **References:** https://docs.aws.amazon.com/data-exchange/latest/userguide/what-is.html https://aws.amazon.com/blogs/awsmarketplace/preparing-your-third-party-data-from-aws-data-exchange-with-aws-glue-databrew Domain: Applications of Foundation Models --- #### 04. A company is preparing an internal compliance review to verify its use of AWS for hosting generative AI models aligns with ISO and SOC 2 requirements. Which AWS service helps access relevant compliance reports? - AWS Trusted Advisor - AWS Artifact - AWS Security Hub - Amazon CloudWatch **CORRECT:** "AWS Artifact" is the correct answer. AWS Artifact is a self-service portal that provides on-demand access to AWS compliance reports, security and compliance documentation, and agreements. It is primarily used by organizations that need to demonstrate compliance with various regulatory standards, such as ISO 27001, SOC 1, SOC 2, SOC 3, PCI DSS, and HIPAA. Through AWS Artifact, customers can download audit artifacts and certifications to support internal audits or regulatory requirements. It plays a crucial role in governance, risk, and compliance (GRC) workflows by offering verified documentation directly from AWS. It is especially useful during audits and internal compliance reviews for regulated industries. Since the company needs to verify its compliance with ISO and SOC 2 for hosting generative AI models, AWS Artifact is the correct service. It provides official, up-to-date compliance documents that can be used as evidence in the company's review process. **INCORRECT:** "AWS Trusted Advisor" is incorrect. AWS Trusted Advisor offers real-time guidance to help optimize your AWS environment. It checks for best practices in areas like cost optimization, performance, security, and fault tolerance. However, it does not provide access to compliance reports or certifications like ISO or SOC 2. **INCORRECT:** "AWS Security Hub" is incorrect. AWS Security Hub helps you manage and improve your security posture by collecting and analyzing security findings from across AWS accounts. It provides a unified view of security alerts but does not provide official compliance reports or audit documents needed for ISO or SOC 2 reviews. **INCORRECT:** "Amazon CloudWatch" is incorrect. Amazon CloudWatch monitors your AWS resources and applications, providing metrics, logs, and alarms. While useful for operational monitoring, it does not offer any compliance documentation or reports, making it unsuitable for a compliance review. **References:** https://aws.amazon.com/artifact https://docs.aws.amazon.com/artifact/latest/ug/what-is-aws-artifact.html Domain: Guidelines for Responsible AI --- #### 05. A company is developing an AI system that is still under development, and the developer team has noticed vulnerabilities related to external data injections during the prompt generation phase. What security concerns should be addressed to protect the system from malicious input? - Encryption at rest - Access control policies - Prompt injection attacks - Data compression **CORRECT:** "Prompt injection attacks" is the correct answer. The concern of prompt injection attacks should be addressed to protect the AI system from malicious input during the prompt generation phase. Prompt injection attacks occur when external actors inject harmful or manipulated inputs that can alter the behavior of the AI system, potentially causing it to generate incorrect or inappropriate outputs. These attacks exploit the generative nature of the AI by introducing data that the system interprets and acts upon, possibly leading to unintended or harmful consequences. To mitigate these risks, developers can implement input validation, sanitization processes, and security checks to ensure that only appropriate and safe prompts are processed by the AI system. Additionally, monitoring and logging input data can help detect and respond to potential injection attacks. **INCORRECT:** "Access control policies" is incorrect. While access control policies are essential for securing any system by ensuring that only authorized users can access specific resources, they do not directly address the risk of prompt injection attacks. Access controls protect against unauthorized access to the system but do not prevent authorized users or external interfaces from introducing harmful inputs once they have legitimate access. **INCORRECT:** "Encryption at rest" is incorrect. Encryption at rest is a security measure that protects data stored on disk by converting it into an unreadable form for anyone who does not have the decryption key. While it is crucial for protecting data from unauthorized access or theft, encryption at rest does not prevent prompt injection attacks because it does not address the security of data in transit or being processed. **INCORRECT:** "Data compression" is incorrect. Data compression involves reducing the size of data to save storage space or decrease transmission time. While useful for efficiency, data compression does not contribute to security against prompt injection attacks. It is unrelated to the specific security measures needed to guard against the manipulation of input data in AI systems. **References:** https://docs.aws.amazon.com/prescriptive-guidance/latest/llm-prompt-engineering-best-practices/common-attacks.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 06. An enterprise AI team is building a machine learning platform on AWS to support multiple departments. The platform includes pre-trained models hosted on Amazon SageMaker, training pipelines, and data stored in Amazon S3. To comply with organizational security policies, the team must implement granular access control such that: Data scientists can invoke inference endpoints but cannot modify model artifacts. ML engineers can retrain models and update endpoints. Analysts can only view model performance metrics and logs without accessing underlying training data. As an AI practitioner, which AWS feature or mechanism is most appropriate to enforce resource-level and action-specific permissions across this distributed ML environment? - Use Amazon SageMaker Role-Based Authorization to create roles within SageMaker for each user type and assign those roles directly to individual services. - Use custom AWS IAM policies and scoped permissions to define fine-grained access for each role across SageMaker, S3, and CloudWatch resources. - Use Amazon Macie to classify model artifacts and enforce access restrictions based on data sensitivity levels defined through automated discovery. - Use AWS Resource Access Manager (RAM) to share SageMaker resources between departments and assign access scopes using service-linked roles. **CORRECT:** "Use custom AWS IAM policies and scoped permissions to define fine-grained access for each role across SageMaker, S3, and CloudWatch resources" is the correct answer. Custom AWS IAM (Identity and Access Management) policies and scoped permissions allow you to define precise, fine-grained access controls for AWS services and resources. These policies are written in JSON and specify which actions (like sagemaker:InvokeEndpoint, s3:GetObject, or cloudwatch:GetMetricData) are allowed or denied, on which specific resources, and under what conditions. Scoped permissions ensure that users or roles only have access to exactly what they need—nothing more. For example, data scientists can be restricted to only invoke SageMaker endpoints, while ML engineers can be granted permission to retrain models and update endpoints. This approach provides a secure, flexible way to manage access across a distributed ML environment by aligning with the principle of least privilege. **INCORRECT:** "Use Amazon SageMaker Role-Based Authorization to create roles within SageMaker for each user type and assign those roles directly to individual services" is incorrect. SageMaker itself doesn't provide native role-based authorization internally for managing user-level access. Instead, it depends on IAM for access control. There is no built-in role assignment inside SageMaker that handles fine-grained, service-wide permissions across training, inference, and logs. IAM policies remain the appropriate tool for this level of control. **INCORRECT:** "Use Amazon Macie to classify model artifacts and enforce access restrictions based on data sensitivity levels defined through automated discovery" is incorrect. Amazon Macie is a data security service that helps discover and protect sensitive data in S3 by classifying PII and other sensitive content. While it can identify risks and alert users, it doesn't provide access control mechanisms or enforce action-specific permissions on AWS services like SageMaker. **INCORRECT:** "Use AWS Resource Access Manager (RAM) to share SageMaker resources between departments and assign access scopes using service-linked roles" is incorrect. AWS RAM is used to share AWS resources (like VPC subnets or license configurations) across AWS accounts or organizations. It is not designed to enforce fine-grained, user-specific permissions within a single account or across SageMaker actions. It lacks the granularity and flexibility needed for this use case. **References:** https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies.html https://docs.aws.amazon.com/sagemaker/latest/dg/security-iam.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 07. A company is exploring the use of Amazon Fraud Detector to enhance its fraud detection capabilities across different parts of the business. Which of the following is not a typical use case for Amazon Fraud Detector? - Preventing fraud in online payment transactions - Identifying unusual inventory discrepancies - Flagging suspicious login attempts - Detecting fake account registrations **CORRECT:** "Identifying unusual inventory discrepancies" is the correct answer. Amazon Fraud Detector is designed to detect online fraud patterns using machine learning models trained on historical data. It specializes in identifying behaviors that indicate fraudulent activity, such as fake account creation, payment fraud, and suspicious logins. However, inventory discrepancies typically involve supply chain issues, stock management, or accounting errors, which are better handled by business intelligence or inventory management systems, not fraud detection services. Therefore, identifying inventory discrepancies is not a typical use case for Amazon Fraud Detector. **INCORRECT:** "Detecting fake account registrations" is incorrect. Amazon Fraud Detector provides prebuilt model types that help detect fraudulent account registrations, such as bots or bad actors attempting to create multiple fake profiles. This is one of the primary use cases supported by the service. **INCORRECT:** "Preventing fraud in online payment transactions" is incorrect. One of the most common use cases of Amazon Fraud Detector is to help prevent fraud in e-commerce and financial transactions. It can analyze patterns in transaction data and flag potentially fraudulent activity, helping businesses reduce chargebacks and financial loss. **INCORRECT:** "Flagging suspicious login attempts" is incorrect. Amazon Fraud Detector can analyze login behaviors to detect unusual or suspicious login attempts, such as logins from new devices, IP addresses, or locations. This helps prevent account takeover and unauthorized access. **References:** https://aws.amazon.com/fraud-detector https://docs.aws.amazon.com/frauddetector/latest/ug/what-is-frauddetector.html Domain: Applications of Foundation Models --- #### 08. An energy provider uses AI agents to control power distribution based on demand data collected every second from smart meters. How does using real-time data help the AI agents manage power distribution? - The agents dynamically balance supply based on real-time demand. - The agents generate static reports after peak usage periods. - The agents rely only on monthly billing data to manage power supply. - The agents require human approval before taking any action. **CORRECT:** "The agents dynamically balance supply based on real-time demand" is the correct answer. Real-time data means information is collected and processed instantly or within seconds as events happen. In the case of power distribution, smart meters provide real-time data about how much electricity customers are using at any given moment. This allows AI agents to analyze the current demand and adjust the power supply immediately to keep the system stable. If the demand increases, the AI can increase supply. If the demand drops, the AI can reduce supply. This helps prevent power outages, reduces waste, and keeps the energy system running efficiently. Dynamic balancing of supply and demand is critical for energy providers to deliver reliable service while optimizing resources. **INCORRECT:** "The agents rely only on monthly billing data to manage power supply" is incorrect. Monthly billing data shows the total energy used over a long period, such as 30 days. This data is too old and slow for real-time management. By the time monthly data is available, the energy usage patterns have already changed. Relying only on monthly data would not allow the AI agents to react quickly to changes in demand, making this option incorrect. **INCORRECT:** "The agents generate static reports after peak usage periods" is incorrect. Static reports are summaries that are created after an event has already happened. While these reports can help analyze past usage, they do not help the AI manage power supply in real-time. The goal is to adjust power supply during usage, not after it. This makes static reporting unsuitable for real-time energy management. **INCORRECT:** "The agents require human approval before taking any action" is incorrect. Requiring human approval slows down the process. Real-time energy management needs to happen in seconds or milliseconds, faster than a human can review and approve decisions. AI agents are designed to act automatically based on real-time data to keep the system balanced without waiting for human intervention. **References:** https://aws.amazon.com/what-is/ai-agents Domain: Applications of Foundation Models --- #### 09. Which of the following best describes the key difference between few-shot prompting and zero-shot prompting in natural language processing? - Zero-shot prompting requires pre-training on task-specific data before use. - Few-shot prompting includes example inputs and outputs within the prompt, while zero-shot prompting does not provide examples. - Few-shot prompting applies reinforcement learning techniques, whereas zero-shot prompting does not. - Few-shot prompting is only supported by models that have been fine-tuned with custom datasets. **CORRECT:** "Few-shot prompting includes example inputs and outputs within the prompt, while zero-shot prompting does not provide examples" is the correct answer. Few-shot and zero-shot prompting are two techniques used to guide language models in performing specific tasks. The main difference lies in how much information is given in the prompt. In few-shot prompting, the prompt includes a few examples of how inputs should be mapped to outputs. This helps the model understand the task by learning from the examples directly in the prompt. On the other hand, zero-shot prompting provides only the instruction or question without examples. It assumes that the model already has enough general knowledge to generate a correct or relevant response without being shown any examples first. **INCORRECT:** "Few-shot prompting applies reinforcement learning techniques, whereas zero-shot prompting does not" is incorrect. Few-shot and zero-shot prompting are not based on reinforcement learning. They are techniques for designing prompts during inference (i.e., when the model is being used), not training. Reinforcement learning may be involved during the model training phase (e.g., RLHF – Reinforcement Learning from Human Feedback), but it is unrelated to few-shot vs. zero-shot prompting. **INCORRECT:** "Zero-shot prompting requires pre-training on task-specific data before use" is incorrect. Zero-shot prompting works without any task-specific training or examples. It leverages the general knowledge of the pre-trained model. You just ask a question or give an instruction, and the model responds based on what it has learned during pretraining — no task-specific fine-tuning is needed. **INCORRECT:** "Few-shot prompting is only supported by models that have been fine-tuned with custom datasets" is incorrect. Most large language models, like GPT or Claude, can support few-shot prompting without additional fine-tuning. The model's ability to interpret examples in the prompt is part of its general capabilities from pretraining, not from being fine-tuned on a specific dataset. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-engineering-guidelines.html Domain: Applications of Foundation Models --- #### 10. A customer service company is using an AI chatbot to handle user queries. However, The team is considering Amazon Augmented AI (A2I) to ensure quality and tone. What is the key benefit of using Amazon A2I? - A2I fine-tunes the model using reinforcement signals without requiring any human intervention. - A2I blocks model responses that violate sentiment constraints, using automated sentiment filters. - A2I enables human review of low-confidence chatbot responses to ensure accuracy and improve user experience. - A2I automatically re-trains the chatbot on real-time user inputs without needing labeled data. **CORRECT:** "A2I enables human review of low-confidence chatbot responses to ensure accuracy and improve user experience" is the correct answer. Amazon Augmented AI (A2I) makes it easy to implement human review for machine learning predictions, including responses from AI chatbots. In scenarios where a model's output may lack confidence or require judgment—such as ensuring tone, accuracy, or compliance—A2I allows human reviewers to step in. This is particularly useful for customer service applications where sensitive or complex responses must be checked to maintain high quality. With A2I, businesses can set up human review workflows triggered by confidence thresholds, flagged keywords, or business rules, ensuring that responses meet desired standards before being sent to users. This helps balance automation with human oversight, improving trust and user satisfaction. **INCORRECT:** "A2I automatically re-trains the chatbot on real-time user inputs without needing labeled data" is incorrect. A2I is not a training tool. It enables human review in real-time or batch workflows, but it does not retrain models automatically. Retraining would require a separate process using Amazon SageMaker or another ML pipeline. **INCORRECT:** "A2I blocks model responses that violate sentiment constraints, using automated sentiment filters" is incorrect. A2I doesn't apply automated sentiment filtering or block responses. Instead, it facilitates human review when needed. If sentiment analysis is required, it must be implemented as part of the workflow logic using services like Amazon Comprehend, not A2I. **INCORRECT:** "A2I fine-tunes the model using reinforcement signals without requiring any human intervention" is incorrect. This is misleading. A2I exists specifically to include human intervention. It is not a reinforcement learning tool. If your goal is to fine-tune a model automatically, you'd use other ML techniques such as RLHF or supervised learning, not A2I. **References:** https://aws.amazon.com/augmented-ai Domain: Guidelines for Responsible AI --- #### 11. A software company is developing generative AI applications for content generation and automated responses. To ensure accurate performance tracking and efficient deployment, the team must understand the key differences between model inference and model evaluation. Model inference is responsible for generating outputs based on input data, while model evaluation assesses the model's performance using specific metrics. Which of the following statements correctly summarizes the differences between model inference and model evaluation? (Select TWO.) - Model inference and model evaluation are interchangeable terms that describe the same process of generating AI-based content. - Model inference is the process of generating predictions or outputs using a trained model, whereas model evaluation measures the model's accuracy and effectiveness using predefined metrics. - Model inference is responsible for training new AI models, while model evaluation generates responses from the trained model. - Model evaluation involves testing a model's performance with various metrics, while model inference is used to produce real-time or batch outputs based on input data. - Model evaluation is only required during training, whereas model inference is only used during deployment. **CORRECT:** "Model inference is the process of generating predictions or outputs using a trained model, whereas model evaluation measures the model's accuracy and effectiveness using predefined metrics" is a correct answer. Model inference refers to using a trained machine learning or AI model to make predictions or generate responses based on new input data. This is what happens during deployment when the model is serving real users. Model evaluation, on the other hand, is a step used to assess how well a model performs by comparing its predictions to actual results using metrics such as accuracy, precision, recall, or F1 score. Evaluation is typically done during model development and testing to ensure the model meets performance goals before deployment. **CORRECT:** "Model evaluation involves testing a model's performance with various metrics, while model inference is used to produce real-time or batch outputs based on input data" is a correct answer. Model evaluation helps data scientists understand how well the model is doing by using performance metrics. These tests are often done on validation or test datasets. Model inference is the phase where the model is applied in the real world, either making predictions in real-time (e.g., chatbot replies) or in batch mode (e.g., generating weekly reports). Both processes are essential but serve different roles in the AI lifecycle. **INCORRECT:** "Model inference and model evaluation are interchangeable terms that describe the same process of generating AI-based content" is incorrect. Model inference and evaluation are two separate stages. Inference is about making predictions using the model, while evaluation is about checking how good those predictions are using metrics. **INCORRECT:** "Model evaluation is only required during training, whereas model inference is only used during deployment" is incorrect. While evaluation is mostly done during training, it can also be performed after deployment to monitor ongoing model performance. Similarly, inference can be tested during development as well as used in production. **INCORRECT:** "Model inference is responsible for training new AI models, while model evaluation generates responses from the trained model" is incorrect. Training is a separate process from inference. Model inference uses a trained model to make predictions. Evaluation does not generate responses—it measures how well the responses match the expected output. