AWS Certified AI Practitioner Practice Test (AIF-C01)

Bias is the systematic error that causes a model's predictions to be consistently skewed toward certain values or groups. This usually stems from assumptions within the algorithm or from imbalanced or representative issues in the training data. Random noise is variance, not bias. The count of parameters merely reflects model size, and weight updates during backpropagation describe the learning process, not bias itself.

Conversion rate directly ties customer actions to revenue. If a higher percentage of site visitors who see the recommendation banner complete a purchase, the model is creating measurable value for the business. Average handle time measures support efficiency, not sales impact. Token usage cost tracks model-execution expenses, not income. Embedding vector dimensionality is a technical attribute that has no inherent connection to revenue generation.

Amazon Bedrock is a fully managed service that hosts foundation models and exposes them through an API. Because Bedrock handles the underlying compute infrastructure, the company can deploy and scale its fine-tuned model without setting up or operating GPU instances.

• Amazon S3 is an object-storage service; it cannot run inference.
• Amazon EC2 Auto Scaling adds or removes virtual machines but still requires customers to build and manage the inference stack on those instances.
• AWS CloudFormation automates infrastructure provisioning but does not provide a managed model-hosting endpoint by itself.

Amazon Transcribe is a fully managed AI service that performs automatic speech recognition to convert audio streams or files into accurate text transcripts. Amazon Polly does the opposite (text-to-speech), Amazon Comprehend analyzes existing text for sentiment and entities, and Amazon Rekognition analyzes images and video, so none of those services meet the need to turn speech into text.

AI/ML adds value when patterns must be learned from historical data and can change over time, such as fraud detection, demand forecasting, or personalized recommendations. Calculating sales tax, however, follows fixed percentages defined by law. Because the rules are explicit, infrequently updated, and produce a single deterministic answer, a simple rule-based program is cheaper, easier to maintain, and guarantees the exact outcome required for regulatory compliance. Therefore, an AI/ML approach would be unnecessary and unjustified for this task, whereas the other scenarios inherently involve prediction under uncertainty where ML can provide clear benefits.

The automatic search of multiple hyperparameter combinations corresponds to the hyperparameter tuning stage of the ML pipeline. Amazon SageMaker Automatic Model Tuning (sometimes called hyperparameter tuning jobs) runs many training jobs in parallel, evaluates them against a chosen metric, and returns the model that performs best. SageMaker Ground Truth is for data labeling, Batch Transform is for offline inference, and a Model Package is a container format used for sharing or deploying an already trained model, so none of these handle hyperparameter optimization.

Artificial intelligence is the overarching discipline focused on building computer systems that can perform tasks normally requiring human intellect, such as perception, reasoning, and decision-making. This broad goal may be achieved with many techniques, including-but not limited to-machine learning and deep learning. Choice describing AI as the ability to learn from data using multi-layer neural networks is actually deep learning, a subset of machine learning. The option stating that AI automatically tunes weights with back-propagation also refers to machine or deep learning specifics, not AI as a whole. A database technology that stores text and image data is unrelated to the definition of AI.

Recommender systems analyze users' historical interactions-such as viewing and rating behaviour-to predict and surface items each user is likely to enjoy. This personalization objective is not addressed by fraud detection (focused on anomalous activity), speech recognition (turning audio into text), or image classification (labelling visual content).

The scenario describes creating new images from textual descriptions, which is characteristic of the image generation use case. Summarization condenses text, translation converts language, and recommendation engines suggest items based on user data; none of these tasks involve synthesizing new visual content from prompts.

Customer lifetime value (CLV) captures the total revenue a business expects from an individual customer across the entire relationship and therefore shows whether generative-AI recommendations are increasing long-term spending. Average order value (AOV) looks only at a single transaction, customer acquisition cost (CAC) measures how much it costs to obtain a new buyer, and net promoter score (NPS) reflects satisfaction rather than revenue. None of these alternatives directly indicate the cumulative revenue impact per customer that CLV provides.

