AWS Certified AI Practitioner Practice Test (AIF-C01)

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.

Service control policies (SCPs) in AWS Organizations allow a central administrator to set account-wide guardrails. An SCP can include a condition that denies all API calls targeting Regions outside the EU, ensuring data cannot be stored or processed elsewhere. IAM inline policies apply only to individual principals, not across accounts. Amazon Macie discovers sensitive data but does not prevent resource creation. AWS CloudFormation StackSets automate deployments but cannot enforce geographic restrictions on their own.

Zero-shot prompting is used when a user supplies an instruction but no worked examples. The model must rely on its pre-training to perform the requested task. Few-shot prompting adds several labeled examples to guide the model, so it would not meet the requirement of omitting sample reviews. Chain-of-thought prompting asks the model to reveal its reasoning steps and still requires an instruction (and optionally examples). Retrieval-Augmented Generation is not a standalone prompting technique; it enriches prompts with retrieved documents. Therefore, zero-shot prompting is the correct choice for giving only the instruction and no examples.

In the shared responsibility model, AWS is always responsible for security "of" the cloud, which includes the physical protection of the infrastructure that runs AWS services. Safeguarding the data-center facilities-such as controlling building access and monitoring the hardware environment-falls entirely to AWS. By contrast, tasks like encrypting training data, defining IAM permissions, and rotating user access keys are customer responsibilities because they relate to how resources are configured and used in the cloud.

Under guidance from the U.S. Copyright Office, copyright can only subsist in works that include "human authorship." Material created solely by an artificial intelligence system, with no meaningful human creative input, is therefore not eligible for copyright registration. Because no copyright arises, the text enters the public domain immediately and can be freely copied by others. Providing a prompt or owning the infrastructure does not itself create authorship, so neither the startup nor AWS holds copyright. There is no waiting period-works that lack protection never acquire it-so a 70-year delay is irrelevant.

To know if the foundation model is helping achieve the stated business goal-longer user sessions-you must measure the KPI itself. An A/B test that compares average session duration for users who receive the new AI-generated summaries versus a control group directly ties model performance to the engagement metric the business cares about. ROUGE scores and BLEU scores evaluate linguistic quality against references but may not correlate with user engagement. Inference latency measures system performance, while parameter count says nothing about the business impact. Neither of those metrics reveals whether the summaries keep users in the app longer.

Forecasting future electricity demand is a predictive problem that depends on patterns in historical, time-series data. Such problems benefit from ML algorithms that can learn complex relationships and improve accuracy over time. The other tasks follow fixed, explicit rules: overtime pay is calculated with a known formula, image compression relies on deterministic encoding algorithms, and enforcing a password-length requirement is a simple conditional check. Because these tasks do not involve pattern discovery or prediction, traditional programming is more efficient than ML.

Systematically changing a prompt, observing the model's response, and repeating the process is an example of iterative experimentation and testing. This cycle of modify-test-measure helps teams converge on a prompt that yields consistent, high-quality answers. The other options describe separate techniques-negative prompting constrains content, zero-shot prompting supplies no examples, and stop sequences control response length-but none rely on a repeated test-and-refine loop.

Simple, transparent models such as linear or logistic regression expose how each input contributes to the prediction, making them easy for auditors to understand. However, these models usually cannot capture complex patterns as well as deep neural networks, so accuracy can be lower. Deep neural networks often achieve higher predictive performance but are considered "black-box" and provide little built-in interpretability. The other options describe consequences that are not generally linked to transparency (such as needing more data, slower inference, or increased interpretability with deep models).

Guardrails for Amazon Bedrock let builders define policies that detect and filter inappropriate, hateful, or violent content before it reaches end users. Provisioned throughput reserves model capacity but does not moderate content. Agents for Bedrock add orchestration logic, not safety controls. Amazon CloudWatch logs can record events but cannot block unsafe text in real time.

Encrypting the dataset with AWS Key Management Service (KMS) keys protects it at rest, while enforcing TLS encryption secures data in transit. This approach directly addresses privacy and security requirements. Publishing a model card improves transparency, subgroup analysis addresses bias and fairness, and tuning by F1 score optimizes performance-none of these practices safeguard sensitive data.

Retrieval-Augmented Generation first looks up the most relevant passages from an external knowledge source-such as an indexed collection of company documents-at inference time. It then injects those retrieved snippets into the prompt so the foundation model can generate an informed answer using that fresh context. Because the knowledge is supplied dynamically, no additional model training, pruning, or model chaining is required. The other options describe full fine-tuning, parameter pruning for cost reduction, or dividing work across several models; none of those approaches capture the defining characteristic of RAG, which is real-time retrieval of external information to augment the model's response.

Amazon SageMaker Model Cards act as a single source of truth for machine-learning models. They capture key governance details such as where the training data came from, how the model will be used, performance and bias metrics, and any known limitations or risks. SageMaker Pipelines orchestrate ML workflows but do not produce governance reports. SageMaker Clarify analyzes bias and explainability results but does not create a consolidated document. SageMaker Feature Store stores features for training and inference, not governance metadata. Therefore, Model Cards are the correct choice.

The scenario focuses on converting text from one language to multiple other languages while maintaining the original intent and style. This is the definition of the translation use case in generative AI. Summarization shortens content, code generation produces programming instructions, and image generation creates visual media; none of these directly address language conversion.

Negative prompting adds explicit constraints such as "Do not include code or URLs" to the instruction portion of the prompt. The model then treats these constraints as requirements and avoids producing the specified content. Adjusting temperature, length settings, or stop sequences can influence style and termination but do not directly tell the model what content to avoid, so they are less reliable for blocking code or links.

The total number of parameters (model size) directly influences how much GPU or CPU memory and compute the model requires during inference. A model with fewer parameters has a smaller memory footprint and lower computational demand, making it more suitable for resource-constrained instances. Temperature, top-K sampling, and supported stop-sequence syntax affect response style or control but have little impact on the base hardware resources needed to load the model.

Amazon Bedrock is a fully managed service that exposes foundation models from AWS and leading model providers through a simple API. Developers can experiment or build applications without setting up servers, GPUs, or ML pipelines, giving a very low entry barrier. Amazon SageMaker JumpStart, AWS Lambda, and an EC2 Deep Learning AMI each require more infrastructure knowledge or management effort than Bedrock.

Amazon Bedrock is a fully managed service that gives developers API access to foundation models from multiple providers such as Anthropic, AI21 Labs, and Stability AI. Because the service is serverless, teams do not need to provision, scale, or maintain GPU infrastructure; they simply call the model endpoints and pay per request. Amazon SageMaker Studio supplies an integrated development environment but still requires users to manage underlying compute resources. Amazon Redshift is a data-warehousing service, and AWS Glue is used for extract-transform-load (ETL) workloads, so neither addresses the need for managed access to generative AI models.

Veracity in responsible AI focuses on the truthfulness, accuracy, and reliability of the data used to train a model and of the model's resulting outputs. Ensuring veracity involves validating, cleaning, and verifying data so that predictions are based on sound information. Robustness (consistent performance under varied or adversarial conditions), explainability (clarity of decision-making), and privacy (protection of personal data) are separate features that address different concerns.

A model is the mathematical representation produced by training algorithms on data. Once training finishes, this artifact captures learned patterns (for example, weights in a neural network) so it can accept new input and generate predictions or inferences. The algorithm itself, the raw training data, and the compute resources used are not the model; they are components of the pipeline that leads to creating the model.