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/key-definitions.html Domain: Fundamentals of AI and ML --- #### 12. An AI startup is developing a document summarization tool using Amazon Bedrock. The tool will process long user-uploaded reports and generate concise summaries. As part of optimizing model behavior, the team is exploring various inference parameters, including Response length. What does the Response length parameter control in Amazon Bedrock? - It defines how diverse the model's responses will be by introducing randomness during token generation. - It sets the minimum or maximum number of tokens to control the length of the generated response. - It limits the number of user prompts the model can remember during a single interaction session. - It instructs the model to reference external documents when responses exceed a certain complexity. **CORRECT:** "It sets the minimum or maximum number of tokens to control the length of the generated response" is the correct answer. In Amazon Bedrock, the Response length parameter (commonly controlled using max_tokens) determines how long the generated output can be, measured in tokens (where one token is roughly 4 characters or ¾ of a word). This is important when you want to keep responses concise, such as in document summarization tools. For example, if you want the model to generate a summary that is short and to the point, setting a lower maximum token limit ensures the output stays within a desired length. This allows the AI to stay on-topic and avoid generating overly long or unnecessary content. This parameter is ideal when you need output length control, such as in customer-facing apps or content summaries. **INCORRECT:** "It defines how diverse the model's responses will be by introducing randomness during token generation" is incorrect. This describes the temperature or top-p parameter, not response length. These parameters control how creative or random the output is, not how long the output should be. **INCORRECT:** "It limits the number of user prompts the model can remember during a single interaction session" is incorrect. This refers to context window or context length, which defines how much prior input (prompt history) the model can consider. It is unrelated to controlling the length of the model's response. **INCORRECT:** "It instructs the model to reference external documents when responses exceed a certain complexity" is incorrect. This describes the behavior of retrieval-augmented generation (RAG) or Bedrock Agents with API calls, not the response length. The response length parameter only controls how long the output is, not whether external documents are used. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html Domain: Applications of Foundation Models --- #### 13. A public health research institute needs to analyze a collection of electronic health records to detect mentions of medication usage, medical conditions, and treatment plans in free-text format. Which AWS service offers pretrained models to support this requirement? - Amazon Comprehend Medical - Amazon Translate - Amazon Rekognition - Amazon Polly **CORRECT:** "Amazon Comprehend Medical" is the correct answer. Amazon Comprehend Medical is a natural language processing (NLP) service that uses machine learning to extract relevant medical information from unstructured clinical text. It identifies entities like medications, medical conditions, treatment procedures, and test results. For the public health research institute analyzing electronic health records (EHRs), this service is ideal because it offers pretrained models specifically designed for healthcare and life sciences data. The researchers can automatically detect mentions of diseases, treatments, dosage, and frequency without building or training custom models. This speeds up analysis and ensures compliance with healthcare standards. **INCORRECT:** "Amazon Polly" is incorrect. Amazon Polly is a text-to-speech (TTS) service that converts text into lifelike speech using deep learning. While useful for accessibility and voice-enabled apps, it doesn't analyze or understand medical content in text. **INCORRECT:** "Amazon Rekognition" is incorrect. Amazon Rekognition is used for analyzing images and videos. It can detect objects, faces, and inappropriate content but is not built to analyze or understand text—especially unstructured medical documents. Thus, it doesn't fit the requirement for analyzing EHR text. **INCORRECT:** "Amazon Translate" is incorrect. Amazon Translate is a neural machine translation service that supports language translation. While it helps in translating medical documents between languages, it doesn't extract or interpret specific medical terms or insights, which is essential for this scenario. **References:** https://aws.amazon.com/comprehend/medical https://docs.aws.amazon.com/comprehend-medical/latest/dev/comprehendmedical-welcome.html Domain: Applications of Foundation Models --- #### 14. Your organization is deploying an AI-powered loan approval engine trained on historical application data. During model evaluation, the data science team discovers that loan decisions disproportionately favor applicants from particular ZIP codes, which correlate with socioeconomic factors. You are tasked with refining the model pipeline to mitigate potential ethical risks and regulatory violations related to geographic and demographic bias. Which responsible AI strategy should you prioritize to proactively detect and mitigate geographic bias in the model while aligning with fairness and ethical AI development frameworks? - Incorporate SHAP-based explainability techniques to identify and exclude highly weighted geographic features from model input. - Deploy adversarial debiasing techniques post-training to reduce the model's reliance on sensitive attributes like geographic location. - Perform fairness-aware preprocessing to reweight training samples or balance subgroups by region before training the model. - Apply differential privacy to all applicant features, ensuring no single individual's data disproportionately influences the model's behavior. **CORRECT:** "Perform fairness-aware preprocessing to reweight training samples or balance subgroups by region before training the model" is the correct answer. Fairness-aware preprocessing is a proactive responsible AI strategy that helps reduce bias before the model is trained. It involves techniques like reweighting or resampling the data to ensure fair representation of different subgroups — in this case, applicants from different ZIP codes or regions. This helps prevent the model from learning biased patterns in historical data that correlate geographic location with loan approval. By balancing the dataset or assigning appropriate weights, the model can be trained in a way that reduces disparities in outcomes across regions. This approach aligns well with fairness and ethical AI principles, and it's especially useful when dealing with sensitive attributes that influence access to financial services. **INCORRECT:** "Incorporate SHAP-based explainability techniques to identify and exclude highly weighted geographic features from model input" is incorrect. SHAP (SHapley Additive exPlanations) helps explain model predictions and identify feature importance. It's a diagnostic tool — not a mitigation strategy by itself. Also, simply removing geographic features can result in proxy bias if other correlated features remain. **INCORRECT:** "Apply differential privacy to all applicant features, ensuring no single individual's data disproportionately influences the model's behavior" is incorrect. Differential privacy is great for protecting individual privacy, but it doesn't directly address fairness or bias in predictions across groups. It's more about data security than reducing geographic or demographic disparities. **INCORRECT:** "Deploy adversarial debiasing techniques post-training to reduce the model's reliance on sensitive attributes like geographic location" is incorrect. Adversarial debiasing is an advanced mitigation method but is typically applied during or after training, often with added model complexity. It's less proactive than fairness-aware preprocessing and may be harder to integrate in regulated environments without clearly explainable mechanisms. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/clarify-configure-processing-jobs.html https://aws.amazon.com/blogs/machine-learning/learn-how-amazon-sagemaker-clarify-helps-detect-bias Domain: Guidelines for Responsible AI --- #### 15. A healthcare provider is developing an AI tool to summarize patient notes and translate clinical reports into multiple languages. The development team is considering Transformer-based models to process unstructured text efficiently. Which statement best describes the Transformer model in this context? - A model that relies entirely on convolutional layers for processing sequential data - A rule-based system designed for structured data classification - A deep learning architecture that uses self-attention mechanisms to understand relationships between words in a sequence - A model limited to analyzing numerical time-series data only **CORRECT:** "A deep learning architecture that uses self-attention mechanisms to understand relationships between words in a sequence" is the correct answer. Transformers are advanced deep learning models specially designed to handle sequential data like text. Unlike traditional models, Transformers rely on a self-attention mechanism that helps the model understand the context and relationships between all words in a sentence, regardless of their position. This makes them especially powerful for tasks like summarizing patient notes or translating clinical documents. In the healthcare use case, a Transformer model can analyze complex, unstructured medical text and extract relevant information or translate it accurately while maintaining the context. This architecture forms the foundation of many state-of-the-art models like BERT, GPT, and T5, which are widely used in natural language processing (NLP) applications. **INCORRECT:** "A model that relies entirely on convolutional layers for processing sequential data" is incorrect. This describes Convolutional Neural Networks (CNNs), which are more common in image processing. They are not designed for understanding long-range dependencies in text. **INCORRECT:** "A rule-based system designed for structured data classification" is incorrect. This refers to traditional expert systems or rule-based AI, which don't learn from data and are not effective for complex language understanding tasks. **INCORRECT:** "A model limited to analyzing numerical time-series data only" is incorrect. This would describe models like ARIMA or certain LSTMs, but Transformers are designed for a wide range of text-based and sequential tasks—not just numerical time-series. **References:** https://aws.amazon.com/what-is/generative-ai Domain: Fundamentals of Generative AI --- #### 16. A healthcare provider is deploying an AI model to predict patient risk factors. To ensure transparency and explainability for regulatory compliance, they decided to use Amazon SageMaker Model Cards. Which of the following is a key feature of Amazon SageMaker Model Cards that makes it useful for this purpose? - Automatic hyperparameter tuning for improved model accuracy - Real-time prediction logging - Detailed documentation of the model's training data, parameters, and performance - Visual representation of model biases **CORRECT:** "Detailed documentation of the model's training data, parameters, and performance" is the correct answer. Amazon SageMaker Model Cards provide a comprehensive way to document critical information about a machine learning model, including the training data, parameters, and performance metrics. This detailed documentation is essential for transparency and explainability, especially in regulated industries like healthcare, where clear and accessible records of how a model works are required for compliance. SageMaker Model Cards help the healthcare provider maintain accountability and ensure the model meets regulatory standards by providing insights into its development and evaluation. **INCORRECT:** "Real-time prediction logging" is incorrect. SageMaker Model Cards focus on documenting the model's characteristics and do not handle real-time prediction logging. Real-time monitoring is a different aspect of model deployment, typically managed by other tools like SageMaker Model Monitor. **INCORRECT:** "Automatic hyperparameter tuning for improved model accuracy" is incorrect. Hyperparameter tuning is related to model optimization but is not a feature of SageMaker Model Cards. Model Cards are focused on documenting and explaining the model, not on improving accuracy. **INCORRECT:** "Visual representation of model biases" is incorrect. SageMaker Clarify is a tool for detecting bias, SageMaker Model Cards provide detailed documentation but do not specifically offer visual representations of biases. Their primary purpose is to document the model's lifecycle and characteristics. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/model-cards.html Domain: Guidelines for Responsible AI --- #### 17. A data science team is using Amazon SageMaker Data Wrangler to prepare data for a machine learning model. They need to split their dataset into training, validation, and test sets. Which of the following statements BEST describes how Data Wrangler facilitates this process? - Data Wrangler automatically splits the dataset using a fixed 70/20/10 ratio without any customization options. - Data Wrangler only supports splitting data into training and test sets, requiring additional tools for creating a validation set. - Data Wrangler allows users to define custom split ratios and provides built-in transformations to create the necessary data subsets. - Data Wrangler relies on external services to perform data splitting, and users must manually integrate the results. **CORRECT:** "Data Wrangler allows users to define custom split ratios and provides built-in transformations to create the necessary data subsets" is the correct answer. Amazon SageMaker Data Wrangler is a tool that simplifies the process of preparing data for machine learning. It provides a visual interface that enables data scientists and analysts to import, clean, transform, and analyze data from various sources without writing extensive code. Data Wrangler supports integrations with services like Amazon S3, Athena, Redshift, and Snowflake, and it streamlines the entire data preparation workflow, including handling missing values, encoding categorical variables, and feature engineering. It also allows exporting the entire data preparation workflow as a pipeline that can be used in SageMaker training jobs or inference endpoints. One of its powerful features is the ability to split datasets into training, validation, and test sets. Users can specify custom ratios (for example, 80/10/10 or 70/15/15), making the process flexible to suit the needs of various ML workflows. Data Wrangler includes a built-in "Split data" transformation that lets you choose how many splits you want and what percentage of the dataset each split should contain. This built-in functionality helps streamline the data preparation process and removes the need to perform these operations manually using code. **INCORRECT:** "Data Wrangler automatically splits the dataset using a fixed 70/20/10 ratio without any customization options" is incorrect. Data Wrangler does not restrict users to a fixed split ratio. It provides flexibility to define your own custom ratios depending on your project's needs. You can also choose whether to randomize the split or keep it sequential, which adds more control to the data preparation process. **INCORRECT:** "Data Wrangler only supports splitting data into training and test sets, requiring additional tools for creating a validation set" is incorrect. Data Wrangler supports multiple splits, not just training and test. It allows users to create three or more subsets easily, including a validation set, all within the tool's interface. There's no need for external tools to create a validation set. **INCORRECT:** "Data Wrangler relies on external services to perform data splitting, and users must manually integrate the results" is incorrect. Data Wrangler performs data splitting within the tool itself. It does not depend on any external service for this function. Once the splits are created, they can be used in downstream SageMaker pipelines without requiring manual integration. **References:** https://aws.amazon.com/blogs/machine-learning/create-train-test-and-validation-splits-on-your-data-for-machine-learning-with-amazon-sagemaker-data-wrangler Domain: Fundamentals of AI and ML --- #### 18. A fintech company is using Amazon Bedrock to generate financial reports from internal data. The team wants to ensure their data privacy obligations are met and clarify AWS's responsibilities in their generative AI pipeline. Which of the following best describes the shared responsibility model in this use case? - The company is responsible for securing all customer data and access permissions; AWS secures the Bedrock infrastructure and APIs. - AWS manages the content generated by the models, while the company is responsible for training and tuning them. - The company must secure the foundation model source code used by Bedrock and store it in a VPC environment. - AWS handles all security configurations, including access control and encryption of client data sent to the models. **CORRECT:** "The company is responsible for securing all customer data and access permissions; AWS secures the Bedrock infrastructure and APIs" is the correct answer. This option best reflects the AWS Shared Responsibility Model. In this shared responsibility model, AWS is responsible for "security of the cloud," which includes the physical infrastructure, managed services like Amazon Bedrock, and APIs. The customer (in this case, the fintech company) is responsible for "security in the cloud." This includes managing access controls (like IAM roles), encrypting sensitive data before sending it to Bedrock, and ensuring that only authorized users can access the system. Since the company is using Bedrock to process sensitive financial data, it must ensure that proper permissions, encryption settings, and audit logging are configured to meet data privacy obligations. AWS ensures that Bedrock itself is secure and reliable, but the customer's data security remains their responsibility. **INCORRECT:** "AWS handles all security configurations, including access control and encryption of client data sent to the models" is incorrect. While AWS provides tools for encryption and access control (e.g., KMS, IAM), the customer is responsible for configuring and managing these tools. AWS does not automatically enforce access policies or encrypt all data without customer action. **INCORRECT:** "AWS manages the content generated by the models, while the company is responsible for training and tuning them" is incorrect. This is misleading. With Amazon Bedrock, customers typically use pre-trained foundation models provided by AWS or third-party providers. The content generated by the models is managed by the customer, not AWS, especially when it comes to compliance, output filtering, and data governance. **INCORRECT:** "The company must secure the foundation model source code used by Bedrock and store it in a VPC environment" is incorrect. This is incorrect because the foundation models used in Bedrock are hosted and managed entirely by AWS. Customers do not have access to the underlying source code and do not need to store it. Instead, they use APIs to interact with these models. VPC configurations can secure how the Bedrock APIs are accessed, but model source code management is not the customer's responsibility. **References:** https://aws.amazon.com/compliance/shared-responsibility-model Domain: Security, Compliance, and Governance for AI Solutions --- #### 19. A government agency is building a policy summarization tool using Amazon Bedrock. They need to verify that the generated summaries are coherent, fact-based, and aligned with original documents. Which strategies support effective model evaluation on Amazon Bedrock? (Select TWO) - Use prompt chaining to improve summarization logic during evaluation. - Evaluate outputs using ROUGE metrics to compare to human-written policy summaries. - Enable Amazon Translate for multilingual summarization accuracy testing. - Use human evaluations to compare model outputs against reference summaries. - Deploy the model in production and gather user feedback before any formal testing. **CORRECT:** "Evaluate outputs using ROUGE metrics to compare to human-written policy summaries" is a correct answer. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is a standard metric used to evaluate summarization models by comparing generated summaries with high-quality, human-written reference summaries. It measures how much content from the reference is included in the model's summary, using overlaps in words or phrases (like unigrams, bigrams, and longer sequences). For a government policy summarization tool, using ROUGE helps assess whether the model is capturing the essential ideas from original documents. This automated metric is useful for evaluating content similarity and relevance at scale, and it is well-supported in evaluation workflows on Amazon Bedrock. **CORRECT:** "Use human evaluations to compare model outputs against reference summaries" is a correct answer. Human evaluation is a highly reliable method to assess generated summaries for coherence, factual accuracy, and alignment with original documents—qualities that are difficult to measure with automated metrics alone. In Amazon Bedrock, users can perform human-in-the-loop evaluations where evaluators review the outputs and rate them based on quality criteria. This is especially important in sensitive domains like government policies, where understanding context and intent is critical. Human evaluations ensure that summaries are not only factually accurate but also readable and aligned with intended goals. **INCORRECT:** "Use prompt chaining to improve summarization logic during evaluation" is incorrect. Prompt chaining is a technique used to build complex workflows or enhance model reasoning by passing outputs from one prompt as inputs to the next. While useful in improving the logic or structure of summarization, prompt chaining is not an evaluation strategy. It's more of a development or prompting technique. It doesn't directly help assess how well the summary aligns with the original policy documents. **INCORRECT:** "Deploy the model in production and gather user feedback before any formal testing" is incorrect. Deploying a model in production without formal testing is not a recommended practice. Evaluation should be done in pre-production environments to ensure the model is producing accurate, safe, and high-quality outputs. Relying only on user feedback after deployment can lead to issues like the release of misleading or low-quality summaries, which can be risky—especially in government or policy applications. **INCORRECT:** "Enable Amazon Translate for multilingual summarization accuracy testing" is incorrect. Amazon Translate converts text between languages, not an evaluation tool. While translation might be useful if you're working with multilingual documents, it does not directly measure the quality or accuracy of a summary. Evaluation should focus on how well the summary captures meaning, which is best done using metrics like ROUGE or through human evaluation. **References:** https://aws.amazon.com/blogs/machine-learning/evaluate-the-text-summarization-capabilities-of-llms-for-enhanced-decision-making-on-aws Domain: Applications of Foundation Models --- #### 20. An insurance company receives thousands of handwritten and printed claim forms submitted by customers. To streamline processing, the company wants to automate the extraction of key fields such as claim amount, policy number, and date of claim from these scanned documents. After extracting the data, the company also wants to build a machine learning model to analyze the extracted information and predict the likelihood of fraudulent claims. Which of the following AWS services are best suited to implement this end-to-end solution? (Select TWO.) - Amazon Transcribe - Amazon SageMaker - Amazon Textract - Amazon Comprehend - Amazon Rekognition **CORRECT:** "Amazon Textract" is a correct answer. Amazon Textract is a fully managed machine learning service by AWS that automatically extracts printed or handwritten text, tables, forms, and other structured data from scanned documents. Unlike traditional OCR (optical character recognition) tools, Textract not only detects text but also understands the layout and relationships between fields, such as key-value pairs in forms or rows and columns in tables. This makes it ideal for automating document processing tasks in industries like insurance, finance, and healthcare. **CORRECT:** "Amazon SageMaker" is a correct answer. Amazon SageMaker is a comprehensive machine learning service that enables developers and data scientists to build, train, and deploy ML models at scale. It provides tools for every step of the ML lifecycle, including data labeling, model building with built-in algorithms or custom code, training with managed infrastructure, automatic model tuning, and hosting for real-time predictions. SageMaker is used across industries to solve problems like fraud detection, customer churn, demand forecasting, and more. Amazon Textract is ideal for this use case because it can accurately extract structured data such as claim amounts, policy numbers, and dates from scanned insurance forms, whether they are typed or handwritten. This extracted data can then be fed into Amazon SageMaker, which provides the tools needed to build and deploy a machine learning model that analyzes historical claim data and predicts the likelihood of fraud. Together, these services automate and enhance the entire workflow—from data extraction to intelligent decision-making—making them the best-fit choices for the insurance company's needs. **INCORRECT:** "Amazon Transcribe" is incorrect. Amazon Transcribe is a speech-to-text service that converts audio input into text. It is useful for transcribing customer service calls or voice notes but is not suitable for processing scanned documents or analyzing structured data. It does not help in this use case. **INCORRECT:** "Amazon Rekognition" is incorrect. Amazon Rekognition is used for analyzing images and videos to detect objects, people, text, and inappropriate content. While it can recognize text in images, it is not designed for structured form or handwriting extraction like Textract. So it is not suitable for extracting detailed claim information from scanned forms. **INCORRECT:** "Amazon Comprehend" is incorrect. Amazon Comprehend is a natural language processing (NLP) service used to analyze and extract insights from unstructured text, such as sentiment analysis, entity recognition, and topic modeling. It is not designed for processing scanned images or extracting form data, which is the primary requirement in this use case. **References:** https://aws.amazon.com/textract https://aws.amazon.com/sagemaker Domain: Applications of Foundation Models --- #### 21. A financial auditing firm has integrated a large language model (LLM) into its workflow to automatically generate compliance summaries from lengthy contractual documents. This tool is intended to reduce manual effort and increase consistency across reports. However, the compliance team has raised concerns that the model might introduce unintended bias by favoring specific business terms or contractual language patterns, which could influence downstream legal interpretations or decisions. To address these concerns, the firm wants to evaluate the model for potential bias in a way that requires minimal administrative overhead while still providing meaningful insights into fairness and representational balance. What is the most suitable approach to evaluate the model for bias with minimal administrative effort? - Deploy the model to a small group of users and collect structured feedback on perceived bias through targeted surveys. - Use pre-built prompt datasets in platforms like AWS Bedrock that are specifically designed to evaluate model outputs for bias and fairness. - Continuously fine-tune the model using recent responses from users with diverse professional and cultural backgrounds. - Conduct manual audits by reviewing model-generated summaries for bias across different document types and scenarios. **CORRECT:** "Use pre-built prompt datasets in platforms like AWS Bedrock that are specifically designed to evaluate model outputs for bias and fairness" is the correct answer. AWS Bedrock provides evaluation tools and pre-built prompt datasets that are specifically curated to assess foundational model outputs for bias, toxicity, and fairness. These tools are designed to reduce the manual overhead of bias evaluation by offering consistent, repeatable testing against standard prompts. Since the firm is looking for a solution that minimizes administrative effort while still offering meaningful insight, leveraging AWS Bedrock's evaluation features is ideal. It allows the compliance team to detect biased behavior without building complex custom pipelines, thus saving time and resources. This approach supports responsible AI practices and aligns with AWS's recommendations for model auditing. **INCORRECT:** "Conduct manual audits by reviewing model-generated summaries for bias across different document types and scenarios" is incorrect. While manual audits can be thorough, they require significant administrative effort and time. This approach does not scale well and is prone to inconsistencies, especially when multiple reviewers are involved. It also lacks the repeatability and automation needed for efficient model evaluation in a production setting. **INCORRECT:** "Deploy the model to a small group of users and collect structured feedback on perceived bias through targeted surveys" is incorrect. This method introduces user-based evaluation, which can be valuable for real-world insights, but it's not ideal for minimal administrative effort. Designing, distributing, and analyzing surveys requires coordination, and feedback may vary due to subjective interpretations of bias, which can complicate results. **INCORRECT:** "Continuously fine-tune the model using recent responses from users with diverse professional and cultural backgrounds" is incorrect. Fine-tuning can improve model behavior over time but is not a lightweight or evaluation-focused solution. It's a training approach that requires labeled data, model retraining, and validation, making it more complex and resource-intensive. It also does not directly assess existing bias, but rather attempts to correct it. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/model-evaluation-prompt-datasets.html Domain: Guidelines for Responsible AI --- #### 22. A self-driving car company is developing AI-powered vision systems to detect pedestrians, road signs, and obstacles. The engineering team must differentiate between computer vision and image processing techniques to implement the best approach for each aspect of the vehicle's perception system. Which of the following statements correctly describes the differences between computer vision and image processing? - Image processing can make autonomous vehicles interpret street signs without any additional AI-based recognition models. - Computer vision is designed for high-level understanding, such as object detection and scene recognition, while image processing focuses on modifying images through techniques like edge detection. - Computer vision relies only on pixel modifications and does not require machine learning models. - Image processing and computer vision are identical and interchangeable in autonomous vehicle applications. **CORRECT:** "Computer vision is designed for high-level understanding, such as object detection and scene recognition, while image processing focuses on modifying images through techniques like edge detection" is the correct answer. Computer vision and image processing are closely related but serve different purposes. Computer vision involves enabling machines to understand and interpret visual data, such as detecting objects (like pedestrians and signs) and recognizing scenes. It often relies on machine learning and deep learning models. On the other hand, image processing focuses on improving image quality or extracting low-level features (like edges, color corrections, or filters) without necessarily understanding the content. These techniques are typically used before feeding the data into a computer vision model to enhance accuracy. **INCORRECT:** "Image processing and computer vision are identical and interchangeable in autonomous vehicle applications" is incorrect. While they work together, image processing is about improving or manipulating images, while computer vision extracts meaning from images. They are not interchangeable. **INCORRECT:** "Image processing can make autonomous vehicles interpret street signs without any additional AI-based recognition models" is incorrect. Image processing alone cannot recognize or interpret signs. It can enhance the image, but AI models in computer vision are required to identify and understand what the sign means. **INCORRECT:** "Computer vision relies only on pixel modifications and does not require machine learning models" is incorrect. Modern computer vision heavily relies on machine learning and deep learning models (e.g., CNNs) to detect, classify, and interpret objects in images—not just simple pixel changes. **References:** https://aws.amazon.com/what-is/computer-vision Domain: Fundamentals of AI and ML --- #### 23. You are working for a global e-commerce company tasked with building an AI-powered customer service chatbot. The chatbot must understand and respond intelligently in multiple languages, handle a variety of customer intents, and generate human-like, context-aware answers. The solution must support custom prompts, integrate easily with existing AWS services, and scale across multiple regions. Which solution best supports these requirements? - Build the chatbot using Amazon Comprehend and Amazon Polly, translating text with Amazon Translate and analyzing sentiment before generating responses through custom rules. - Use Amazon Lex combined with Amazon Translate to handle multilingual input and generate responses using AWS Lambda functions for business logic. - Train a custom transformer-based NLP model on Amazon SageMaker using multilingual training data and deploy the model via an endpoint with inference logic written in Python. - Deploy a foundation model via Amazon Bedrock (e.g., Anthropic Claude or Amazon Titan), and use its multilingual capabilities with prompt engineering to generate intelligent, natural responses across multiple languages. **CORRECT:** "Deploy a foundation model via Amazon Bedrock (e.g., Anthropic Claude or Amazon Titan), and use its multilingual capabilities with prompt engineering to generate intelligent, natural responses across multiple languages" is the correct answer. Amazon Bedrock is a fully managed AWS service that provides access to high-performing foundation models (FMs) from leading AI providers such as Anthropic (Claude), Amazon (Titan), and others. These models support multiple languages, understand diverse customer intents, and generate human-like responses with advanced context handling. With prompt engineering, you can customize the chatbot's behavior without training your own models. Bedrock also integrates easily with other AWS services (like Lambda, S3, and API Gateway), supports multi-region deployment, and automatically scales to handle varying workloads. This makes it ideal for building sophisticated AI-powered multilingual chatbots that are fast to deploy and easy to manage. **INCORRECT:** "Use Amazon Lex combined with Amazon Translate to handle multilingual input and generate responses using AWS Lambda functions for business logic" is incorrect. Amazon Lex is designed for building rule-based chatbots with predefined intents and slots. While combining it with Translate allows multilingual input handling, it lacks deep contextual understanding and advanced natural language generation like foundation models offer. It's best for simpler bots, not complex, human-like conversations. **INCORRECT:** "Build the chatbot using Amazon Comprehend and Amazon Polly, translating text with Amazon Translate and analyzing sentiment before generating responses through custom rules" is incorrect. Comprehend and Polly are powerful services for language analysis and text-to-speech, but they are not designed to generate intelligent chatbot responses. This setup requires heavy custom logic, lacks natural contextual response capabilities, and doesn't scale well across languages and use cases compared to using a foundation model. **INCORRECT:** "Train a custom transformer-based NLP model on Amazon SageMaker using multilingual training data and deploy the model via an endpoint with inference logic written in Python" is incorrect. While SageMaker is powerful for training and deploying custom models, this approach requires significant time, data, and ML expertise. It is costly and complex for chatbot development, especially when foundation models via Bedrock already provide advanced multilingual NLP capabilities out of the box. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html https://aws.amazon.com/blogs/machine-learning/multilingual-content-processing-using-amazon-bedrock-and-amazon-a2i Domain: Applications of Foundation Models --- #### 24. Select and order the following Amazon SageMaker training options from the LOWEST to the HIGHEST training cost. Each option should be selected one time. (Select and order THREE.) Note: Select only the correct options, as the type of "Ordering" question is not supported here. - Spot Training - On-Demand Training - Multi-GPU Distributed Training **CORRECT:** The correct order from LOWEST to HIGHEST training cost is: 1. Spot Training 2. On-Demand Training 3. Multi-GPU Distributed Training Spot Training is the most cost-effective option for training machine learning models in Amazon SageMaker. It uses spare AWS computing capacity, which can be interrupted and resumed later. Because it takes advantage of discounted instances, Spot Training can reduce training costs by up to 90% compared to On-Demand Training. However, it may take longer to complete if interruptions occur. On-Demand Training provides dedicated computing resources without interruptions, making it more expensive than Spot Training. It charges users for the exact amount of compute time used and ensures consistent training without delays. This option is suitable for time-sensitive workloads where interruptions are not acceptable. Multi-GPU Distributed Training is the most expensive option because it uses multiple GPUs across multiple instances to speed up model training. While this reduces training time significantly, it also increases the overall cost due to the high pricing of GPU instances. This method is commonly used for training deep learning models on large datasets. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/model-managed-spot-training.html https://aws.amazon.com/sagemaker-ai/pricing Domain: Applications of Foundation Models --- #### 25. A company using AWS to build a generative AI application needs to decide between using pre-built models or custom models. While custom models provide flexibility, the company is concerned about cost. What is a key cost-related factor to consider when choosing custom models over pre-trained ones? - Custom models require significant compute resources for training, increasing cost. - Custom models offer fixed pricing, reducing overall cost. - Pre-trained models are less scalable, which increases long-term costs. - Custom models automatically reduce operational costs due to better optimization. **CORRECT:** "Custom models require significant compute resources for training, increasing cost" is the correct answer. When you choose to use custom models, one of the major cost-related factors is the significant compute resources needed for training the model. Custom models need large amounts of data and computational power, such as using GPUs or specialized hardware like AWS's EC2 instances, to fine-tune the model. Training these models often takes time and requires ongoing monitoring and optimization, which increases the overall cost. Compared to pre-trained models, which are ready to use out-of-the-box, the custom model path can be much more expensive due to these resource-intensive training needs. **INCORRECT:** "Custom models offer fixed pricing, reducing overall cost" is incorrect. Custom models do not offer fixed pricing; the costs are based on the resources used for training and deployment. The flexibility and complexity of custom models often result in higher costs. **INCORRECT:** "Pre-trained models are less scalable, which increases long-term costs" is incorrect. Pre-trained models are highly scalable and can handle increased workloads without significant cost changes, making them cost-effective over time. **INCORRECT:** "Custom models automatically reduce operational costs due to better optimization" is incorrect. Custom models can be optimized, but they do not automatically reduce operational costs. In fact, they usually increase costs due to training and maintenance needs. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/canvas-build-model.html Domain: Fundamentals of Generative AI --- #### 26. A retail company is developing an AI agent that integrates a generative AI model with its inventory management systems. The agent needs to retrieve product details, update stock quantities, and generate order summaries in response to user queries. What is the primary function of this AI agent? - Serving as a middleware to interact with backend systems - Expanding the underlying generative model's architecture - Archiving historical inventory data for reporting - Generating synthetic data for model training **CORRECT:** "Serving as a middleware to interact with backend systems" is the correct answer. An AI agent is an intelligent software system designed to autonomously perceive its environment, interpret data, and take actions to achieve specific goals. It can interact with users, digital systems, or the physical world, using techniques such as machine learning, natural language processing, and reasoning. AI agents are capable of making decisions, learning from experience, and adapting their behavior based on context. They are used in various domains including customer service, robotics, logistics, and healthcare. In modern applications, AI agents often connect to APIs, databases, or other tools to retrieve information, perform tasks, and provide intelligent responses or actions. In this scenario, the generative AI model is not just generating text, but is also retrieving product details, updating stock levels, and generating order summaries—functions that require interacting with backend APIs or databases. This makes the AI agent a smart interface or middleware layer that bridges natural language interactions with system-level tasks, providing a seamless user experience. **INCORRECT:** "Generating synthetic data for model training" is incorrect. This involves creating fake but realistic data to improve or expand training datasets. This is not the task described. The AI agent is performing live actions, not generating training data. **INCORRECT:** "Archiving historical inventory data for reporting" is incorrect. Archiving and reporting are data storage and analysis tasks. They are not interactive and do not involve real-time responses to user queries or updates to systems. Hence, this is not the agent's main function. **INCORRECT:** "Expanding the underlying generative model's architecture" is incorrect. This refers to changing or scaling the internal design of the AI model itself, such as adding layers or modifying algorithms. This is a development task and has nothing to do with how the agent interacts with inventory systems. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html Domain: Applications of Foundation Models --- #### 27. A retail company is developing a generative AI chatbot for customer service. They are concerned about maintaining the responsibility and safety of their AI models. Which AWS infrastructure feature best supports these concerns? - AWS provides tools like Amazon SageMaker Clarify and Model Monitor to help detect and mitigate biases, ensuring responsible AI usage. - AWS infrastructure is not responsible for the ethical concerns of AI model deployment. - AWS guarantees that all models deployed through its services are free from bias by default. - AWS leaves it to the users to monitor bias and safety without providing specific tools. **CORRECT:** "AWS provides tools like Amazon SageMaker Clarify and Model Monitor to help detect and mitigate biases, ensuring responsible AI usage" is the correct answer. AWS offers tools like Amazon SageMaker Clarify to help developers detect and mitigate bias in their AI models. This is critical for companies concerned about maintaining responsible and safe AI usage. SageMaker Clarify provides insights into model predictions to ensure fairness, and Model Monitor helps continuously track the behavior of AI models in production, allowing for proactive management of any safety or ethical concerns. These tools ensure that AI models are deployed responsibly and that any issues can be addressed in real-time. **INCORRECT:** "AWS leaves it to the users to monitor bias and safety without providing specific tools" is incorrect. AWS offers specific tools like SageMaker Clarify and Model Monitor to help with bias detection and model safety. **INCORRECT:** "AWS infrastructure is not responsible for the ethical concerns of AI model deployment" is incorrect. AWS provides tools and services to help users address ethical concerns, even though the responsibility is shared between AWS and the user. **INCORRECT:** "AWS guarantees that all models deployed through its services are free from bias by default" is incorrect. While AWS provides tools to detect and mitigate bias, it does not guarantee that all models are free from bias by default. Users need to actively monitor and manage bias. **References:** https://aws.amazon.com/sagemaker/clarify https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor.html Domain: Fundamentals of Generative AI --- #### 28. A prestigious photography contest has found evidence that some participants are submitting images enhanced or entirely generated by AI tools. Contest organizers aim to detect and disqualify such submissions to uphold fairness and genuine creative expression. What key concern is the photography contest primarily addressing by detecting AI-generated images? - Mitigating model hallucination in AI-generated text descriptions of photographs. - Protecting copyrighted material from being redistributed without permission. - Ensuring fair competition and genuine creativity among participants. - Preventing malware embedded in image files from compromising contest systems. **CORRECT:** "Ensuring fair competition and genuine creativity among participants" is the correct answer. This option directly reflects the main concern of the photography contest. The organizers want to make sure all submitted images are created by real photographers, not enhanced or generated by AI tools. This ensures the integrity of the competition, where human creativity and skill are judged—not machine-generated content. Detecting AI-generated images helps prevent unfair advantages and maintains the contest's reputation for valuing authentic artistic expression. **INCORRECT:** "Protecting copyrighted material from being redistributed without permission" is incorrect. While copyright is a valid issue in digital media, this is not the primary concern here. The contest is focusing on how the images were created (human vs. AI), not whether someone has reused copyrighted content. **INCORRECT:** "Preventing malware embedded in image files from compromising contest systems" is incorrect. Although security is always important, this is unrelated to the ethics and fairness aspect that the question highlights. The focus is on creative authenticity, not cybersecurity threats. **INCORRECT:** "Mitigating model hallucination in AI-generated text descriptions of photographs" is incorrect. Model hallucination refers to AI generating incorrect or misleading information—usually in text responses. This is not relevant to a photography contest focused on image submissions. **References:** https://www.amazon.science/blog/responsible-ai-in-the-generative-era Domain: Guidelines for Responsible AI --- #### 29. Your company runs an online marketplace that allows users to upload product images. Management is exploring the use of AWS AI services to automate image analysis for better content moderation and product categorization. You are asked to evaluate whether Amazon Rekognition can support these use cases. Which of the following tasks can Amazon Rekognition perform when analyzing images? (Select TWO.) - Extracting facial attributes such as emotions or age range - Performing sentiment analysis based on image content - Identifying inappropriate or unsafe content in images - Detecting objects such as people, cars, or furniture - Generating product descriptions based on image content **CORRECT:** "Detecting objects such as people, cars, or furniture" is a correct answer. Amazon Rekognition can analyze images and detect a wide variety of objects, such as people, cars, furniture, animals, and more. This capability helps businesses build automation features like product categorization, inventory management, or scene understanding. For an online marketplace, this means uploaded product images can be analyzed to automatically identify the type of object, improving the search experience and product listings. This makes it the right choice for supporting automated product categorization. **CORRECT:** "Identifying inappropriate or unsafe content in images" is a correct answer. Amazon Rekognition also includes content moderation features that can detect inappropriate or unsafe content, such as nudity, violence, or offensive material. This helps businesses maintain a safe and professional environment for their users. In an online marketplace where users upload product images, this feature helps automate content moderation to ensure that all uploaded images meet community standards and legal requirements. **INCORRECT:** "Generating product descriptions based on image content" is incorrect. Amazon Rekognition can detect objects in images but does not generate full product descriptions. Product description generation typically requires natural language generation services like Amazon Bedrock with foundation models or Amazon SageMaker with custom models. Therefore, this is not the correct option. **INCORRECT:** "Performing sentiment analysis based on image content" is incorrect. Sentiment analysis evaluates text to determine whether the sentiment is positive, negative, or neutral. While Rekognition can analyze facial expressions, sentiment analysis is generally used on text data, not images. AWS services like Amazon Comprehend are designed for text sentiment analysis, not Rekognition. **INCORRECT:** "Extracting facial attributes such as emotions or age range" is incorrect. While Amazon Rekognition does support facial analysis to detect attributes like emotions, age range, or gender, this is typically used for face-based applications such as user verification or audience analysis, not general product image analysis. In the context of your online marketplace focused on product images, this is not the most relevant capability. **References:** https://docs.aws.amazon.com/rekognition/latest/dg/labels.html https://docs.aws.amazon.com/rekognition/latest/dg/moderation.html Domain: Fundamentals of AI and ML --- #### 30. You are responsible for implementing fine-grained access control for your AI system to ensure that only authorized users can interact with specific models and datasets. Which approach best aligns with AWS's least-privilege security principle? - Use a single IAM policy for all users and services. - Define IAM roles with resource-specific permissions. - Assign IAM roles with wide permissions across all resources. - Implement resource-based policies for all AI models. **CORRECT:** "Define IAM roles with resource-specific permissions" is the correct answer. AWS Identity and Access Management (IAM) roles are entities that define a set of permissions for making AWS service requests. IAM roles are used to grant specific permissions to AWS resources without needing to share credentials. They can be assumed by users, applications, or AWS services, enabling controlled access to resources. Roles are useful for delegating access to resources securely, with customizable policies that define actions, conditions, and resource scopes. They are essential for implementing least-privilege access and managing security in cloud environments, ensuring only authorized entities can access and interact with specific resources. Defining IAM roles with resource-specific permissions aligns best with AWS's least-privilege security principle. This approach ensures that each role has only the permissions necessary to perform specific actions on designated resources. By using fine-grained permissions, you can minimize the risk of unauthorized access and reduce the impact of potential security breaches. AWS recommends using least-privilege access to limit the scope of permissions granted to each user, service, or application, improving overall security. Implementing resource-specific permissions also helps in better auditing and controlling access across your AI system. **INCORRECT:** "Assign IAM roles with wide permissions across all resources" is incorrect. This approach violates the least-privilege principle because it grants users more permissions than they need, increasing the risk of unauthorized actions or security breaches. **INCORRECT:** "Use a single IAM policy for all users and services" is incorrect. Using a single policy for all users and services lacks granularity, making it difficult to enforce the least-privilege principle. It can lead to unnecessary permissions being granted. **INCORRECT:** "Implement resource-based policies for all AI models" is incorrect. While resource-based policies are useful, they are not the best choice for managing fine-grained access at a user or service level. They work better in combination with IAM roles and policies. **References:** https://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 31. An educational technology company uses a generative AI tutor app. To protect it from malicious prompt attacks that trick the model into revealing confidential data, which action is most effective? - Use prompt templates to teach the LLM to detect and handle suspicious prompts. - Expand model input size to accommodate complex user queries. - Allow open-ended prompts to foster creativity and flexibility in responses. - Rely on human moderators to review every response in real-time. **CORRECT:** "Use prompt templates to teach the LLM to detect and handle suspicious prompts" is the correct answer. Prompt templates are structured inputs that guide the LLM on how to respond safely and appropriately. By designing templates with clear instructions and rules, developers can help the model recognize suspicious or harmful prompts and respond cautiously. This method creates a controlled interaction space, making it harder for attackers to trick the model into revealing confidential information. Templates often include safety instructions like "If the prompt asks for private data, decline to answer politely." AWS best practices recommend prompt engineering and templating to increase model robustness and reduce the risk of prompt injection attacks. **INCORRECT:** "Allow open-ended prompts to foster creativity and flexibility in responses" is incorrect. Open-ended prompts encourage diverse and imaginative responses, which is great for creativity but dangerous for security. They leave the model vulnerable to prompt injection attacks because there are fewer boundaries on what the model can output. Open-ended prompting is not a security measure. **INCORRECT:** "Rely on human moderators to review every response in real-time" is incorrect. Human moderation can catch many issues but is not scalable or fast enough for real-time applications like AI tutors. It also introduces operational challenges and delays. While human review is useful for audit and escalation, it is not the primary defense mechanism against malicious prompts. **INCORRECT:** "Expand model input size to accommodate complex user queries" is incorrect. Increasing the input size allows users to submit longer queries, but does not add any security protections. In fact, longer inputs could make it easier for attackers to embed hidden prompts. This option does not help detect or prevent malicious behavior. **References:** https://docs.aws.amazon.com/prescriptive-guidance/latest/llm-prompt-engineering-best-practices/best-practices.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 32. When deploying a generative AI model in a business application, which challenge might arise due to the model's non-deterministic nature? - Lack of multilingual support - Inconsistent responses to the same input - Limited scalability - High computational costs **CORRECT:** "Inconsistent responses to the same input" is the correct answer. Generative AI models are often non-deterministic, meaning they can generate different outputs for the same input each time. This variability can be a challenge when deploying these models in business applications, as inconsistent responses may confuse users or provide unreliable results. For example, in customer support chatbots or automated code generation, consistency is crucial. To mitigate this, businesses can adjust the model's temperature settings or apply post-processing steps to ensure more consistent outputs. **INCORRECT:** "Lack of multilingual support" is incorrect. While some generative AI models may not have multilingual capabilities by default, this is a separate challenge. Non-deterministic behavior does not directly affect whether a model supports multiple languages. **INCORRECT:** "High computational costs" is incorrect. Generative AI models, especially large ones, can be expensive to run due to their computational needs. However, this issue is related to model size and complexity, not the non-deterministic nature of the model. **INCORRECT:** "Limited scalability" is incorrect. Scalability refers to a model's ability to handle increasing workloads. This issue is influenced by the infrastructure and design but is not directly tied to the non-deterministic nature of generative AI models. **References:** https://aws.amazon.com/what-is/generative-ai Domain: Fundamentals of Generative AI --- #### 33. An autonomous drone company is deploying machine learning models to predict the shortest possible routes during live operations and also to analyze bulk telemetry data collected periodically from thousands of flights. The company needs to understand the appropriate inference methods for these scenarios. Which statement accurately describes the key differences between real-time inference and batch inference? - Batch inference offers lower accuracy than real-time inference due to its inability to adapt to live data. - Real-time inference is optimized for large-scale data processing over long periods, while batch inference handles time-sensitive predictions during operations. - Real-time inference is more cost-effective than batch inference because it processes large amounts of data at once. - Real-time inference provides an immediate response with low latency, while batch inference processes large volumes of data with higher latency. **CORRECT:** "Real-time inference provides an immediate response with low latency, while batch inference processes large volumes of data with higher latency" is the correct answer. Real-time inference and batch inference are two common methods for running machine learning predictions, each suited for different use cases. Real-time inference is used when you need instant results, typically during live operations. For example, an autonomous drone navigating in real-time needs low-latency responses to make quick decisions. This type of inference is optimized for speed and responsiveness, often handling one data point at a time. On the other hand, batch inference is used for analyzing large volumes of data at once, such as reviewing thousands of drone flight logs. It's more efficient when processing large datasets but does not provide instant results. It runs at scheduled times or on-demand and has higher latency because it's not required to be fast for individual predictions. Understanding these differences helps determine the right inference method for the right scenario. **INCORRECT:** "Real-time inference is optimized for large-scale data processing over long periods, while batch inference handles time-sensitive predictions during operations" is incorrect. It reverses the definitions. Real-time inference is used for time-sensitive operations, not large-scale batch processing. Batch inference, not real-time, is optimized for handling large datasets over time. **INCORRECT:** "Real-time inference is more cost-effective than batch inference because it processes large amounts of data at once" is incorrect. Batch inference is typically more cost-effective for large volumes of data because it processes everything in bulk, reducing overhead. Real-time inference often requires dedicated resources for low-latency response, which can be more expensive. **INCORRECT:** "Batch inference offers lower accuracy than real-time inference due to its inability to adapt to live data" is incorrect. The accuracy of inference is determined by the underlying model, not the inference method. Batch inference can be just as accurate; it just doesn't react instantly to new data. It doesn't mean the predictions are less reliable. **References:** https://aws.amazon.com/blogs/architecture/batch-inference-at-scale-with-amazon-sagemaker Domain: Fundamentals of AI and ML --- #### 34. Arrange the different types of machine learning approaches from LEAST to MOST complex. (Select and order THREE.) - Supervised learning - Reinforcement learning - Unsupervised learning Note: Select only the correct options, as the type of "Ordering" question is not supported here. ``` Reinforcement learning Supervised learning Unsupervised learning ``` ``` Unsupervised learning Supervised learning Reinforcement learning ``` ``` Unsupervised learning Reinforcement learning Supervised learning ``` ``` Supervised learning Unsupervised learning Reinforcement learning ``` **CORRECT:** The correct order from LEAST to MOST complex is: 1. Supervised learning 2. Unsupervised learning 3. Reinforcement learning Supervised learning is the simplest form of machine learning. In this approach, the model is trained using labeled data, meaning the input comes with the correct output. The model learns by finding patterns in the data and mapping inputs to outputs. Since it follows a structured learning process with clear guidance, it is the easiest to understand and implement. Examples include image classification and spam detection. Unsupervised learning is more complex than supervised learning because it works with unlabeled data. The model must find hidden patterns and relationships in the data without explicit instructions. It is commonly used for clustering and anomaly detection, such as customer segmentation and fraud detection. Since the model has no predefined labels, it requires more advanced algorithms and interpretation. Reinforcement learning (RL) is the most complex approach. Instead of learning from labeled data, RL models learn by interacting with an