Sign in
Home/AWS/AWS Certified AI Practitioner (AIF-C01)/Free questions

AWS Certified AI Practitioner (AIF-C01) — Free Practice Questions

10 free sample questions from a bank of 154, with the correct answers and explanations. No signup required — start practising right now.

1A company makes forecasts each quarter to decide how to optimize operations to meet expected demand. The company uses ML models to make these forecasts. An AI practitioner is writing a report about the trained ML models to provide transparency and explainability to company stakeholders. What should the AI practitioner include in the report to meet the transparency and explainability requirements?
  • Code for model training
  • Partial dependence plots (PDPs)
  • Sample data for training
  • Model convergence tables
Answer: B

The short version

B — Partial dependence plots show how features drive predictions. They visualize marginal effects so stakeholders can see why a forecast model behaves as it does.

Key concepts in this question

  • Transparency: stakeholders need understandable insight into model inputs, behavior and limits.
  • Explainability: tools must link feature values to output changes in an auditable way.
  • Partial dependence plots: charts showing average predicted outcome as one feature varies, holding others steady.
  • Forecast governance: operations planning from ML forecasts requires trust and documented reasoning.
  • Artifact choice: visual behavior evidence beats raw code or data dumps for business audiences.

Why B is correct

PDPs directly address transparency and explainability by revealing how key drivers such as demand signals affect the forecast. They let the practitioner include intuitive charts that connect input changes to output changes without exposing full internals. Raw code, sample rows or convergence tables do not explain behavior to stakeholders. PDPs therefore meet the requirement to provide accessible, decision-relevant transparency in the report.

Why the others are wrong

  • A. Training code shows implementation detail, but does not explain model behavior to stakeholders.
  • C. Sample training data shows inputs only, without linking features to predictions.
  • D. Convergence tables show optimization progress, not human-understandable decision logic.

AIF-C01 exam tip

For transparency reports, choose visual behavior tools like PDPs over code or raw data.

2A law firm wants to build an AI application by using large language models (LLMs). The application will read legal documents and extract key points from the documents. Which solution meets these requirements?
  • Build an automatic named entity recognition system.
  • Create a recommendation engine.
  • Develop a summarization chatbot.
  • Develop a multi-language translation system.
Answer: C

The short version

C — Extracting key points from legal documents is summarization. A summarization chatbot built on LLMs condenses long text into salient obligations, dates and arguments.

Key concepts in this question

  • Large language models: pretrained on vast text, strong at comprehension, abstraction and fluent rewriting.
  • Summarization: produces concise outputs preserving essential meaning from lengthy source documents.
  • Legal use case: contracts and filings are long and dense, users need rapid issue-spotting and review.
  • Chatbot interface: lets lawyers ask follow-ups, request shorter versions or focus on specific clauses.
  • Solution fit: extraction of key points matches summarization, not recommendation or translation.

Why C is correct

LLMs excel at reading long documents and generating faithful condensed versions that surface key points. A summarization chatbot therefore meets the need to ingest legal files and return digestible highlights with interactive refinement. Named entity systems only tag spans, recommendation engines suggest items and translation systems change languages. Only summarization delivers the requested comprehension plus condensation for legal review.

Why the others are wrong

  • A. Named entity recognition tags people and dates, but does not produce key-point summaries.
  • B. A recommendation engine suggests relevant items, it does not distill document meaning.
  • D. Translation converts between languages, it does not extract key legal points.

AIF-C01 exam tip

When the task is read long text and extract key points, choose summarization.

3A company wants to classify human genes into 20 categories based on gene characteristics. The company needs an ML algorithm to document how the inner mechanism of the model affects the output. Which ML algorithm meets these requirements?
  • Decision trees
  • Linear regression
  • Logistic regression
  • Neural networks
Answer: A

The short version

A — Decision trees are inherently interpretable classifiers. Their split-by-split logic can be traced and documented, showing exactly how gene traits lead to categories.

Key concepts in this question

  • Interpretability: ability to document inner mechanism and link inputs to outputs for review.
  • Decision trees: flowchart models that split on feature thresholds, paths are readable rules.
  • Multiclass classification: assigning genes to twenty categories needs a model supporting many classes.
  • Linear and logistic limits: linear regression predicts continuous values, logistic is less transparent for many classes.
  • Neural opacity: deep networks achieve accuracy but hide reasoning in distributed weights.

Why A is correct

Decision trees meet both needs at once, handling multiclass gene classification while remaining explainable through explicit split rules. Each branch names the gene characteristic, threshold and resulting group, so the team can document how internal logic affects output. Linear regression suits continuous prediction rather than categories, logistic adds complexity across many classes and neural networks lack clear traceability. Trees therefore best satisfy interpretable classification.

Why the others are wrong

  • B. Linear regression predicts numeric outcomes, not twenty discrete gene categories.
  • C. Logistic regression is built for binary or extended multiclass use, with less direct rule documentation.
  • D. Neural networks act as black boxes, making inner-mechanism documentation difficult.