The described scenario is prompt poisoning. In a prompt-poisoning attack, adversaries inject harmful or misleading instructions into the data a foundation model later consumes (for example, documents indexed for RAG). When the model retrieves this tainted content, the malicious instructions become part of the prompt, leading the model to generate compromised outputs. Prompt hijacking instead refers to an attacker taking over an in-progress conversation, while jailbreaking tries to bypass a model's safety filters. Model drift is an unrelated performance change that happens gradually as data or requirements evolve, not from deliberate malicious instructions.

Amazon DocumentDB offers native vector search indexes, letting teams persist embeddings and run similarity queries while using familiar MongoDB drivers and tools. Amazon Neptune is a graph database. Amazon DynamoDB is a key-value/NoSQL service that does not have native vector search capabilities but can integrate with other services like Amazon OpenSearch Service to provide it. Amazon ElastiCache for Redis provides in-memory caching rather than being primarily a managed document database.

Detecting sentiment in chat transcripts is a natural language processing task, so the model must learn from examples of written language. Labeled text documents that include sentiment tags (for example, positive, neutral, or negative) provide the necessary input-output pairs for supervised learning. Time-series sensor data, numeric tabular records, and labeled images do not contain the textual information required to teach a model about language sentiment, making them unsuitable for this scenario.

Machine learning is defined as the field of study that enables computers to improve their performance at a task by learning patterns from data, rather than relying on explicitly coded instructions. This distinguishes ML from approaches where developers manually encode every decision rule. Manually labeling data is often part of preparing supervised ML, but it is not the definition of ML itself. Likewise, ML does not specifically refer to specialized hardware acceleration.

RAG solutions first retrieve content relevant to the user's question and then pass that content to a foundation model for generation. Amazon Bedrock supports this by connecting to a vector database where embeddings of the source documents are stored. When a query arrives, the vector store performs similarity search, returns the most relevant passages, and those passages are appended to the prompt that Bedrock sends to the model. Fine-tuning the model, granting the model blanket S3 access, or only adjusting inference parameters does not provide retrieval capabilities. Therefore, adding a vector store such as an Amazon OpenSearch Service index for embeddings is required to implement RAG.

Amazon Bedrock charges by the number of tokens processed by the model during inference. This includes both the input tokens contained in the prompt and the output tokens returned by the model. Storage, regional deployment, and IAM user counts incur separate charges (such as S3 storage or data transfer) or no charge at all, but they do not change the token fee that Bedrock applies for each model invocation.

Removing personal identifiers and filtering out policy-violating or off-topic conversations creates a domain-specific, compliant corpus. This improves the model's ability to answer retail queries accurately while reducing privacy risks and the chance that it learns undesirable behavior. Simply retaining all raw logs, relying on unrelated public articles, or using synthetic data without review either adds irrelevant content, preserves sensitive data, or risks compounding model errors, all of which hurt fine-tuning quality.

Politeness and helpfulness are subjective, qualitative attributes that automated metrics such as BLEU, latency, or token counts cannot reliably capture. Instead, organizations typically rely on human evaluation: reviewers read sample responses and rate them against a rubric that defines desired qualities (for example, politeness levels or usefulness of information). Human judgment can account for nuance, tone, and context in ways automated statistics cannot. BLEU focuses on n-gram overlap with reference text, latency measures performance speed, and token counts track output length-none of which directly indicates whether answers are polite or helpful.

Using a fully managed foundation model through simple prompt calls removes the need for data scientists, infrastructure setup, and model training. This ease of adoption demonstrates the simplicity benefit of generative AI. Redundancy and regional coverage are infrastructure considerations, while interpretability refers to model transparency-not to how easy a model is to consume.

Adding a negative instruction (negative prompt) inside the prompt explicitly tells the model what it must not include, such as personal data. This operates as a guardrail before generation begins. Simply lowering temperature, using chain-of-thought examples, or enlarging the token limit affects creativity or verbosity but does not directly block disallowed content.