Free AWS Certified AI Practitioner (AIF-C01) Sample Questions

Correct answer: A — Amazon SageMaker

• A. Correct. Amazon SageMaker is a fully managed service specifically designed to provide every developer and data scientist with the ability to build, train, and deploy machine learning models at scale. It offers a broad set of capabilities, including IDEs, debuggers, and profilers, to manage the entire ML lifecycle.
• B. Incorrect. Amazon Q is a generative AI-powered assistant designed for business intelligence, coding assistance, and answering questions about AWS services. It is not a platform for building, training, and deploying custom machine learning models.
• C. Incorrect. Amazon Bedrock is a fully managed service that makes foundation models (FMs) from leading AI startups and Amazon available via an API. While it facilitates building generative AI applications, it is not the primary service for the full build-train-deploy lifecycle of custom models.
• D. Incorrect. AWS CloudFormation is an Infrastructure as Code (IaC) service used to model and provision AWS resources. It is not a service for developing machine learning models.

Correct answer: B — Recommendation System

• A. Incorrect. A Speech Recognition System (or Automatic Speech Recognition - ASR) is designed to convert spoken language into text. It processes audio data and is unrelated to analyzing user behavior for product suggestions.
• B. Correct. A Recommendation System suggests products or content to users based on data such as past behavior (browsing history) and the preferences of similar users (collaborative filtering). This directly addresses the e-commerce use case described.
• C. Incorrect. A Fraud Detection System is used to identify anomalies or malicious activities in transactions to prevent financial loss. While it analyzes historical data, its primary goal is security and classification of legitimacy, not personalizing product offers.
• D. Incorrect. An Optical Character Recognition (OCR) System is used to extract text from images or scanned documents. It is a computer vision application and does not involve modeling user preferences or recommending items.

Correct answers: A, B — Scalability of operations.; Automation of manual translation tasks.

• A. AI enables the agency to scale operations by processing large volumes of text across multiple languages simultaneously. This allows for rapid global expansion without the proportional increase in time and labor costs that purely human translation would require.
• B. AI automates the repetitive process of translating text, drastically reducing manual effort and speeding up the time-to-market for news stories. This allows human translators to shift their focus to post-editing and quality assurance rather than translating every sentence from scratch.
• C. AI translation (such as Neural Machine Translation) is not 100% accurate and often lacks the cultural nuance, idiomatic understanding, and context sensitivity of professional human translators.
• D. The scenario specifically asks about localizing (translating) existing content, not generating new content. While Generative AI can write stories, that capability does not address the specific requirement of translating the agency's existing reporting.
• E. Identifying objects in images is a Computer Vision task. This is unrelated to the Natural Language Processing (NLP) task of translating written text.

Correct answer: B — SageMaker Data Wrangler

• A. Incorrect. Amazon EC2 is a web service providing resizable compute capacity. While it can host data preparation tools, it does not inherently offer a unified, low-code interface for data cleaning, feature engineering, or visualization; utilizing it would require significant manual setup and coding.
• B. Correct. Amazon SageMaker Data Wrangler provides a unified, visual interface that allows users to import, prepare, transform, feature engineer, and analyze data distributions with little to no code. It is specifically designed to accelerate the data preparation phase of the ML lifecycle.
• C. Incorrect. Amazon S3 is an object storage service used for scalability, data availability, and security. It acts as a storage layer for datasets but does not provide the compute capabilities or interface required for interactive data preparation or visualization.
• D. Incorrect. SageMaker Inference is focused on the deployment phase of the ML lifecycle, enabling the hosting and serving of models to generate predictions. It is not used for the initial data preparation or exploratory data analysis phases.

Correct answer: B — Real-time Inference Endpoint (Managed API)

• A. Incorrect. Batch Transform is designed for processing large datasets offline in bulk, typically on a schedule. It is not suitable for scenarios requiring immediate, low-latency responses to individual user actions.
• B. Correct. A Real-time Inference Endpoint (Managed API) is designed to handle synchronous requests and provide low-latency predictions. This is the appropriate method for delivering instant product recommendations while a user is actively browsing a website.
• C. Incorrect. Offline Reporting is used for analyzing historical data to generate business insights or dashboards. It is not capable of serving live, per-request predictions needed for a real-time user experience.
• D. Incorrect. Manual processing involves human intervention or ad-hoc workflows, which cannot meet the speed, scalability, and automation requirements of a real-time recommendation system.