environment, receiving rewards or penalties for their actions. This trial-and-error learning process makes RL powerful but challenging to implement. It is used in robotics, game AI, and self-driving cars. RL requires more computational resources and a well-defined reward system to work effectively. **References:** https://aws.amazon.com/compare/the-difference-between-machine-learning-supervised-and-unsupervised https://aws.amazon.com/what-is/reinforcement-learning Domain: Fundamentals of AI and ML --- #### 35. You're designing an AI-driven system that dynamically adapts its strategy by interacting with its environment, receiving feedback in the form of rewards or penalties, and refining its actions over time based on those outcomes. Which AI paradigm is most suitable for enabling this type of continuous, feedback-based learning? - Deep Learning for hierarchical feature extraction - Reinforcement Learning for decision-making via trial and error - Supervised Learning with labeled historical data - Natural Language Processing for understanding text inputs **CORRECT:** "Reinforcement Learning for decision-making via trial and error" is the correct answer. Reinforcement Learning (RL) is a type of machine learning where an agent learns by interacting with an environment. It makes decisions, receives feedback as rewards or penalties, and adjusts its actions to maximize the total reward over time. This process is like learning through trial and error. RL is ideal for systems that need to dynamically adapt, such as robots navigating a space, game-playing AI, or recommendation engines that update based on user behavior. Because it continuously learns from feedback and refines its strategy, it's the perfect match for scenarios requiring adaptive decision-making over time. **INCORRECT:** "Supervised Learning with labeled historical data" is incorrect. Supervised learning uses labeled datasets to train a model to make predictions or classifications. While powerful for static tasks like spam detection or image classification, it doesn't adapt over time or respond to feedback. It also doesn't support learning through trial and error, making it unsuitable for dynamic, feedback-driven environments. **INCORRECT:** "Deep Learning for hierarchical feature extraction" is incorrect. Deep learning is a type of machine learning that uses neural networks to learn complex patterns in data, often for images, text, or audio. While it can be used within reinforcement learning, deep learning itself doesn't include the feedback-based learning loop needed for dynamic strategy adjustment. It's more about pattern recognition than real-time adaptation. **INCORRECT:** "Natural Language Processing for understanding text inputs" is incorrect. NLP is used for understanding and processing human language, such as chatbots or sentiment analysis. It's a domain-specific tool, not a learning paradigm like reinforcement learning. NLP can work with RL (e.g., in conversational AI), but by itself, it doesn't support decision-making through feedback and rewards. **References:** https://aws.amazon.com/what-is/reinforcement-learning Domain: Fundamentals of AI and ML --- #### 36. A business is sitting on a vast pool of raw, unlabeled customer reviews and wants to extract common themes and hidden patterns without prior annotations or labels. Which AI method is best suited for this requirement? - Supervised learning with pre-classified training data - Reinforcement learning using reward-based feedback - Unsupervised learning techniques such as topic modeling or clustering - Natural language processing (NLP) for intent classification **CORRECT:** "Unsupervised learning techniques such as topic modeling or clustering" is the correct answer. Unsupervised learning is a machine learning approach that discovers patterns or structures in data without needing labeled examples. This makes it perfect for analyzing large sets of unlabeled customer reviews. Techniques like topic modeling (e.g., LDA – Latent Dirichlet Allocation) and clustering (e.g., k-means) help group reviews by similar themes or topics, even when the specific categories aren't known ahead of time. These methods can uncover hidden patterns, such as what customers commonly talk about (e.g., delivery time, product quality, or customer service), giving businesses valuable insights without manual labeling. **INCORRECT:** "Supervised learning with pre-classified training data" is incorrect. Supervised learning requires labeled data, meaning each example must be tagged with the correct output (e.g., positive or negative review). Since the business has raw, unlabeled reviews, supervised learning isn't appropriate here. **INCORRECT:** "Reinforcement learning using reward-based feedback" is incorrect. Reinforcement learning involves training agents to make decisions through trial and error, using rewards or penalties. It's commonly used in robotics, gaming, or navigation tasks—not for analyzing text data like customer reviews. It also requires a feedback loop, which doesn't exist in this scenario. **INCORRECT:** "Natural language processing (NLP) for intent classification" is incorrect. NLP is a great tool for working with human language, but intent classification is a supervised NLP task, which needs labeled examples (e.g., labeling each review with an intent like "complaint" or "praise"). Since the business doesn't have labeled data, this isn't the right approach for discovering unknown patterns or themes. **References:** https://aws.amazon.com/compare/the-difference-between-machine-learning-supervised-and-unsupervised https://docs.aws.amazon.com/comprehend/latest/dg/topic-modeling.html Domain: Fundamentals of AI and ML --- #### 37. A company is deploying an AI-powered customer support chatbot using Amazon Bedrock. The chatbot needs to generate responses while controlling the percentage of most-likely candidates that the model considers for the next token to balance creativity and coherence. Which of the following inference parameters would you recommend for the given use case? - Epochs - Top K - Top-p - Temperature **CORRECT:** "Top-p" is the correct answer. In AI text generation, especially when using foundation models like those on Amazon Bedrock, Top-p (also known as nucleus sampling) is a key inference parameter used to control how creative or deterministic the model's responses are. Top-p controls the cumulative probability of the most likely tokens. The model considers only the smallest set of tokens whose cumulative probability is greater than or equal to p (e.g., 0.9). This means you're balancing between creativity and relevance, which is ideal for generating chatbot responses that are both coherent and flexible. **INCORRECT:** "Temperature" is incorrect. Temperature also affects creativity, but it controls the randomness of token selection rather than focusing on the top-probability candidates. A higher temperature increases randomness, while a lower one makes outputs more predictable. However, Top-p gives more fine-tuned control in many cases. **INCORRECT:** "Epochs" is incorrect. Epochs refer to how many times a model sees the training data during training. They are not used during inference, so they don't apply here. **INCORRECT:** "Top K" is incorrect. Top K sampling chooses from the top K most likely next tokens, regardless of their combined probability. It's useful but less flexible than Top-p, which adjusts the candidate set based on cumulative probability rather than a fixed number. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html Domain: Applications of Foundation Models --- #### 38. Case Study: A multinational e-commerce company wants to improve its customer support experience by implementing an AI-powered chatbot. The company has historical chat logs from customer interactions in multiple languages and wants to enhance its chatbot to provide personalized, accurate, and context-aware responses. The AI model should also be able to handle product inquiries, order status checks, and refund requests efficiently. Question: The company wants to ensure that the chatbot provides more accurate responses based on past customer inquiries. Which combination of techniques should be used? (Select TWO.) - Fine-tune the foundation model with historical customer interactions - Reduce the model size to improve inference latency - Increase the temperature parameter for diverse responses - Implement Retrieval-Augmented Generation (RAG) to fetch relevant context - Use a zero-shot learning approach for unseen queries **CORRECT:** "Fine-tune the foundation model with historical customer interactions" is a correct answer. Fine-tuning is the process of training an existing AI model on a specific dataset to improve its performance on domain-specific tasks. In this case, fine-tuning the chatbot with historical customer interactions allows it to learn from past inquiries, making responses more accurate, personalized, and context-aware. By incorporating past conversations, the chatbot can better understand customer intent, common issues, and preferred solutions. This helps improve response quality and ensures a better customer support experience. **CORRECT:** "Implement Retrieval-Augmented Generation (RAG) to fetch relevant context" is a correct answer. Retrieval-Augmented Generation (RAG) is a technique that enhances AI models by integrating a retrieval mechanism that fetches relevant documents or data before generating a response. This approach is beneficial for customer support because it enables the chatbot to pull up-to-date product details, order history, and policy information before responding. By retrieving relevant customer-specific context from databases or knowledge bases, RAG ensures that the chatbot provides more precise and informed responses, improving customer satisfaction. **INCORRECT:** "Use a zero-shot learning approach for unseen queries" is incorrect. Zero-shot learning allows an AI model to handle queries it has never seen before by making inferences based on general knowledge. While this approach can help with new inquiries, it is not as reliable as fine-tuning or RAG for improving chatbot accuracy. Without historical training or external context retrieval, responses may lack depth and specificity. **INCORRECT:** "Reduce the model size to improve inference latency" is incorrect. Reducing the model size can make responses faster, but it does not directly improve accuracy. A smaller model may lose critical information and context, leading to less precise responses. While latency is important, accuracy in customer support is a higher priority. **INCORRECT:** "Increase the temperature parameter for diverse responses" is incorrect. The temperature parameter controls the randomness of AI responses. Increasing it leads to more varied and creative answers, which is useful in creative applications but not ideal for customer support. Inaccurate or unpredictable responses can frustrate customers and reduce trust in the chatbot. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html https://aws.amazon.com/what-is/retrieval-augmented-generation Domain: Applications of Foundation Models --- #### 39. A healthcare technology company is developing an AI model to analyze and summarize electronic health records (EHRs). The company needs to ensure that patient data remains encrypted and secure while complying with HIPAA regulations. The AI system must prevent unauthorized access to sensitive information and track all API interactions for auditing purposes. Which combination of AWS services will best address these security and compliance requirements? (Select TWO.) - AWS Secrets Manager - Amazon DynamoDB - AWS Key Management Service (KMS) - Amazon Rekognition - AWS CloudTrail **CORRECT:** "AWS Key Management Service (KMS)" is a correct answer. AWS Key Management Service (KMS) is a fully managed encryption service that helps organizations securely manage cryptographic keys used to protect sensitive data. In a healthcare AI system handling Electronic Health Records (EHRs), KMS ensures that all patient data is encrypted at rest and in transit. Since HIPAA compliance mandates strong encryption for Protected Health Information (PHI), KMS provides centralized key management, automatic key rotation, and access control via AWS Identity and Access Management (IAM). This prevents unauthorized access to sensitive records while meeting compliance requirements. KMS integrates with AWS services like Amazon S3, AWS Lambda, and DynamoDB, ensuring seamless encryption without requiring custom implementations. **CORRECT:** "AWS CloudTrail" is a correct answer. AWS CloudTrail is a logging and monitoring service that records all API activity in an AWS environment. For a healthcare AI system analyzing EHRs, CloudTrail ensures compliance by tracking who accessed patient data, when, and from where. Since HIPAA requires audit logging for all interactions with PHI, CloudTrail enables organizations to monitor and review access logs, detect unauthorized attempts, and generate compliance reports. CloudTrail integrates with AWS services like AWS Security Hub and Amazon CloudWatch, allowing real-time alerts on suspicious activities. By maintaining detailed logs, healthcare companies can ensure regulatory adherence and security best practices. **INCORRECT:** "Amazon DynamoDB" is incorrect. Amazon DynamoDB is a NoSQL database that provides fast, scalable storage. While it can store EHR data, it does not directly provide encryption or compliance tracking features. Without additional security measures like KMS encryption and IAM policies, using DynamoDB alone would not be sufficient for HIPAA compliance. **INCORRECT:** "Amazon Rekognition" is incorrect. Amazon Rekognition is an AI service for image and video analysis, such as facial recognition and object detection. It is not designed for processing or securing Electronic Health Records (EHRs). Since the question focuses on data encryption and compliance, Rekognition does not address these concerns. **INCORRECT:** "AWS Secrets Manager" is incorrect. AWS Secrets Manager securely stores and manages sensitive credentials like API keys and passwords. While useful for protecting authentication data, it does not encrypt EHRs or provide compliance auditing. KMS and CloudTrail are better suited for meeting HIPAA security requirements in this scenario. **References:** https://docs.aws.amazon.com/kms/latest/developerguide/overview.html https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 40. A healthcare provider plans to use AWS machine learning to analyze patient medical records and predict potential health risks. The team is evaluating AWS services to determine the best fit for their predictive analytics use case. Which of the following statements accurately describe AWS ML services suitable for predictive healthcare analytics? (Select TWO.) - Amazon Rekognition forecasts patient diagnoses based on visual medical records. - Amazon Translate analyzes medical data to identify future health conditions. - Amazon SageMaker provides tools to train and deploy models predicting patient health risks. - Amazon Comprehend Medical extracts medical information from patient records, assisting predictive analytics. - Amazon Polly directly predicts patient health outcomes from text-to-speech conversion. **CORRECT:** "Amazon SageMaker provides tools to train and deploy models predicting patient health risks" is a correct answer. Amazon SageMaker is a fully managed service that lets data scientists and developers build, train, and deploy machine learning models quickly. It supports end-to-end machine learning workflows and offers built-in algorithms, notebooks, model tuning, and deployment tools. This makes it ideal for healthcare predictive analytics—such as building models to forecast potential health risks using historical patient data. Because it supports custom ML models, it's suitable for analyzing complex medical records and generating predictions about patient health outcomes. **CORRECT:** "Amazon Comprehend Medical extracts medical information from patient records, assisting predictive analytics" is a correct answer. Amazon Comprehend Medical is a natural language processing (NLP) service designed specifically for healthcare. It extracts key medical information—like conditions, medications, tests, and treatments—from unstructured medical text such as doctor's notes or clinical reports. This structured data can then be used for predictive analytics, such as identifying patients at high risk of developing certain conditions. By transforming text into useful data, Comprehend Medical supports better and faster decision-making in healthcare. **INCORRECT:** "Amazon Translate analyzes medical data to identify future health conditions" is incorrect. Amazon Translate is a language translation service that automatically translates text between languages. It doesn't analyze data or make predictions—it simply helps users understand content in other languages. While it can translate medical records into English or another language, it does not support predictive analytics or health risk forecasting. **INCORRECT:** "Amazon Rekognition forecasts patient diagnoses based on visual medical records" is incorrect. Amazon Rekognition is used for image and video analysis—such as detecting objects, scenes, and faces. It's not specialized for medical images and cannot predict diagnoses. While it might be used in some visual healthcare scenarios (like analyzing X-rays), it lacks the clinical accuracy and medical model training required for predictive healthcare tasks. **INCORRECT:** "Amazon Polly directly predicts patient health outcomes from text-to-speech conversion" is incorrect. Amazon Polly is a text-to-speech (TTS) service that turns written text into natural-sounding speech. It's useful for accessibility or reading out patient instructions but doesn't have any predictive analytics capabilities. Polly does not analyze data or generate predictions—it just reads the text aloud. **References:** https://aws.amazon.com/sagemaker https://aws.amazon.com/comprehend/medical Domain: Applications of Foundation Models --- #### 41. A media streaming company wants to use pre-built foundation models to auto-generate movie summaries and customize these outputs using their own movie scripts and descriptions. They need an AWS service that provides access to multiple models and also allows them to fine-tune these models privately with their proprietary data. Which AWS service or feature will meet these requirements? - Amazon Rekognition enables model customization for personalized summaries. - Amazon Comprehend provides pre-trained models for customization. - Amazon Q Business offers pre-built models with private fine-tuning capabilities. - Amazon Bedrock offers access to multiple FMs with private customization features. **CORRECT:** "Amazon Bedrock offers access to multiple FMs with private customization features" is the correct answer. Amazon Bedrock is a managed AWS service that allows you to build and scale generative AI applications using foundation models (FMs) from leading providers like Anthropic, Meta, and Amazon itself. With Bedrock, you can access multiple models through a simple API without needing to manage infrastructure. Most importantly, Bedrock allows private customization — meaning you can fine-tune or adapt these models securely using your own data (like movie scripts and descriptions) without exposing that data to the model provider. This fits perfectly with the company's needs for model access and private fine-tuning. **INCORRECT:** "Amazon Rekognition enables model customization for personalized summaries" is incorrect. Amazon Rekognition is mainly used for analyzing images and videos — detecting objects, faces, activities, and inappropriate content. It is not built for text generation or summarization tasks, and it does not support using foundation models or fine-tuning them with custom text data. Thus, it doesn't meet the company's requirement for movie summaries. **INCORRECT:** "Amazon Comprehend provides pre-trained models for customization" is incorrect. Amazon Comprehend is a natural language processing (NLP) service that helps extract insights like sentiment, key phrases, or entities from text. It is not designed for foundation model-based text generation or fine-tuning large models with new training data. It's mainly for extracting meaning, not for generating new content. **INCORRECT:** "Amazon Q Business offers pre-built models with private fine-tuning capabilities" is incorrect. Amazon Q Business is focused on building intelligent assistants for internal company knowledge, using conversational AI. It does not offer foundation model access or creative text generation like movie summaries. It is mainly aimed at enhancing productivity and answering enterprise questions. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html Domain: Applications of Foundation Models --- #### 42. Which of the following is a primary advantage of diffusion models over traditional generative adversarial networks (GANs)? - Diffusion models are significantly faster in generating content compared to GANs. - Diffusion models require less training data than GANs. - Diffusion models use supervised learning instead of unsupervised learning. - Diffusion models generate higher-quality images with better control over the generation process. **CORRECT:** "Diffusion models generate higher-quality images with better control over the generation process" is the correct answer. Diffusion models have gained popularity over GANs due to their ability to generate high-quality images with superior fine-grained control. They work by iteratively refining noise through a step-by-step denoising process, leading to more stable and realistic outputs. Unlike GANs, which can suffer from mode collapse (where the model generates limited variations of an image), diffusion models explore the entire data distribution effectively, producing more diverse and high-fidelity results. Additionally, techniques like classifier-guided and classifier-free guidance provide better control over image attributes, making them ideal for applications in AI-generated art and image synthesis. **INCORRECT:** "Diffusion models require less training data than GANs" is incorrect. Diffusion models typically require substantial computational resources and large datasets for training. While they generate high-quality outputs, their training process involves learning a denoising schedule, making it computationally expensive and data-intensive compared to GANs. **INCORRECT:** "Diffusion models are significantly faster in generating content compared to GANs" is incorrect. Diffusion models are generally slower than GANs during inference. Since they generate images through multiple iterative steps of denoising, they require more computation time compared to GANs, which generate images in a single forward pass of a neural network. However, optimizations such as latent diffusion models (LDMs) and accelerated sampling techniques are improving inference speeds. **INCORRECT:** "Diffusion models use supervised learning instead of unsupervised learning" is incorrect. Like GANs, diffusion models are typically trained in an unsupervised or self-supervised manner, learning patterns from unlabeled data. They do not inherently require labeled data, as their training process involves modeling the distribution of real data through iterative noise removal. **References:** https://aws.amazon.com/what-is/stable-diffusion https://aws.amazon.com/what-is/gan Domain: Fundamentals of Generative AI --- #### 43. A customer service firm is developing an AI-powered chatbot using foundation models on AWS. To improve response quality and user satisfaction, the team is considering Reinforcement Learning from Human Feedback (RLHF). They want to understand the primary benefit of using RLHF. Which of the following best describes why RLHF is used in chatbot development? - RLHF replaces the need for prompt engineering by automatically generating optimal prompts. - RLHF eliminates the use of supervised learning and relies entirely on reward scores. - RLHF adds more training data from web crawlers to improve the model's factual accuracy. - RLHF aligns model outputs with human preferences by using feedback to fine-tune responses over time. **CORRECT:** "RLHF aligns model outputs with human preferences by using feedback to fine-tune responses over time" is the correct answer. Reinforcement Learning from Human Feedback (RLHF) is a training approach that helps AI systems, such as chatbots, generate responses that better align with human expectations and values. It works by first generating responses using a base model, then collecting human feedback on which responses are more appropriate, helpful, or relevant. This feedback is used to train a reward model that guides the chatbot during further fine-tuning. The result is a model that not only performs well based on statistical patterns but also responds in a way that feels more natural, safe, and satisfying to users. In chatbot development, RLHF is especially useful for reducing harmful, biased, or irrelevant responses by reinforcing behaviors that align with what humans prefer. **INCORRECT:** "RLHF adds more training data from web crawlers to improve the model's factual accuracy" is incorrect. RLHF is not about gathering new data from web sources. Instead, it focuses on using human-labeled preferences to guide model improvements. While external data may improve general accuracy, RLHF specifically enhances alignment with human intent, not factual correctness. **INCORRECT:** "RLHF replaces the need for prompt engineering by automatically generating optimal prompts" is incorrect. RLHF does not eliminate prompt engineering. While it helps models produce better outputs, prompt engineering is still valuable, especially in foundation model use. RLHF adjusts the model's behavior through feedback, but it doesn't generate or automate prompts. **INCORRECT:** "RLHF eliminates the use of supervised learning and relies entirely on reward scores" is incorrect. RLHF is usually a multi-step process that starts with supervised fine-tuning using labeled data. The reward model is then trained using human preferences, and reinforcement learning is applied afterward. It doesn't fully replace supervised learning; it builds upon it. **References:** https://aws.amazon.com/what-is/reinforcement-learning-from-human-feedback Domain: Applications of Foundation Models --- #### 44. A developer team is building a text generation app and wants to experiment with multiple foundation models while keeping the architecture simple. Which two Amazon Bedrock features make this possible? (Select TWO.) - Fully managed infrastructure - On-device inference engine - Model switching without retraining - Built-in GPU provisioning control - Direct access to training data **CORRECT:** "Fully managed infrastructure" is a correct answer. Amazon Bedrock provides a fully managed infrastructure that allows developers to experiment with and deploy foundation models without needing to manage servers, provision GPUs, or handle model hosting manually. This helps simplify the architecture significantly. Developers can focus on building their applications while AWS handles the scalability, security, and availability of the infrastructure, making it easier to experiment with generative AI models. **CORRECT:** "Model switching without retraining" is a correct answer. One of the unique features of Amazon Bedrock is the ability to switch between different foundation models (such as Anthropic Claude, AI21 Labs Jurassic, Meta Llama, and Amazon Titan) without retraining. This is ideal for developers who want to compare model outputs or optimize for different use cases. Since all models are accessible through the same API structure, developers can test different models with minimal code changes, enabling faster experimentation and development cycles. **INCORRECT:** "Direct access to training data" is incorrect. Amazon Bedrock does not provide direct access to the training data of foundation models. This is by design to protect proprietary model data and user security. Developers can fine-tune or customize models using their own data but cannot view or access the original training datasets. **INCORRECT:** "Built-in GPU provisioning control" is incorrect. Amazon Bedrock abstracts away the need for users to manage GPU provisioning. While this simplifies usage, it also means developers do not have direct control over GPU resources. This feature is managed by AWS to provide a seamless, scalable experience. **INCORRECT:** "On-device inference engine" is incorrect. Amazon Bedrock is a cloud-based service. It does not support on-device inference. All inference happens within the AWS infrastructure, so this option is not applicable for Bedrock's serverless model usage. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html Domain: Fundamentals of Generative AI --- #### 45. A startup is launching an AI chatbot that experiences unpredictable spikes in user traffic. To minimize infrastructure management and control costs, they are considering Amazon Bedrock. Which of the following is the primary benefit of using Amazon Bedrock in this scenario? - Amazon Bedrock is a fully managed service that automatically handles provisioning and scaling of infrastructure based on application demand. - Amazon Bedrock accelerates model inference by defaulting to high-performance GPU instances. - Amazon Bedrock charges only for data storage, making it ideal for cost-sensitive AI applications. - Amazon Bedrock provides customizable chat interface templates for rapid chatbot deployment. **CORRECT:** "Amazon Bedrock is a fully managed service that automatically handles provisioning and scaling of infrastructure based on application demand" is the correct answer. Amazon Bedrock is a fully managed AWS service that makes it easy to build and scale generative AI applications using foundation models from leading AI companies without managing any infrastructure. One of the key benefits of Bedrock is that it abstracts away infrastructure provisioning, scaling, and maintenance. This is particularly valuable for startups or applications with unpredictable traffic, like a chatbot experiencing sudden spikes. Bedrock automatically scales based on demand, so the developers don't need to worry about capacity planning, hardware selection, or managing server loads. This reduces operational overhead and allows the startup to focus more on innovation and user experience, while also optimizing costs since they pay only for what they use. **INCORRECT:** "Amazon Bedrock provides customizable chat interface templates for rapid chatbot deployment" is incorrect. Amazon Bedrock does not provide front-end interface templates. It focuses on backend model access and orchestration. Developers are responsible for building the user interface on their own or with other AWS tools like Amazon Lex or AWS Amplify. **INCORRECT:** "Amazon Bedrock accelerates model inference by defaulting to high-performance GPU instances" is incorrect. While Amazon Bedrock uses high-performance infrastructure in the background, it does not expose instance types or guarantee GPU acceleration settings to the user. Its strength lies in simplicity and scalability—not hardware configuration. **INCORRECT:** "Amazon Bedrock charges only for data storage, making it ideal for cost-sensitive AI applications" is incorrect. Bedrock's pricing is primarily based on the number of input/output tokens processed, not data storage. Data storage costs are not the main driver of cost in generative AI use cases. **References:** https://aws.amazon.com/bedrock https://aws.amazon.com/bedrock/pricing Domain: Applications of Foundation Models --- #### 46. Your team is building an AI-based recommendation engine using sensitive customer data that includes names, phone numbers, and financial account details. To meet global privacy regulations (such as GDPR and CCPA), your team must ensure that personally identifiable information (PII) is automatically detected before any data is processed for model training. You also want the ability to generate reports, mask sensitive fields during preprocessing. Which AWS service is best suited to help detect and manage PII in your AI datasets before training? - Use Amazon Macie to scan structured and unstructured data in Amazon S3, identify PII such as names and account numbers, and apply automated alerts and protection policies. - Use Amazon Comprehend to train a custom entity recognizer that can detect PII in natural language text and redact personally sensitive entities during preprocessing. - Use AWS Lake Formation to define data access policies, manage permissions at the column level, and restrict access to sensitive PII fields in shared AI datasets. - Use AWS Identity and Access Management (IAM) to block unauthorized access to S3 datasets and enforce field-level encryption of customer records containing PII. **CORRECT:** "Use Amazon Macie to scan structured and unstructured data in Amazon S3, identify PII such as names and account numbers, and apply automated alerts and protection policies" is the correct answer. Amazon Macie is a data security service that uses machine learning to automatically discover, classify, and protect sensitive data stored in Amazon S3. It is specifically designed to detect personally identifiable information (PII), including names, contact numbers, financial data, and more. Macie automatically scans structured and unstructured datasets and provides detailed reports and dashboards to highlight security risks. It can generate alerts and trigger actions based on findings—for example, masking PII before it's used in AI/ML pipelines. This makes it especially useful for compliance with privacy regulations like GDPR and CCPA. Since your use case involves AI training on sensitive customer data, Macie offers the ideal automated and scalable solution for pre-processing protection. **INCORRECT:** "Use AWS Identity and Access Management (IAM) to block unauthorized access to S3 datasets and enforce field-level encryption of customer records containing PII" is incorrect. While IAM is essential for securing access to AWS resources, it does not detect or classify PII. IAM controls who can access data but does not inspect or modify the data itself. It does not offer functionality for detecting sensitive fields or automatically encrypting/redacting them based on content. **INCORRECT:** "Use Amazon Comprehend to train a custom entity recognizer that can detect PII in natural language text and redact personally sensitive entities during preprocessing" is incorrect. Amazon Comprehend can detect PII in unstructured text, but it requires more setup and training for custom entities. It's effective for natural language processing tasks, but it doesn't support scanning files at scale in Amazon S3 or generate automated privacy compliance reports like Macie does. It's more limited in scope and not designed for general dataset compliance. **INCORRECT:** "Use AWS Lake Formation to define data access policies, manage permissions at the column level, and restrict access to sensitive PII fields in shared AI datasets" is incorrect. AWS Lake Formation is great for managing access to large-scale data lakes, especially in analytics environments. It helps restrict access to specific tables or columns, but it does not automatically detect PII or perform masking/redaction. It's a strong governance tool but doesn't meet the need for automated PII detection before ML training. **References:** https://docs.aws.amazon.com/macie/latest/user/what-is-macie.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 47. An educational institution plans to implement Foundation Models for generating personalized learning content and automated course materials. The educational team seeks to understand the essential features of these models to evaluate their effectiveness for educational content. Which statement correctly describes Foundation Models (FMs) in the generative AI context? - Foundation Models are designed for only one specific use case and cannot be fine-tuned for different applications. - Foundation Models (FMs) are pre-trained on large datasets and can be fine-tuned for specific tasks, making them highly adaptable for various AI applications. - Foundation Models must be trained separately for each subject matter. - Foundation Models cannot create adaptive content based on learner interactions. **CORRECT:** "Foundation Models (FMs) are pre-trained on large datasets and can be fine-tuned for specific tasks, making them highly adaptable for various AI applications" is the correct answer. Foundation Models (FMs) are a type of large-scale machine learning model trained on diverse and massive datasets. These models, such as GPT, BERT, and others, learn a broad understanding of language and concepts during their initial pre-training. Once trained, they can be fine-tuned on smaller, domain-specific datasets to perform specific tasks, such as generating personalized learning content or automating educational materials. This makes them highly flexible and efficient for many use cases, including education, healthcare, customer service, and more. In an educational context, FMs can generate adaptive learning experiences by analyzing student interactions, preferences, and performance, then tailoring content accordingly. This adaptability helps institutions offer more engaging and effective learning paths for individual students. **INCORRECT:** "Foundation Models are designed for only one specific use case and cannot be fine-tuned for different applications" is incorrect. Foundation Models are not limited to one specific use case. Their primary strength lies in their ability to be adapted (fine-tuned) for many different applications after the initial training. **INCORRECT:** "Foundation Models cannot create adaptive content based on learner interactions" is incorrect. FMs are capable of generating adaptive and dynamic content. They can analyze user behavior and feedback to tailor responses or educational content to suit different learners. **INCORRECT:** "Foundation Models must be trained separately for each subject matter" is incorrect. One of the main advantages of FMs is that they don't need to be trained from scratch for each new subject. After initial training, they can be fine-tuned with minimal data to work on various subject areas. **References:** https://aws.amazon.com/what-is/foundation-models Domain: Fundamentals of Generative AI --- #### 48. A team working in environmental science is developing a model to classify satellite images to detect forest loss and water body expansion. These images include complex textures and varying resolutions. Which type of deep learning model is most effective for extracting hierarchical spatial features from satellite imagery? - Graph Neural Networks (GNN) - Feedforward Neural Networks (FNNs) - Convolutional Neural Networks (CNN) - Recurrent Neural Networks (RNN) **CORRECT:** "Convolutional Neural Networks (CNN)" is the correct answer. Convolutional Neural Networks (CNNs) are a class of deep learning models specifically designed to process data with a grid-like topology, such as images. They are especially effective at capturing spatial hierarchies in data, making them ideal for tasks involving image recognition, classification, and segmentation. In CNNs, layers of convolutional filters scan over the input image to detect local patterns, such as edges, textures, and shapes. These patterns are then combined through deeper layers to form more abstract and complex features, allowing the network to understand high-level structures in an image. CNNs typically include three main types of layers: convolutional layers, pooling layers, and fully connected layers. Convolutional layers apply a set of filters to the input to extract features, while pooling layers reduce the spatial dimensions, making the model more computationally efficient and less sensitive to small shifts in the image. Finally, fully connected layers interpret the extracted features to produce the final output, such as classifying an image as showing forest loss or water body expansion. This hierarchical feature learning ability makes CNNs particularly powerful for analyzing satellite imagery with complex textures and varying resolutions. **INCORRECT:** "Graph Neural Networks (GNN)" is incorrect. GNNs are designed for data represented as graphs, like social networks, molecular structures, or transportation systems. While powerful for analyzing relationships and connections, GNNs are not designed for image data. They don't work well with raw pixel data like CNNs do for image classification. **INCORRECT:** "Recurrent Neural Networks (RNN)" is incorrect. RNNs are used for sequential data, such as time series, speech, or text. They are good at remembering previous steps in a sequence, which makes them useful for tasks like language translation or forecasting — not for analyzing image data. **INCORRECT:** "Feedforward Neural Networks (FNNs)" is incorrect. Feedforward Neural Networks are the most basic type of neural network. While they can handle image inputs in theory, they do not scale well with high-dimensional image data. They lack the ability to recognize spatial patterns or local features like CNNs, making them inefficient for image classification. **References:** https://aws.amazon.com/what-is/neural-network Domain: Fundamentals of AI and ML --- #### 49. A company wants to design a chatbot that can answer customer queries about its financial products. The company is evaluating whether to use standard machine learning (ML) techniques or to implement a large language model (LLM) solution. What is the key difference between these two approaches? - ML techniques are only for regression tasks, while LLMs can be used for any advanced AI tasks. - ML techniques cannot handle language processing, while LLMs can handle images and text simultaneously. - ML techniques rely on smaller, task-specific datasets, while LLMs are pre-trained on massive, diverse datasets. - ML techniques and LLMs have the same architectural structure, but LLMs use only supervised learning. **CORRECT:** "ML techniques rely on smaller, task-specific datasets, while LLMs are pre-trained on massive, diverse datasets" is the correct answer. Machine Learning (ML) is a branch of artificial intelligence (AI) that enables computers to learn patterns from data and make predictions or decisions without being explicitly programmed. It uses algorithms to improve performance over time through experience. Machine learning (ML) techniques typically require carefully labeled, task-specific datasets to train models for specific use cases, such as fraud detection or customer segmentation. These models are trained on a limited dataset focused on solving a particular problem. On the other hand, Large Language Models (LLMs) are advanced artificial intelligence (AI) models trained on massive datasets to understand, generate, and process human language. They use deep learning techniques, particularly transformer architectures, to analyze and predict text patterns. LLMs power applications like chatbots, text summarization, code generation, and translation. Examples include OpenAI's GPT, Google's Gemini, and Meta's Llama. This fundamental difference makes LLMs more adaptable and capable of answering a wide range of questions, while standard ML techniques require more tailored training for each specific problem. **INCORRECT:** "ML techniques and LLMs have the same architectural structure, but LLMs use only supervised learning" is incorrect. This statement is incorrect because ML models can have different architectures, including decision trees, neural networks, and transformers. LLMs, such as GPT and BERT, specifically