AIF-C01 exam tip

For document-how-it-decides plus classification, pick decision trees first.

4A company has built an image classification model to predict plant diseases from photos of plant leaves. The company wants to evaluate how many images the model classified correctly. Which evaluation metric should the company use to measure the model's performance?
  • R-squared score
  • Accuracy
  • Root mean squared error (RMSE)
  • Learning rate
Answer: B

The short version

B — Accuracy measures share of correct classifications. For plant disease image labels, it directly answers how many photos were classified correctly.

Key concepts in this question

  • Accuracy: correct predictions divided by total predictions, intuitive for balanced classification.
  • Image classification: predicts discrete disease labels from leaf photos, success means exact label match.
  • Regression metrics: R-squared and RMSE assess continuous error, not categorical correctness.
  • Training controls: learning rate guides optimization steps, it does not evaluate final quality.
  • Metric alignment: choose the metric whose definition matches the business question asked.

Why B is correct

The company asks how many images were classified correctly, which is precisely the definition of accuracy. It counts correct disease labels over all evaluated photos and reports model performance in understandable terms. R-squared explains variance for regression, RMSE measures numeric deviation and learning rate is a hyperparameter. Only accuracy evaluates categorical correctness for this image classifier.

Why the others are wrong

  • A. R-squared evaluates variance explained in regression, not correct image labels.
  • C. RMSE measures average numeric error magnitude, not classification correctness.
  • D. Learning rate controls training step size, it is not a performance evaluation metric.

AIF-C01 exam tip

How many classified correctly always points to accuracy, regression errors point elsewhere.

5A company is using a pre-trained large language model (LLM) to build a chatbot for product recommendations. The company needs the LLM outputs to be short and written in a specific language. Which solution will align the LLM response quality with the company's expectations?
  • Adjust the prompt.
  • Choose an LLM of a different size.
  • Increase the temperature.
  • Increase the Top K value.
Answer: A

The short version

A — Prompt adjustment controls length, language and style. Clear instructions for brevity and target language align LLM outputs without retraining or resampling tweaks.

Key concepts in this question

  • Prompt engineering: instructions, context and constraints that steer tone, format and language.
  • Response quality: short outputs in a specific language are explicit output requirements.
  • Temperature and Top K: sampling controls affecting creativity and diversity, not reliable language control.
  • Model size: larger or smaller models change capability and cost, not precise formatting.
  • Cheapest fix first: prompt changes are fast and reversible compared with model swaps.

Why A is correct

Adjusting the prompt lets the team state be concise and answer in the required language, with examples if needed. That directly enforces short length and specific language for product recommendations. Raising temperature or Top K increases randomness and often lengthens or diversifies output, working against brevity. Choosing a different size does not guarantee language or length compliance. Prompt adjustment is therefore the targeted solution.

Why the others are wrong

  • B. Changing model size alters capacity and latency, but does not enforce short specific-language replies.
  • C. Higher temperature adds randomness and verbosity, opposite of short controlled answers.
  • D. Higher Top K broadens token choices, increasing variation rather than ensuring brevity.

AIF-C01 exam tip

For short plus specific language or format, fix the prompt before touching sampling.

6A company uses Amazon SageMaker for its ML pipeline in a production environment. The company has large input data sizes up to 1 GB and processing times up to 1 hour. The company needs near real-time latency. Which SageMaker inference option meets these requirements?
  • Real-time inference
  • Serverless inference
  • Asynchronous inference
  • Batch transform
Answer: C

The short version

C — Asynchronous inference fits large payloads and long processing with near real-time needs. It queues requests, scales compute and returns results without blocking the client.

Key concepts in this question

  • SageMaker real-time inference: persistent endpoint for low-latency small payloads, limited payload size and timeout.
  • Serverless inference: auto-scaled ephemeral option for intermittent traffic, still bound by payload and duration caps.
  • Asynchronous inference: queues up to 1 GB inputs, supports up to one hour processing, ideal for near real-time.
  • Batch transform: offline bulk processing for large datasets, not designed for request-response latency.
  • Requirement match: 1 GB plus one hour plus near real-time removes all but async.

Why C is correct

The workload specifies input sizes up to 1 GB and processing up to one hour with near real-time latency. Asynchronous inference is purpose-built for exactly those limits, accepting large payloads, running long jobs and notifying on completion while scaling queues. Real-time and serverless cannot accept that size and duration, while batch transform targets offline jobs without near real-time response. Async therefore uniquely meets all three constraints together.

Why the others are wrong

  • A. Real-time inference needs persistent low-latency calls, but rejects 1 GB payloads and hour-long runs.
  • B. Serverless inference suits sporadic small requests, not gigabyte inputs with hour processing.
  • D. Batch transform handles bulk offline work, not near real-time request workflows.