Correct answers: A, B — F1 Score; Accuracy

• A. The F1 Score is a standard metric for evaluating classification models, particularly when class distributions are imbalanced. It represents the harmonic mean of precision and recall, providing a balanced view of the model's ability to minimize both false positives and false negatives.
• B. Accuracy is a fundamental metric for classification models that measures the proportion of correct predictions (both true positives and true negatives) relative to the total number of predictions. While useful, it is often analyzed alongside other metrics like the F1 score to account for dataset imbalances.
• C. Customer Satisfaction Score (CSAT) is a business performance metric focused on user experience and sentiment. While it may be influenced by model performance, it does not directly measure the technical accuracy or predictive capability of the model itself.
• D. Server CPU Utilization is an operational or infrastructure metric. It measures the resource consumption of the environment hosting the model but does not provide insight into the quality or correctness of the model's predictions.
• E. Total Cloud Spend is a financial metric related to the cost of operations and infrastructure. It is critical for budgeting and optimization but is not a measure of the technical performance or predictive quality of a classification model.

Correct answer: A — The output artifact generated after an algorithm trains on data.

• A. Correct. In machine learning, a model is the output artifact generated after a specific algorithm (the logic) is trained on a dataset. It contains the learned patterns, parameters (such as weights and biases), and structure required to make predictions on new data.
• B. Incorrect. This describes the training data or dataset. While data is an essential input for creating a model, it is not the model itself.
• C. Incorrect. This refers to the compute resources (such as CPUs, GPUs, or AWS Trainium) used to execute the training process. The model is software/data, not physical infrastructure.
• D. Incorrect. This describes the machine learning algorithm (e.g., Linear Regression, Random Forest). The algorithm is the set of rules or mathematical logic used to learn from data, whereas the model is the result of applying that algorithm to specific data.

Correct answer: A — Computer Vision

• A. Computer Vision is the field of AI specifically dedicated to enabling computers to identify, process, and derive meaningful information from images, videos, and other visual inputs. It uses techniques like object detection and image classification to recognize defects.
• B. Natural Language Processing (NLP) focuses on the interaction between computers and humans through natural language (text or speech). It is designed for linguistic analysis, not for processing visual data from cameras.
• C. Reinforcement Learning is a type of machine learning where an agent learns to make a sequence of decisions by performing actions and receiving rewards. While it can be used for robot control, it is not the primary field for analyzing visual content.
• D. Speech-to-Text is a technology that converts spoken language into written text. It processes auditory data and is unrelated to visual analysis.

Correct answer: B — The model is unaware of events or data created after the training period ended.

• A. Incorrect. The cutoff date refers to the temporal limit of the training data, not the character or token length of the input text the model can process.
• B. Correct. The cutoff date represents the point in time when the training data collection ended. Consequently, the model's internal knowledge base is static and does not include any information, events, or developments that occurred after that date.
• C. Incorrect. The model remains operational after the cutoff date. The limitation is strictly regarding the currency of its knowledge, not its ability to function or generate responses.
• D. Incorrect. Pre-trained models rely on static weights derived from training; they do not actively delete data. The cutoff date simply means no new data was added after that point, not that previous data is removed.

Correct answers: A, B — Hallucinations; Lack of interpretability

• A. Hallucinations are a significant risk in Generative AI, where the model produces confident-sounding but factually incorrect, fabricated, or nonsensical information. This can lead to misinformation and requires rigorous verification mechanisms.
• B. Lack of interpretability, often referred to as the 'black box' nature of deep learning models, is a recognized disadvantage. It is difficult to trace exactly how the model processes input to arrive at a specific output, making auditing, debugging, and explaining decisions challenging.
• C. This is incorrect. Generative AI models, particularly Large Language Models (LLMs), are explicitly designed to process, understand, and generate text efficiently.
• D. This is incorrect. Unlike traditional machine learning models that often require structured tabular data, Generative AI excels at processing unstructured data types such as text, images, and audio.
• E. This is incorrect. Generative AI models are inherently probabilistic (stochastic) rather than deterministic. They often produce different outputs for the same input depending on sampling parameters (like temperature), rather than a single fixed outcome.