use transformer-based architectures. Additionally, LLMs use both supervised and unsupervised learning techniques during their training, including self-supervised learning for pretraining and supervised fine-tuning for task-specific improvements. **INCORRECT:** "ML techniques cannot handle language processing, while LLMs can handle images and text simultaneously" is incorrect. ML techniques include natural language processing (NLP) models that can process text-based tasks. Traditional ML models like Naïve Bayes, Support Vector Machines (SVM), and even neural networks can handle text classification and sentiment analysis. Also, not all LLMs handle both images and text—only multimodal models can process both. **INCORRECT:** "ML techniques are only for regression tasks, while LLMs can be used for any advanced AI tasks" is incorrect. ML techniques are not limited to regression; they also include classification, clustering, reinforcement learning, and more. LLMs are powerful for text-based tasks, but they are not universally superior for every AI application. Some ML models are more efficient for structured data analysis, such as fraud detection or recommendation systems. **References:** https://aws.amazon.com/ai/machine-learning https://aws.amazon.com/what-is/large-language-model Domain: Fundamentals of AI and ML --- #### 50. Case Study: A multinational e-commerce company wants to improve its customer support experience by implementing an AI-powered chatbot. The company has historical chat logs from customer interactions in multiple languages and wants to enhance its chatbot to provide personalized, accurate, and context-aware responses. The AI model should also be able to handle product inquiries, order status checks, and refund requests efficiently. Question: The company wants to ensure the chatbot is unbiased and does not generate inappropriate responses. Which AWS tool should they use for content moderation? - Amazon SageMaker Model Monitor - Amazon Bedrock Agent - Amazon Bedrock Guardrails - Amazon SageMaker Model Cards **CORRECT:** "Amazon Bedrock Guardrails" is the correct answer. Amazon Bedrock Guardrails is a content moderation and safety feature designed to ensure AI-generated responses remain appropriate, unbiased, and aligned with business guidelines. It helps filter out harmful, offensive, or biased language from chatbot interactions, making it ideal for ensuring ethical AI responses. By configuring guardrails, the company can enforce strict moderation policies, prevent misinformation, and enhance customer trust in the chatbot. Since the chatbot interacts with customers in multiple languages, this feature ensures that responses remain safe and contextually appropriate across different cultures and regions. **INCORRECT:** "Amazon Bedrock Agent" is incorrect. Amazon Bedrock Agent is designed to automate complex workflows by integrating AI models with business applications and APIs. While it helps chatbots perform tasks like retrieving order details or processing refunds, it does not provide content moderation features. **INCORRECT:** "Amazon SageMaker Model Monitor" is incorrect. Amazon SageMaker Model Monitor tracks the performance of machine learning models over time to detect data drift and quality degradation. While useful for monitoring AI performance, it does not filter inappropriate content or manage chatbot biases. **INCORRECT:** "Amazon SageMaker Model Cards" is incorrect. Amazon SageMaker Model Cards help document machine learning models, providing transparency on their training data, biases, and performance. However, it does not actively prevent biased or inappropriate responses in chatbot interactions. **References:** https://aws.amazon.com/bedrock/guardrails Domain: Applications of Foundation Models --- #### 51. A startup is developing a language model for real-time translation services. They want to adopt responsible AI practices, especially focusing on sustainability. The development team is considering whether to train a new model from scratch or fine-tune an existing pre-trained model. Which option would be more responsible from a sustainability perspective? - Use only open-source datasets without considering the model's environmental impact - Fine-tune an existing pre-trained model - Focus on creating the largest possible model for better performance - Train a completely new model from scratch **CORRECT:** "Fine-tune an existing pre-trained model" is the correct answer. From a sustainability perspective, fine-tuning an existing pre-trained model is the more responsible option. Training a new model from scratch typically requires vast computational resources, which translates into a higher carbon footprint due to energy consumption during the process. By leveraging a pre-trained model and fine-tuning it for specific tasks such as real-time translation, the startup can significantly reduce the energy and resources needed for model training. Pre-trained models already contain generalized knowledge, so fine-tuning only requires updating the model's parameters for the specific task, making the process much more energy-efficient. This approach aligns with responsible AI practices as it minimizes the environmental impact while achieving the desired performance. By using pre-trained models, companies can reduce their carbon footprint while still delivering high-quality AI solutions. **INCORRECT:** "Train a completely new model from scratch" is incorrect. Training a new model from scratch involves a large computational effort and increases the environmental impact. It is not the most sustainable option, especially when an existing pre-trained model can be fine-tuned to meet the translation service needs. **INCORRECT:** "Use only open-source datasets without considering the model's environmental impact" is incorrect. While using open-source datasets can help save costs and ensure transparency, this option does not directly address the sustainability concerns related to the environmental impact of model training. The focus should be on minimizing the energy required, which is better achieved by using pre-trained models. **INCORRECT:** "Focus on creating the largest possible model for better performance" is incorrect. Larger models tend to require more computational resources, which increases the energy consumption and carbon footprint. While larger models may offer better performance, prioritizing sustainability involves optimizing for both performance and energy efficiency. **References:** https://aws.amazon.com/sustainability Domain: Guidelines for Responsible AI --- #### 52. A retail company is under pressure to launch a generative AI–powered recommendation engine within a few weeks to align with an upcoming promotional campaign. Which key advantage of generative AI enables the company to accelerate development and meet its tight deadline with minimal overhead? - Access to pre-built APIs and managed infrastructure through services like Amazon Bedrock. - The ability to train large foundation models from scratch on proprietary datasets. - Custom container orchestration using Amazon ECS for model deployment. - Full control over GPU cluster configurations and custom ML pipelines. **CORRECT:** "Access to pre-built APIs and managed infrastructure through services like Amazon Bedrock" is the correct answer. Amazon Bedrock is a fully managed service that lets developers build and scale generative AI applications quickly by offering easy access to foundation models from leading providers (like Anthropic, AI21 Labs, and more) via API. The biggest advantage is that you don't need to manage infrastructure or train models from scratch. This makes it perfect for businesses under time pressure because you can prototype, integrate, and deploy generative AI features with minimal overhead. Instead of worrying about hardware setup or model training, teams can focus on building business-specific features. This drastically shortens development timelines — ideal for meeting tight deadlines like a promotional campaign launch. **INCORRECT:** "The ability to train large foundation models from scratch on proprietary datasets" is incorrect. Training large foundation models from scratch is resource-intensive, time-consuming, and typically requires deep expertise and massive datasets. It's not practical for a company under time pressure to launch within weeks. AWS provides this capability, but it's not the fastest path for quick deployment. **INCORRECT:** "Full control over GPU cluster configurations and custom ML pipelines" is incorrect. While having control over GPU clusters is valuable for advanced customization, it introduces setup and management overhead. This flexibility is better suited for custom ML research or long-term projects, not for businesses trying to move fast with minimal setup. **INCORRECT:** "Custom container orchestration using Amazon ECS for model deployment" is incorrect. Amazon ECS is great for deploying containerized applications, including ML models, but it still requires you to build, containerize, and manage the model lifecycle. This adds development complexity and doesn't offer the speed and simplicity that generative AI services like Amazon Bedrock provide. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html https://docs.aws.amazon.com/bedrock/latest/userguide/foundation-models-reference.html Domain: Fundamentals of Generative AI --- #### 53. An online retail company aims to proactively forecast its upcoming sales volume by leveraging historical transaction data, evolving marketing campaign performance, and patterns of seasonal consumer demand. Which AI/ML technique is best suited to generate accurate predictive insights in this case? - Sentiment analysis using NLP to evaluate customer reviews across time - Time-series forecasting models trained on multi-variable temporal datasets - Unsupervised clustering to segment customer behavior - Convolutional neural networks (CNNs) for product image classification **CORRECT:** "Time-series forecasting models trained on multi-variable temporal datasets" is the correct answer. Time-series forecasting is a type of machine learning technique used to predict future values based on previously observed data points collected over time. These models are especially effective when dealing with trends, seasonality, and cyclic patterns, making them ideal for forecasting sales. In this case, the company wants to predict future sales based on past transactions, marketing campaign data, and seasonal demand. Time-series forecasting models like ARIMA, Prophet, or machine learning-based ones such as LSTM (Long Short-Term Memory) networks can learn from multiple time-dependent variables (multi-variable datasets). This approach helps generate accurate and actionable predictions, allowing the business to plan inventory, staffing, and marketing campaigns accordingly. **INCORRECT:** "Sentiment analysis using NLP to evaluate customer reviews across time" is incorrect. Sentiment analysis uses Natural Language Processing (NLP) to determine whether text data (like reviews) express positive, negative, or neutral opinions. While this technique can help understand customer satisfaction or brand perception, it doesn't directly help forecast numerical values like future sales volume. **INCORRECT:** "Unsupervised clustering to segment customer behavior" is incorrect. Unsupervised clustering (e.g., k-means or DBSCAN) groups similar data points together based on patterns. It's useful for customer segmentation—understanding which customers behave alike. While this technique helps target marketing campaigns, it doesn't predict future values or trends. It's more exploratory than predictive, making it less suitable for forecasting sales. **INCORRECT:** "Convolutional neural networks (CNNs) for product image classification" is incorrect. CNNs are great for tasks involving images, such as recognizing products or detecting visual defects. However, sales forecasting is a numerical and temporal problem, not an image classification one. Therefore, CNNs are not suitable here as they don't process time-based or sales-related data effectively. **References:** https://aws.amazon.com/blogs/machine-learning/time-series-forecasting-with-amazon-sagemaker-automl Domain: Fundamentals of AI and ML --- #### 54. A healthcare AI company is developing a solution that uses Amazon Bedrock to fine-tune a foundation model for summarizing patient case notes. As part of their model validation process, they need to securely store large volumes of sensitive patient data (e.g., medical histories and physician notes). The validation process must be highly scalable and compliant with data protection standards such as HIPAA. The company is seeking a storage solution that integrates well with Amazon Bedrock, ensures data security, and can handle parallel access for efficient model validation at scale. Which storage solution best meets the company's requirements for scalability, security, and compatibility with Amazon Bedrock during the model validation process? - Use Amazon EBS volumes attached to EC2 instances to store and stream validation data during model evaluation. - Use Amazon DynamoDB to store validation datasets with millisecond latency for frequent read/write operations. - Use Amazon Redshift to store structured validation data and leverage its high-performance analytics capabilities. - Use Amazon S3 with server-side encryption (SSE) to store and manage large-scale, secure datasets used in validation. **CORRECT:** "Use Amazon S3 with server-side encryption (SSE) to store and manage large-scale, secure datasets used in validation" is the correct answer. Amazon S3 (Simple Storage Service) is a scalable object storage service used for storing and retrieving large amounts of data. It is ideal for unstructured or semi-structured data like medical notes or reports. With server-side encryption (SSE), S3 automatically encrypts data at rest using AWS-managed keys (SSE-S3) or customer-managed keys (SSE-KMS), helping organizations meet strict data protection and compliance requirements like HIPAA. For a healthcare AI company using Amazon Bedrock, S3 is the best choice because it is highly scalable, integrates well with other AWS services including Bedrock, and allows secure, parallel access to data for model validation tasks. It supports audit logging via AWS CloudTrail and access control using IAM policies, which further enhances security and compliance. **INCORRECT:** "Use Amazon Redshift to store structured validation data and leverage its high-performance analytics capabilities" is incorrect. Amazon Redshift is a data warehouse used mainly for complex analytical queries on structured data. It's not designed for storing large unstructured datasets like medical histories or physician notes. Also, Redshift doesn't provide the same object-level storage and flexibility as S3 for parallel access or integration with Amazon Bedrock. **INCORRECT:** "Use Amazon DynamoDB to store validation datasets with millisecond latency for frequent read/write operations" is incorrect. DynamoDB is a NoSQL key-value and document database optimized for low-latency read/write operations. It's excellent for transactional workloads, but not ideal for storing and accessing large unstructured files (like PDFs or case notes) used in ML validation. It also has limitations in storage size per item, making it unsuitable for large-scale document storage. **INCORRECT:** "Use Amazon EBS volumes attached to EC2 instances to store and stream validation data during model evaluation" is incorrect. Amazon EBS provides block-level storage for EC2 instances and is suitable for high-performance local workloads. However, it does not scale as easily as S3, lacks the parallel access capabilities, and requires you to manage storage provisioning and instance availability. It's not cost-effective or scalable for large-scale data validation in an AI workflow. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/custom-model-import-prereq.html https://docs.aws.amazon.com/bedrock/latest/userguide/s3-bucket-access.html Domain: Applications of Foundation Models --- #### 55. You are developing a chatbot to interact with customers and provide informative, human-like responses about your company's products. To ensure the chatbot can understand various customer inputs and respond with contextually relevant answers, which AI solution should be used? - Natural Language Processing (NLP) with context-aware transformer models - Deep Reinforcement Learning for adaptive conversation management - Hybrid Neural-Symbolic Systems for integrated reasoning and language understanding - Advanced Generative Adversarial Networks (GANs) for dynamic text synthesis **CORRECT:** "Natural Language Processing (NLP) with context-aware transformer models" is the correct answer. Natural Language Processing (NLP) is the AI subfield focused on enabling machines to understand, interpret, and generate human language. Modern NLP relies heavily on transformer-based models like BERT, GPT, or T5, which are capable of capturing context and intent from customer queries to produce relevant and natural-sounding responses. For a chatbot to engage in human-like conversations, it needs to understand nuances, slang, sentiment, and user intent—capabilities that transformer models offer. These models use attention mechanisms to focus on relevant parts of the input and generate context-rich replies, making NLP with transformers the best fit for a smart, conversational chatbot. **INCORRECT:** "Advanced Generative Adversarial Networks (GANs) for dynamic text synthesis" is incorrect. GANs are designed for generating data like images, music, and sometimes text, by pitting two neural networks against each other. While GANs are powerful for creative generation tasks, they are not ideal for conversational understanding or dialogue management. They lack the sequential processing and context retention needed for chatbot applications. **INCORRECT:** "Deep Reinforcement Learning for adaptive conversation management" is incorrect. Reinforcement learning can help optimize decision strategies in interactive environments (like choosing the best response in a game or chatbot). However, it doesn't specialize in understanding or generating natural language. While it might enhance a chatbot's response selection, it cannot be the core method for interpreting customer queries. **INCORRECT:** "Hybrid Neural-Symbolic Systems for integrated reasoning and language understanding" is incorrect. These systems combine neural networks with symbolic reasoning, aiming to bring logic and interpretability to AI. While promising for explainable AI or complex reasoning tasks, they are not the current standard for building responsive and fluent chatbots. NLP with transformers is more suited and widely adopted for this use case. **References:** https://aws.amazon.com/what-is/nlp Domain: Fundamentals of AI and ML --- #### 56. A call center is developing an AI-powered IVR system using Amazon Polly. They want to fine-tune the speech output to sound more natural by adjusting tone, pitch, and speaking rate. Which Amazon Polly feature should they use to accomplish this? - Voice Engines - SSML (Speech Synthesis Markup Language) - Lexicons - Speech Marks **CORRECT:** "SSML (Speech Synthesis Markup Language)" is the correct answer. SSML (Speech Synthesis Markup Language) is a standardized markup language supported by Amazon Polly that allows developers to customize and control various aspects of speech output. With SSML, you can adjust the tone, pitch, volume, rate of speech, pauses, and pronunciation to make the synthesized speech sound more natural and human-like. In the context of an IVR (Interactive Voice Response) system, this is especially useful for delivering information in a clear and friendly tone that improves user experience. For example, you can slow down the speech rate when reading phone numbers or dates, or change pitch and emphasis for better clarity and engagement. SSML makes it easier to fine-tune how the system communicates with callers. **INCORRECT:** "Voice Engines" is incorrect. Voice Engines in Amazon Polly refer to the underlying technology (standard or neural) used to produce speech. While neural voices do sound more natural, they don't allow custom tuning of pitch, tone, or speed directly. SSML is required for that kind of control. **INCORRECT:** "Speech Marks" is incorrect. Speech Marks provide metadata like sentence boundaries, word timings, or phoneme positions during speech synthesis. They're useful for syncing speech with visuals or captions, but they don't alter how the voice sounds. **INCORRECT:** "Lexicons" is incorrect. Lexicons in Amazon Polly are used to customize the pronunciation of words using phonetic symbols. They help correct or standardize how Polly says specific words, but they do not control tone, pitch, or speed. **References:** https://docs.aws.amazon.com/polly/latest/dg/ssml.html https://docs.aws.amazon.com/polly/latest/dg/supportedtags.html Domain: Applications of Foundation Models --- #### 57. Order and arrange the following technologies from the most general to the most specialized. Deep Learning Artificial Intelligence Generative AI Machine Learning Note: Select only the correct options, as the type of "Ordering" question is not supported here. ``` Generative AI Deep Learning Machine Learning Artificial Intelligence ``` ``` Artificial Intelligence Machine Learning Deep Learning Generative AI ``` ``` Deep Learning Machine Learning Generative AI Artificial Intelligence ``` ``` Machine Learning Deep Learning Generative AI Artificial Intelligence ``` **CORRECT:** The correct order from most general to most specialized is: 1. Artificial Intelligence 2. Machine Learning 3. Deep Learning 4. Generative AI Artificial Intelligence - Artificial Intelligence (AI) is the broadest concept. It refers to machines designed to simulate human intelligence, such as