Correct answer: B — Tokens

• A. Incorrect. Pixels are the smallest unit of a digital image or display. While multimodal models may process pixels, they are the fundamental unit for computer vision, not for text processing.
• B. Correct. Tokens are the fundamental units of text that LLMs process and generate. Before an LLM processes text, a tokenizer breaks the input down into tokens, which can represent whole words, sub-words, or characters.
• C. Incorrect. Bytes are the basic units of digital information and storage. Although text is encoded in bytes and some tokenization strategies are byte-based (like Byte-Pair Encoding), the model itself operates on the discrete tokens derived from the data.
• D. Incorrect. Vectors (or embeddings) are internal high-dimensional numerical representations of data. While tokens are converted into vectors for the model to perform mathematical operations, vectors are the format for computation rather than the fundamental unit of text itself.

Correct answer: B — Prompt engineering

• A. Incorrect. Model training refers to the computational process of updating a model's weights and parameters using a large dataset to learn patterns. The scenario describes using an existing, pre-trained model for inference, not training it.
• B. Correct. Prompt engineering is the practice of crafting and structuring input text (prompts) to guide a generative AI model to produce a specific desired output, such as performing a sentiment classification task.
• C. Incorrect. Data embedding involves converting high-dimensional data (like text) into lower-dimensional numerical vectors to capture semantic meaning. It is a data representation technique, not the method of instructing a model via natural language.
• D. Incorrect. Vectorization is the general process of converting non-numeric data into numerical formats (vectors) that algorithms can process. While part of the AI pipeline, it is not the technique of guiding model behavior through instructions.

Correct answers: A, C — Eliminates the need to manage physical infrastructure and patching; Accelerates speed to market via accessible APIs

• A. Correct. A primary benefit of using AWS managed services is the abstraction of the underlying infrastructure. AWS handles the undifferentiated heavy lifting, such as procuring physical servers, managing data center facilities, and performing OS and software patching. This reduces operational overhead and allows teams to focus on developing and deploying their Generative AI applications rather than managing hardware.
• B. Incorrect. AWS does not provide 100% free access to all Foundation Models. While some open-source models may not have licensing fees, using them through managed services like Amazon Bedrock or Amazon SageMaker incurs costs for compute, storage, and API invocations. Proprietary models also have specific pricing models.
• C. Correct. AWS managed services for GenAI, such as Amazon Bedrock, provide high-level, accessible APIs and SDKs. This simplifies the integration of powerful foundation models into applications, reducing development complexity and allowing developers to build and deploy features more quickly, thereby accelerating the time to market.
• D. Incorrect. AWS maintains strict security over its physical infrastructure, and customers are not granted physical access to data centers. Customers interact with AWS resources virtually through the AWS Management Console, APIs, and SDKs.
• E. Incorrect. The AWS shared responsibility model is a fundamental concept that always applies. While managed services shift more responsibility for the infrastructure (security *of* the cloud) to AWS, the customer remains responsible for their data, access controls, and configuration (security *in* the cloud). The model is adjusted, not removed.

Correct answer: B — It provides a hub to access, deploy, and fine-tune a wide range of open-source and proprietary models.

• A. Incorrect. SageMaker JumpStart is not limited to Amazon-proprietary models like Titan. It provides access to a broad collection of models, including a large number of popular open-source and third-party foundation models.
• B. Correct. Amazon SageMaker JumpStart acts as a centralized ML hub where users can discover, deploy, and fine-tune a diverse set of pre-trained models. It simplifies the process by offering pre-built solutions, model packages, deployment templates, and integrated fine-tuning workflows, significantly accelerating the path from model selection to deployment.
• C. Incorrect. A key benefit of SageMaker JumpStart is that it eliminates the need for users to manually find, download, and configure model weights from various external websites. It provides pre-packaged model artifacts and integrates directly with model repositories for a streamlined experience.
• D. Incorrect. SageMaker JumpStart supports a wide variety of tasks and model types, including text generation (LLMs), image generation, and multimodal models. Limiting it to only one modality like image generation is a false statement.

Correct answer: B — Jailbreaking

• A. Incorrect. Model hallucination refers to instances where a model generates factually incorrect, nonsensical, or ungrounded information presented as fact. It is typically a model error or limitation rather than an intentional adversarial attack designed to bypass safety filters.
• B. Correct. Jailbreaking is a specific form of prompt injection attack where a user crafts sophisticated or deceptive prompts (such as role-playing scenarios) to trick the AI into ignoring its safety training and guardrails, thereby allowing the generation of restricted or harmful content.
• C. Incorrect. Data poisoning involves corrupting or manipulating the training dataset (e.g., inserting malicious samples) to alter the model's behavior during the learning phase. It is a training-time attack, whereas jailbreaking is an inference-time prompt attack.
• D. Incorrect. Prompt routing is an architectural pattern used to direct prompts to specific models or components based on the complexity or type of the query (e.g., sending simple queries to a smaller model). It is a legitimate engineering technique, not an adversarial attack.

Correct answers: A, B — Prompt Injection; Jailbreaking

• A. Correct. Prompt Injection is a security vulnerability where an attacker crafts malicious inputs to override the original instructions of an LLM, causing it to perform unintended actions or reveal sensitive information.
• B. Correct. Jailbreaking refers to specific prompt engineering techniques designed to bypass an LLM's safety guardrails and content filters, tricking the model into generating restricted, harmful, or unethical content.
• C. Incorrect. High latency is a performance and operational issue regarding the time it takes for a model to respond. While it affects user experience, it is not a security risk specific to prompt engineering.
• D. Incorrect. Overfitting is a machine learning concept where a model learns the training data too specifically, leading to poor generalization on new data. It is a model training issue, not a prompt engineering security risk.
• E. Incorrect. Data normalization is a standard data preprocessing technique used to scale and structure data efficiently. It is a best practice for data preparation, not a security threat.

Correct answer: B — Modality

• A. Incorrect. Latency refers to the time delay between sending a request and receiving a response. It is a performance metric measuring speed, not the type of data.
• B. Correct. Modality refers to the specific types of data (e.g., text, images, audio, video) that a foundation model can process as input or generate as output. A model that handles more than one type is often called multimodal.
• C. Incorrect. Throughput measures the amount of data or number of requests processed within a specific time period (e.g., tokens per second). It is a metric for capacity and speed, not data classification.
• D. Incorrect. Embeddings are numerical vector representations of data used to capture semantic meaning. While embeddings are derived from data of various modalities, the term itself refers to the vector format, not the data type.

Correct answer: B — Retrieval Augmented Generation (RAG)

• A. Incorrect. Pre-training a new model is the most resource-intensive approach, involving training a model from scratch on a large dataset. It is not feasible for frequently changing documents because the model would need to be retrained constantly to incorporate new information.
• B. Correct. Retrieval Augmented Generation (RAG) connects the foundation model to an external knowledge source (such as a vector database) that can be updated independently of the model. When a user asks a question, the system retrieves the most current relevant documents to generate the answer. This ensures up-to-date responses without modifying the model's weights.
• C. Incorrect. Instruction fine-tuning involves further training a pre-trained model on a smaller, specific dataset to improve performance on specific tasks. While less expensive than pre-training, it still requires a training cycle to update knowledge, making it inefficient for data that changes frequently.
• D. Incorrect. Reinforcement Learning from Human Feedback (RLHF) is an alignment technique used to align a model's behavior with human preferences (e.g., helpfulness and safety). It is not a mechanism for providing real-time access to dynamic external data.

Correct answers: A, B — Access to proprietary or real-time data; Reduced risk of hallucinations by grounding answers in retrieved data

• A. RAG enables models to incorporate information from external documents or data sources during inference. This allows them to access proprietary, domain-specific, or up-to-date (real-time) information that is not present in the static, pre-trained model's knowledge base, leading to more relevant and current responses.
• B. By grounding responses in retrieved documents, RAG significantly reduces the likelihood of hallucinations. The model bases its answers on explicit source content provided in the context, rather than relying solely on its internal parametric memory, which can sometimes generate plausible but incorrect information.
• C. RAG introduces an additional retrieval step before the generation phase, which typically increases, rather than decreases, the overall end-to-end latency. While some complex queries might see faster generation due to concise context, it does not guarantee faster inference for all queries.
• D. The ability to generate images from text is a function of specialized image-generation models, such as diffusion models, not a feature of RAG. RAG is a technique focused on retrieving textual context to augment text generation.
• E. RAG does not eliminate the need for prompt engineering. In fact, effective prompt engineering is crucial for both formulating the initial query to the retrieval system and for structuring the final prompt to the large language model, which includes the retrieved context, to generate a high-quality response.