problem-solving, decision-making, and perception. AI includes many techniques, from rule-based systems to more advanced learning methods. Machine Learning - Machine Learning (ML) is a subset of AI. It focuses on enabling machines to learn from data without being explicitly programmed. ML algorithms improve their performance over time as they are exposed to more data. Deep Learning - Deep Learning is a specialized area within ML. It uses artificial neural networks with many layers (hence "deep") to model complex patterns in large datasets. Deep learning powers many modern AI capabilities like image recognition, speech processing, and natural language understanding. Generative AI - Generative AI is a specific application of deep learning focused on creating new content—like text, images, or audio. It uses models such as Generative Adversarial Networks (GANs) or Large Language Models (LLMs) to generate human-like outputs. **References:** https://docs.aws.amazon.com/whitepapers/latest/aws-caf-for-ai/aws-caf-for-ai.html Domain: Fundamentals of AI and ML --- #### 58. A global ecommerce company wants to explore generative AI applications for producing real-time product recommendations, creating chatbot responses, and translating user-generated reviews. Which factors should the company primarily consider when selecting an appropriate generative AI model? (Select TWO.) - The availability of open source ML frameworks - Performance requirements, such as latency and accuracy - Governance and compliance requirements - The number of CPU cores available - Consumer device operating systems **CORRECT:** "Governance and compliance requirements" is a correct answer. Since the company operates globally, it must ensure that its generative AI applications comply with data privacy laws, content moderation policies, and regional regulations like GDPR, CCPA, and local consumer protection laws. Governance includes ensuring AI-generated recommendations are fair, unbiased, and non-discriminatory, and that user data is handled securely. This is particularly important for AI chatbots, product recommendations, and content translation, which can directly impact customer experience and trust. **CORRECT:** "Performance requirements, such as latency and accuracy" is a correct answer. For real-time applications like product recommendations, chatbot interactions, and translation, low latency and high accuracy are critical. The AI model must be able to generate responses quickly while maintaining relevant and personalized recommendations. If latency is too high, user experience suffers, and if accuracy is low, product recommendations or translations may be irrelevant or misleading. The company should consider model efficiency, deployment infrastructure (e.g., AWS Inferentia, SageMaker), and optimization techniques (e.g., quantization, distillation) to ensure optimal performance. **INCORRECT:** "The number of CPU cores available" is incorrect. While hardware resources affect model performance, generative AI workloads typically rely on GPUs or TPUs, rather than CPU cores. Cloud-based AI services allow scaling without direct dependency on local CPU cores. **INCORRECT:** "Consumer device operating systems" is incorrect. Generative AI models are usually deployed on server-side infrastructure (e.g., AWS, Google Cloud, or on-premise data centers). While device compatibility is important for on-device AI (e.g., Apple Neural Engine, TensorFlow Lite), most real-time e-commerce AI models run in the cloud and are accessible across all operating systems. **INCORRECT:** "The availability of open-source ML frameworks" is incorrect. While open-source frameworks like Hugging Face Transformers, TensorFlow, and PyTorch are helpful for development, they are not the primary concern for selecting an AI model. Instead, the company should focus on performance, scalability, and compliance when choosing a model for production use. **References:** https://aws.amazon.com/ai/generative-ai Domain: Fundamentals of Generative AI --- #### 59. Your organization is deploying a machine learning (ML) application that transmits encrypted financial datasets between Amazon SageMaker and a private analytics application hosted on Amazon EC2. The datasets contain highly confidential information and must not traverse the public internet to meet internal data residency and regulatory compliance requirements. As the AI practitioner, which AWS-native mechanism is best suited to ensure secure, private, and compliant data transmission in transit between services across VPCs? - Use AWS PrivateLink to create a private endpoint for SageMaker to communicate with the EC2 application over an internal VPC network, preventing exposure to the public internet. - Use Amazon Macie to automatically classify and encrypt the sensitive data before transfer, ensuring all data is masked and compliant during transit. - Use Amazon CloudFront with signed URLs and HTTPS to accelerate data transfer while encrypting traffic using TLS across edge locations. - Use AWS Identity and Access Management (IAM) to enforce policies that allow secure service-to-service communication and prevent unauthorized access during data transfer. **CORRECT:** "Use AWS PrivateLink to create a private endpoint for SageMaker to communicate with the EC2 application over an internal VPC network, preventing exposure to the public internet" is the correct answer. AWS PrivateLink enables private connectivity between VPCs, services, and AWS services without exposing data to the public internet. It uses interface VPC endpoints powered by AWS PrivateLink to establish secure communication between services inside AWS. In this scenario, where encrypted financial datasets must remain within private networks to meet compliance requirements, PrivateLink is ideal. It ensures that traffic between Amazon SageMaker and the EC2-hosted analytics application stays on the AWS backbone network, which aligns with internal data residency and regulatory standards. It prevents data from traveling the public internet, reducing the risk of exposure or interception, making it the most secure and compliant AWS-native solution for in-transit protection. **INCORRECT:** "Use Amazon Macie to automatically classify and encrypt the sensitive data before transfer, ensuring all data is masked and compliant during transit" is incorrect. Amazon Macie is a data security and privacy service that uses machine learning to discover, classify, and protect sensitive data. While it helps with identifying and managing sensitive information (such as PII), it does not handle network-level security or create private communication paths between services. Macie does not ensure secure data transfer across VPCs or prevent public internet exposure. **INCORRECT:** "Use AWS Identity and Access Management (IAM) to enforce policies that allow secure service-to-service communication and prevent unauthorized access during data transfer" is incorrect. IAM is essential for controlling access to AWS resources but does not handle network routing or encryption of data in transit. It defines who can access what, but it doesn't establish private network paths or enforce encryption between services. It should be used in combination with PrivateLink, not as a replacement for network-level security. **INCORRECT:** "Use Amazon CloudFront with signed URLs and HTTPS to accelerate data transfer while encrypting traffic using TLS across edge locations" is incorrect. Amazon CloudFront is a content delivery network (CDN) designed for public distribution of web content, often used for serving media, websites, and APIs with low latency. While it uses TLS for encryption and supports signed URLs for access control, it routes data through public edge locations. This does not meet the requirement of not traversing the public internet, thus violating the data residency compliance need. **References:** https://docs.aws.amazon.com/vpc/latest/privatelink/what-is-privatelink.html https://docs.aws.amazon.com/sagemaker/latest/dg/interface-vpc-endpoint.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 60. Your team is deploying a multi-step AI model that involves processing various data sources and requires a secure and scalable environment to orchestrate the deployment of multiple machine learning models. Which AWS service is best suited to help manage and automate this multi-step machine learning workflow? - Amazon SageMaker Autopilot - Amazon SageMaker Canvas - Amazon SageMaker Pipelines - Amazon SageMaker Clarify **CORRECT:** "Amazon SageMaker Pipelines" is the correct answer. Amazon SageMaker Pipelines is a fully managed service that helps you build, automate, and manage end-to-end machine learning (ML) workflows. It is designed to handle multi-step processes, such as data preparation, model training, evaluation, and deployment. SageMaker Pipelines makes it easy to define the steps of your ML workflow in a secure and repeatable way, ensuring consistency and reducing errors. You can version, track, and reuse your workflows, which is especially useful when working with complex foundation model applications that require secure and organized orchestration. **INCORRECT:** "Amazon SageMaker Canvas" is incorrect. SageMaker Canvas is a no-code machine learning tool that allows business analysts and non-technical users to build ML models by simply uploading data and following a visual interface. While it is great for simplifying model creation without writing code, it does not provide orchestration capabilities for multi-step workflows or managing multiple models in a secure environment. Therefore, it is not suitable for complex ML pipelines. **INCORRECT:** "Amazon SageMaker Autopilot" is incorrect. SageMaker Autopilot automatically builds, trains, and tunes machine learning models based on your input data. It is ideal for automating model selection and hyperparameter tuning without deep ML expertise. However, it focuses on single-model automation, not on managing multi-step or multi-model workflows. **INCORRECT:** "Amazon SageMaker Clarify" is incorrect. SageMaker Clarify helps detect bias in machine learning models and provides explainability insights. It supports fairness and transparency by analyzing data and model predictions for potential bias. However, it does not manage ML workflows or orchestrate multi-step model processing. Its role is focused on ethical AI practices rather than workflow management. **References:** https://aws.amazon.com/sagemaker/pipelines Domain: Applications of Foundation Models --- #### 61. An animation studio is exploring generative AI to produce concept art. They've heard diffusion models are effective but want to know how these models work. Which statement best explains the core process of diffusion models? - Diffusion models begin with random noise and use a learned reverse process to form detailed images. - Diffusion models retrieve similar content from training data using semantic similarity. - Diffusion models generate images by combining encoder-decoder pairs with attention heads. - Diffusion models enhance images by sharpening pixel edges through supervised learning. **CORRECT:** "Diffusion models begin with random noise and use a learned reverse process to form detailed images" is the correct answer. Diffusion models are a type of generative AI that creates images by starting from pure random noise and gradually transforming that noise into a clear and detailed image. They do this using a process that has two main parts: adding noise to training data (forward process) and then learning to reverse that process (reverse process). During training, the model learns how to remove noise step-by-step until it recreates the original image. At inference time, the model starts with just noise and applies the reverse process to generate a brand-new image. This method has been proven very effective for generating high-quality, diverse images — making it popular for tasks like concept art generation. **INCORRECT:** "Diffusion models enhance images by sharpening pixel edges through supervised learning" is incorrect. This option describes an image enhancement technique, not how diffusion models work. Diffusion models are not primarily used to sharpen existing images but to generate new images from noise. Also, diffusion models use unsupervised or self-supervised learning rather than classic supervised learning. **INCORRECT:** "Diffusion models generate images by combining encoder-decoder pairs with attention heads" is incorrect. This statement better fits transformer-based models, such as those used in natural language processing. While some advanced diffusion models may include attention mechanisms, the core process of diffusion models does not rely on encoder-decoder pairs. **INCORRECT:** "Diffusion models retrieve similar content from training data using semantic similarity" is incorrect. This describes how a retrieval-based system works, like those in recommendation engines or semantic search. Diffusion models do not retrieve content from training data — they generate new content by sampling from learned distributions. **References:** https://aws.amazon.com/what-is/generative-ai Domain: Fundamentals of Generative AI --- #### 62. A video streaming service wants to analyze how much they earn from each subscriber over the past quarter. Which metric should they calculate to measure this? - Average Revenue Per User (ARPU) - Customer Satisfaction Score (CSAT) - Net Promoter Score (NPS) - Monthly Active Users (MAU) **CORRECT:** "Average Revenue Per User (ARPU)" is the correct answer. ARPU stands for Average Revenue Per User. It is a financial metric used by businesses, especially subscription-based services like streaming platforms, to measure how much revenue they generate per user over a specific period, such as a month or a quarter. This calculation helps businesses understand their revenue performance at the user level. ARPU is typically calculated by dividing the total revenue by the number of active users during the period. In this case, since the business wants to measure how much they earned from each subscriber over the past quarter, ARPU is the right metric. It provides a clear view of user-level profitability and helps in making data-driven business decisions, such as pricing strategies or marketing investments. **INCORRECT:** "Monthly Active Users (MAU)" is incorrect. Monthly Active Users (MAU) measure the number of unique users who engage with a service at least once within a month. While this metric shows user engagement, it does not provide any insight into how much revenue is earned per user. MAU focuses on usage, not financial performance. **INCORRECT:** "Customer Satisfaction Score (CSAT)" is incorrect. Customer Satisfaction Score (CSAT) measures how satisfied customers are with a service or product, usually collected through surveys. It is useful for understanding user experience but does not directly measure financial outcomes like revenue per user. Therefore, it is not the right metric in this context. **INCORRECT:** "Net Promoter Score (NPS)" is incorrect. Net Promoter Score (NPS) measures customer loyalty by asking how likely they are to recommend the service to others. While NPS helps gauge customer advocacy, it does not indicate the actual revenue generated per user. Thus, it does not fit the need to measure financial performance. Domain: Fundamentals of Generative AI --- #### 63. An AI-driven healthcare application must comply with regulations regarding safety, fairness, and inclusivity. Which step would be most effective to ensure responsible AI use? - Focus on reducing deployment time to maximize ROI. - Select an AI model with the highest accuracy on a public dataset. - Improve the model's performance without considering external impacts. - Implement bias detection tools, conduct human audits, and use diverse and curated data sources. **CORRECT:** "Implement bias detection tools, conduct human audits, and use diverse and curated data sources" is the correct answer. Bias in AI can lead to unfair treatment and inaccurate diagnoses, which is especially dangerous in healthcare applications. Implementing bias detection tools helps identify and mitigate unfair patterns in the data and model predictions. Human audits ensure that AI decisions align with ethical and regulatory standards. Using diverse and curated datasets reduces the risk of biased AI outputs by ensuring that the model is trained on representative data. This approach aligns with principles of Responsible AI and is crucial for compliance with regulations like HIPAA, GDPR, and AI ethics guidelines from AWS and other organizations. **INCORRECT:** "Select an AI model with the highest accuracy on a public dataset" is incorrect. While accuracy is important, it does not guarantee fairness or inclusivity. Public datasets may contain biases that lead to unfair outcomes. A model with high accuracy on one dataset may perform poorly on diverse real-world data, making it unsuitable for responsible AI use in healthcare. **INCORRECT:** "Focus on reducing deployment time to maximize ROI" is incorrect. Prioritizing speed and profit over fairness and safety can lead to regulatory non-compliance and ethical risks. A rushed deployment without proper testing, bias mitigation, and audits can cause harm, legal issues, and reputational damage. **INCORRECT:** "Improve the model's performance without considering external impacts" is incorrect. Optimizing for performance without addressing fairness, bias, and inclusivity can result in AI systems that discriminate against certain groups or make unsafe recommendations. In healthcare, ensuring ethical and safe AI use is more important than just improving performance. **References:** https://aws.amazon.com/ai/responsible-ai Domain: Guidelines for Responsible AI --- #### 64. A company wants to build a recommendation engine that can generate product suggestions based on user behavior and preferences. Which AWS services are most suited for building such a generative AI application? (Select TWO.) - Amazon SageMaker - Amazon Polly - Amazon Lex - Amazon Translate - Amazon Bedrock **CORRECT:** "Amazon SageMaker" is a correct answer. Amazon SageMaker is a comprehensive machine learning service that allows you to build, train, and deploy machine learning models, making it highly suitable for creating recommendation engines. With SageMaker, you can train models using user behavior data, such as past purchases or browsing history, to generate personalized product recommendations. SageMaker provides built-in algorithms like Factorization Machines and XGBoost, which are commonly used in recommendation systems. It also offers tools for data preprocessing and model tuning, making it a powerful solution for developing and deploying AI-driven recommendations. **CORRECT:** "Amazon Bedrock" is a correct answer. Amazon Bedrock is designed for building generative AI applications, offering pre-trained models that can be customized for specific use cases, such as recommendation engines. With Bedrock, you can leverage large language models (LLMs) to understand user preferences from text data, such as reviews or chat interactions. This service makes it easier to integrate generative AI into applications without requiring deep expertise in model training or infrastructure management, accelerating the development of recommendation systems that adapt to user preferences in real time. **INCORRECT:** "Amazon Translate" is incorrect. Amazon Translate is designed for language translation, not for building recommendation engines. It converts text from one language to another but does not analyze user behavior or preferences. **INCORRECT:** "Amazon Lex" is incorrect. Amazon Lex is used for building conversational interfaces, such as chatbots, and is not intended for generating product recommendations based on user data. **INCORRECT:** "Amazon Polly" is incorrect. Amazon Polly is a text-to-speech service that converts text into lifelike speech. While useful for voice-based applications, it does not play a role in generating recommendations from user behavior. **References:** https://aws.amazon.com/sagemaker https://aws.amazon.com/bedrock Domain: Fundamentals of Generative AI --- #### 65. Which sequence of steps best describes the lifecycle of a foundation model (FM) for generative AI? (Select and order FOUR) Note: Select only the correct options, as the type of "Ordering" question is not supported here. **CORRECT:** The correct sequence from first to last is: 1. Data Selection 2. Pre-training 3. Model Fine-tuning 4. Evaluation Data Selection - The first step in the foundation model lifecycle is selecting and preparing a diverse, high-quality dataset. Foundation models require vast amounts of data across multiple domains to learn generalizable patterns. Data preprocessing, tokenization, and filtering are crucial at this stage. Pre-training - Once the data is ready, the model undergoes pre-training on large-scale unlabeled data using self-supervised learning techniques. During this stage, the model learns general linguistic or multimodal patterns without being task-specific. This is the most computationally intensive step. Model Fine-tuning - After pre-training, the model is fine-tuned on task-specific data to improve its performance for a particular application, such as text summarization, code generation, or medical diagnosis. Techniques like LoRA (Low-Rank Adaptation) or parameter-efficient fine-tuning (PEFT) help optimize the model for specific tasks with minimal additional training. Evaluation - The final step involves evaluating the model's performance on test datasets to ensure accuracy, fairness, and robustness. Evaluation metrics vary depending on the task but may include BLEU, ROUGE, perplexity, or F1-score for text-based models. Additional safety and bias checks are performed before deployment. **References:** https://aws.amazon.com/what-is/foundation-models Domain: Fundamentals of Generative AI