Correct answer: C — Model Distillation

• A. Reinforcement Learning (RL) is a technique where a model learns by taking actions in an environment to maximize a reward signal. While used in fine-tuning (e.g., RLHF), it is not a compression method involving a student-teacher architecture.
• B. Continuous Pre-training involves training an already pre-trained model on additional, often domain-specific data to update its knowledge or adapt it to a new domain. It does not involve training a smaller model to mimic a larger one.
• C. Model Distillation (or Knowledge Distillation) is the process of transferring knowledge from a large, complex 'teacher' model to a smaller, more efficient 'student' model. The student is trained to replicate the teacher's outputs, allowing it to achieve similar performance with reduced latency and operational costs.
• D. Instruction Tuning is a fine-tuning process using datasets formatted as instruction-response pairs to improve a model's ability to follow user commands. It focuses on alignment and utility rather than model compression.

Correct answer: A — Training a model on a dataset of prompt-response pairs to teach it to follow directions

• A. This is the correct definition. Instruction tuning is a form of supervised fine-tuning (SFT) where a pre-trained model is further trained on a high-quality dataset of instruction-response pairs. This process explicitly teaches the model how to follow user commands, adhere to specific formats, and generalize its ability to respond appropriately to a wide variety of prompts it hasn't seen before.
• B. This describes model optimization or compression techniques, such as pruning or quantization, not instruction tuning. These methods focus on reducing the model's size and improving inference speed, whereas instruction tuning focuses on aligning the model's behavior with user intent.
• C. This describes a highly specific, task-oriented fine-tuning for a single domain. Instruction tuning is a more general process that uses a diverse dataset of instructions across many different tasks to improve the model's overall instruction-following capabilities, not restrict it to a narrow task like financial forecasting.
• D. This describes data preprocessing or data curation. This is a crucial step that occurs before the initial pre-training of a foundation model. Instruction tuning is a distinct, subsequent phase of training that happens after pre-training to refine the model's behavior.

Correct answer: C — Parameter-Efficient Fine-Tuning (PEFT)

• A. Full fine-tuning involves updating all of the model's parameters. This process is highly resource-intensive, requiring substantial computational power, memory, and time, making it impractical and unsuitable for scenarios with a limited budget.
• B. Pre-training from scratch is the process of training a new model from the beginning on a massive dataset. It is the most computationally expensive and time-consuming approach, requiring extreme amounts of data and compute resources, and is therefore completely inappropriate for an organization with a limited budget looking to adapt an existing model.
• C. PEFT methods are specifically designed for situations with limited computational resources. Instead of updating all model parameters, PEFT techniques (such as LoRA, adapters, or prompt tuning) modify or add only a small subset of parameters. This drastically reduces the compute, memory, and storage requirements for fine-tuning, making it the most appropriate and efficient approach for the retailer's situation.
• D. RLHF is a technique used to align a model's outputs with human preferences, not primarily to reduce computational costs. It typically adds a layer of complexity to the training process, often requiring significant human annotation effort and additional computational resources for the reinforcement learning phase, making it unsuitable for a budget-constrained fine-tuning task.

Correct answer: A — ROUGE (Recall-Oriented Understudy for Gisting Evaluation)

• A. ROUGE is the standard metric for evaluating automatic summarization. It specifically measures the overlap of n-grams (sequences of words) between the generated text and one or more human-written reference summaries, focusing primarily on recall to determine how much of the reference content appears in the generation.
• B. Mean Squared Error (MSE) is a regression metric used to measure the average squared difference between predicted and actual numerical values. It is not designed for text generation tasks and cannot evaluate textual overlap or semantic meaning.
• C. Latency is a performance metric that measures the time delay between a user's request and the model's response. While important for user experience, it evaluates the speed of the system rather than the quality or accuracy of the generated summary.
• D. Tokens per second is a throughput metric that measures the generation speed of the model. Like latency, it assesses operational efficiency rather than the content quality or the similarity of the generated summary to a reference text.

Correct answer: A — To provide a standardized way to compare the general capabilities of different models

• A. Correct. Benchmark datasets like MMLU (Massive Multitask Language Understanding) and HELM (Holistic Evaluation of Language Models) provide a standardized set of tasks, data, and metrics. This allows researchers and practitioners to objectively evaluate and compare the general capabilities, robustness, and performance of different foundation models in a fair and consistent manner.
• B. Incorrect. Foundation Models are trained on massive, broad datasets. Benchmark datasets are specifically designed and reserved for evaluation (testing) after a model has been trained. Training a model on a benchmark dataset would be a form of 'data leakage' or 'teaching to the test,' which would invalidate its scores on that benchmark.
• C. Incorrect. Reducing model inference latency is an optimization task addressed by techniques like model quantization, pruning, using specialized hardware, or improving serving architecture. While some benchmarks might measure latency as a performance metric, their primary purpose is to evaluate model capabilities, not to implement optimizations that reduce latency.
• D. Incorrect. Model compression for deployment on resource-constrained devices like mobile phones is achieved through specific techniques such as quantization, distillation, or pruning. Benchmark datasets are not the method for compression; rather, they can be used to evaluate the performance of the compressed model to see how much accuracy was lost during the compression process.

Correct answer: B — When a model generates confident but factually incorrect information

• A. Incorrect. This describes a safety mechanism or guardrail, where a model abstains from answering based on content policies. This is a deliberate design choice for responsible AI, not an error in content generation like a hallucination.
• B. Correct. A hallucination is a phenomenon where a generative model produces information that is presented confidently but is factually incorrect, nonsensical, or fabricated. This occurs due to factors like overgeneralization or gaps in the training data and is a primary concern for ensuring the trustworthiness and accuracy of AI systems.
• C. Incorrect. This describes memorization or data leakage, where a model reproduces its training data verbatim. While also a concern for responsible AI (posing privacy and copyright risks), it is distinct from hallucination, which involves the creation of new, incorrect information.
• D. Incorrect. A model crashing due to insufficient memory is a technical, system-level failure related to infrastructure or runtime resources. A hallucination is a content-level issue concerning the semantic correctness of the model's output, not its operational stability.

Correct answer: A — The dataset contains an equal number of samples for each target class or category

• A. This is the correct definition. A balanced dataset contains an equal or approximately equal number of samples for each target class or category. This is a crucial aspect of responsible AI because it helps prevent the model from developing a bias towards majority classes, leading to fairer and more reliable performance across all groups.
• B. The size of a dataset is not a determinant of its balance. A dataset can be massive (e.g., petabytes in size) but still be highly imbalanced if one class overwhelmingly outnumbers the others. Balance refers to the distribution of samples across classes, not the total volume of data.
• C. The source of the data, whether real or synthetic, does not determine if a dataset is balanced. While generating synthetic data is a common technique to address class imbalance (a process called oversampling), a dataset composed entirely of synthetic data is not guaranteed to be balanced. Its balance depends on the distribution used during generation.
• D. Encryption is a security measure used to protect data confidentiality and privacy. It has no bearing on the statistical properties of the dataset, such as the distribution of samples across classes. A dataset's balance is a characteristic of its composition, not its security state.

Correct answers: A, C — Intended uses of the model; Performance metrics and limitations

• A. Correct. A key part of a model card is defining its intended use cases and target users. This clarifies the model's scope, helps prevent misuse, and provides essential context for evaluating its performance and limitations, which is fundamental to achieving transparency.
• B. Incorrect. Including sensitive credentials like a root password is a severe security violation and is completely unrelated to the purpose of a model card. Model cards focus on documenting a model's characteristics and performance, not operational secrets or infrastructure credentials.
• C. Correct. This is a core component of a model card. Documenting performance metrics (e.g., accuracy, fairness metrics) and known limitations provides a transparent view of the model's capabilities and potential shortcomings. This information is crucial for stakeholders to make informed decisions and understand the risks associated with the model's deployment.
• D. Incorrect. Sensitive financial data like credit card information is entirely irrelevant to model documentation and transparency. Including it would be a major security and privacy breach. Model cards are for technical and governance documentation, not account billing information.
• E. Incorrect. The underlying operating system's binary code is low-level system information that is irrelevant to the purpose of a model card. A model card focuses on the model itself—its purpose, training, and performance—not the operating system it was trained or deployed on.

Correct answer: B — Display a 'Why am I seeing this?' explanation alongside recommendations

• A. Increasing the training dataset size might improve the model's accuracy or its ability to generalize, but it does not address the core issue of transparency. It fails to provide customers with the reasons behind the recommendations, which is necessary to resolve their confusion.
• B. This is the most direct solution as it directly addresses the principles of transparency and explainability. By providing a clear rationale for each recommendation, the system helps users understand the factors driving the suggestions. This approach builds trust, enhances the user experience, and is a best practice for responsible AI.
• C. Switching to a more complex model, such as a deeper neural network, often has the opposite effect on explainability. These models are typically more of a 'black box' and harder to interpret, which would likely increase, not decrease, user confusion.
• D. While this would technically eliminate the user confusion, it is an extreme and impractical solution. It sacrifices a valuable business feature and fails to solve the underlying problem. The goal should be to improve the feature's transparency, not discard it entirely.

Correct answers: B, C — Compliance with regulatory audit requirements; Increased trust from stakeholders and customers

• A. Incorrect. Explainable AI (XAI) focuses on making model decisions transparent and understandable, not on optimizing performance metrics like training speed. In some cases, implementing explainability can add computational overhead, potentially increasing training time.
• B. Correct. In highly regulated industries like banking, there are strict requirements for auditing and justifying automated decisions (e.g., loan approvals/denials). Explainable AI provides the necessary transparency and documentation to demonstrate compliance with these regulations and to provide clear rationale during audits.
• C. Correct. When stakeholders, including customers, regulators, and internal users, can understand the reasoning behind an AI model's decisions, it fosters trust and confidence. This transparency is crucial for the adoption and acceptance of AI systems, especially for critical applications in banking.
• D. Incorrect. Model explainability is unrelated to cloud infrastructure costs. The techniques used for XAI do not reduce or eliminate the need for cloud storage or other computational resources.
• E. Incorrect. Synthetic data generation is a separate technique used to create artificial data for training models, often to address privacy concerns or data scarcity. It is not a direct benefit or outcome of using explainable AI models.

Correct answer: B — AWS Shared Responsibility Model

• A. Incorrect. The AWS Well-Architected Framework provides architectural best practices and pillars (including Security) for designing and operating reliable and efficient systems. While it helps customers implement security, it defines design principles rather than the formal division of obligations between parties.
• B. Correct. The AWS Shared Responsibility Model explicitly defines the division of security obligations. AWS is responsible for the "security of the cloud" (infrastructure, hardware, foundational services), while the customer is responsible for "security in the cloud" (data, configuration, access management). This concept applies to all workloads, including AI.
• C. Incorrect. The AWS Cloud Adoption Framework (AWS CAF) offers guidance for organizational change and digital transformation strategies during cloud adoption. It focuses on planning and business outcomes rather than the specific delineation of security liabilities.
• D. Incorrect. The AWS Acceptable Use Policy (AUP) outlines prohibited activities and rules for proper conduct when using AWS services. It defines what a user is allowed or forbidden to do, not how security responsibilities are shared between the provider and the consumer.

Correct answer: C — Prompt injection

• A. Incorrect. A Distributed Denial of Service (DDoS) attack focuses on availability by overwhelming a system with traffic to disrupt service. It does not involve crafting inputs to manipulate the logic or safety guidelines of an LLM.
• B. Incorrect. SQL injection involves inserting malicious code into database queries to manipulate backend databases. While strictly defined input validation is important, this attack targets database query logic rather than the instruction-following behavior of an LLM.
• C. Correct. Prompt injection is a specific vulnerability where an attacker uses crafted inputs (prompts) to override the system instructions or safety guardrails of an LLM. This manipulation can trick the model into performing unauthorized actions or revealing sensitive information.
• D. Incorrect. A Man-in-the-middle (MitM) attack involves intercepting and potentially altering communications between two parties. While it compromises confidentiality, it does not rely on crafting specific prompts to manipulate the internal behavior of the AI model.

Correct answer: A — Guardrails for Amazon Bedrock

• A. Correct. Guardrails for Amazon Bedrock allows organizations to implement safeguards customized to their application requirements and responsible AI policies. It specifically includes content filters to detect and block toxic, harmful, or inappropriate content (such as hate speech, insults, and violence) in both user inputs and model responses.
• B. Incorrect. Amazon Bedrock Knowledge Bases are used to implement Retrieval-Augmented Generation (RAG) by connecting foundation models to company data sources to improve relevance and accuracy. While this grounds the model in facts, it does not specifically provide content filtering or toxicity prevention.
• C. Incorrect. Provisioned Throughput is a capacity mode that ensures a defined level of performance and availability for model inference. It relates to scaling and latency management, not content safety or moderation.
• D. Incorrect. Amazon Bedrock Agents are designed to orchestrate and execute multi-step tasks by interacting with external systems and APIs. While they manage workflows, they do not inherently provide the content moderation or safety filtering capabilities required to prevent toxic output.

Correct answers: A, B — Transparency standards; Team training requirements

• A. Correct. Transparency standards are a core element of AI governance. They ensure that AI systems are understandable, interpretable, and auditable. This includes practices like maintaining detailed documentation, using model cards, implementing explainability techniques, and tracking data provenance, which are all essential for building trust, ensuring accountability, and complying with regulations.
• B. Correct. Team training is a fundamental component of AI governance. It ensures that all stakeholders, from data scientists to business leaders, understand the organization's policies, ethical guidelines, legal obligations, and risk controls related to AI. Proper training promotes responsible AI development and deployment, mitigates risks, and ensures consistent application of governance principles.
• C. Incorrect. Hardware acceleration is a technical consideration related to infrastructure and performance, focusing on improving the computational speed and efficiency of AI models. It is not a component of a governance framework, which deals with policies, ethics, risk management, and oversight.
• D. Incorrect. Model latency optimization is an operational performance metric focused on reducing the time it takes for a model to generate a prediction. While important for user experience and system efficiency, it is not a primary concern of an AI governance framework, which prioritizes principles like fairness, accountability, and transparency.
• E. Incorrect. Cost allocation tags are a financial management tool used for tracking cloud resource usage and attributing costs. While they can be part of broader operational governance, they are not a primary component of an AI governance framework, which is concerned with the ethical, legal, and responsible use of AI systems.

Correct answer: A — To classify AI use cases based on risk and identify appropriate security controls

• A. Correct. The Generative AI Security Scoping Matrix is a governance tool designed to help organizations classify their AI use cases by risk level. It provides a structured framework to map these risks to appropriate security, governance, and policy controls, thereby helping to prioritize and implement mitigations effectively.
• B. Incorrect. The matrix is focused on risk assessment and security control selection, which are governance activities. Calculating the financial cost of training models is a cost-management or FinOps activity, not the primary purpose of this security tool.
• C. Incorrect. The Generative AI Security Scoping Matrix is a framework for assessment and planning, not an automation or code generation tool. While its findings might inform the creation of security configurations, it does not automatically produce code artifacts like Python scripts for security groups.
• D. Incorrect. Encrypting data in transit is a specific security control. The matrix is a higher-level governance tool used to determine which controls, such as encryption, are necessary based on a use case's risk profile. The matrix itself does not perform the encryption.

Correct answers: A, B — AWS Trusted Advisor; AWS Artifact

• A. AWS Trusted Advisor is an online tool that provides real-time guidance to help you provision your resources following AWS best practices. It inspects your AWS environment and makes recommendations for saving money, improving system performance and reliability, and closing security gaps. It is the appropriate service for ensuring the environment follows security best practices.
• B. AWS Artifact is a central resource for compliance-related information. It provides on-demand access to AWS's security and compliance reports (such as SOC, ISO, and PCI) and select online agreements. This is the correct service for obtaining compliance documentation to share with stakeholders like investors.
• C. Amazon SageMaker is a fully managed service for building, training, and deploying machine learning models. While it is central to deploying the AI solution itself, it is not a tool for assessing the overall AWS account's security posture or for accessing AWS compliance reports.
• D. AWS Lambda is a serverless, event-driven compute service. It is used for running application code and is not designed for environment-wide security assessments or providing access to compliance documentation.
• E. Amazon EC2 provides scalable computing capacity (virtual servers) in the AWS Cloud. It is an infrastructure component for running workloads, not a governance service that provides automated security best-practice checks or access to compliance reports.