AIF-C01 exam tip

Memorize async as 1 GB plus one hour plus near real-time, batch as offline bulk.

7A company is using domain-specific models. The company wants to avoid creating new models from the beginning. The company instead wants to adapt pre-trained models to create models for new, related tasks. Which ML strategy meets these requirements?
  • Increase the number of epochs.
  • Use transfer learning.
  • Decrease the number of epochs.
  • Use unsupervised learning.
Answer: B

The short version

B — Transfer learning reuses pretrained models for related tasks. It adapts learned features to new domains instead of training from scratch.

Key concepts in this question

  • Pretrained models: capture general patterns from large domain data that transfer to nearby problems.
  • Transfer learning: freezes or fine-tunes base layers, then trains task heads on smaller new datasets.
  • Efficiency gain: saves data, compute and time versus building fresh models from random initialization.
  • Epoch tweaks: changing epoch counts adjusts training length, not knowledge reuse strategy.
  • Learning paradigms: unsupervised learning finds structure without labels, it does not adapt a prior model.

Why B is correct

The company wants to avoid creating new models from the beginning and instead adapt pretrained models for new related tasks. That is the textbook definition of transfer learning, where domain knowledge carries over and only final layers need retraining. Increasing or decreasing epochs merely changes iteration count for the same model. Unsupervised learning describes unlabeled training, not reuse of an existing model. Transfer learning is therefore the required ML strategy.

Why the others are wrong

  • A. More epochs extends training of the current model, it does not reuse a pretrained base.
  • C. Fewer epochs shortens training, it does not adapt existing models to new tasks.
  • D. Unsupervised learning clusters unlabeled data, it does not fine-tune pretrained models.

AIF-C01 exam tip

Reuse pretrained for related tasks always signals transfer learning and fine-tuning.

8A company is building a solution to generate images for protective eyewear. The solution must have high accuracy and must minimize the risk of incorrect annotations. Which solution will meet these requirements?
  • Human-in-the-loop validation by using Amazon SageMaker Ground Truth Plus
  • Data augmentation by using an Amazon Bedrock knowledge base
  • Image recognition by using Amazon Rekognition
  • Data summarization by using Amazon QuickSight Q
Answer: A

The short version

A — Ground Truth Plus adds expert human review for accuracy. Human-in-the-loop validation catches bad annotations before generated eyewear images cause safety risk.

Key concepts in this question

  • High accuracy need: protective eyewear imagery demands reliable labels, errors carry safety consequences.
  • Human-in-the-loop: people review, correct and approve machine outputs where precision matters.
  • Ground Truth Plus: managed workforce service delivering high-quality labeling and validation workflows.
  • Generation risk: synthetic images can look plausible while containing incorrect annotations or defects.
  • Alternative mismatch: augmentation, recognition and dashboards do not assure annotation correctness.

Why A is correct

Minimizing incorrect annotations while demanding high accuracy calls for human validation of labels and generated images. SageMaker Ground Truth Plus provides expert reviewers plus workflows built for quality assurance at scale. Data augmentation expands datasets, Rekognition detects objects and QuickSight Q summarizes data, but none enforce managed human approval. Only human-in-the-loop validation through Ground Truth Plus meets both accuracy and risk requirements.

Why the others are wrong

  • B. Bedrock knowledge base augmentation improves retrieval context, it does not validate annotations.
  • C. Rekognition performs image recognition, it does not provide human approval for labels.
  • D. QuickSight Q enables natural language analytics, it does not assure image annotation quality.

AIF-C01 exam tip

For minimize annotation errors plus high accuracy, choose human review with Ground Truth.

9A company wants to create a chatbot by using a foundation model (FM) on Amazon Bedrock. The FM needs to access encrypted data that is stored in an Amazon S3 bucket. The data is encrypted with Amazon S3 managed keys (SSE-S3). The FM encounters a failure when attempting to access the S3 bucket data. Which solution will meet these requirements?
  • Ensure that the role that Amazon Bedrock assumes has permission to decrypt data with the correct encryption key.
  • Set the access permissions for the S3 buckets to allow public access to enable access over the internet.
  • Use prompt engineering techniques to tell the model to look for information in Amazon S3.
  • Ensure that the S3 data does not contain sensitive information.
Answer: A

The short version

A — Bedrock needs an IAM role allowed to decrypt the SSE-S3 object. Without decrypt permission on the key, knowledge base ingestion and retrieval calls fail even when the bucket exists.

Key concepts in this question

  • SSE-S3 encryption: S3 encrypts objects with managed keys, reads require both S3 access and decrypt rights.
  • Bedrock execution role: the service assumes a customer role to access data sources such as S3.
  • Failure pattern: access denied on encrypted objects usually means missing kms decrypt, not missing data.
  • Least privilege fix: grant precise decrypt on the key rather than opening broad access.
  • Security anti-pattern: public buckets or prompt tricks do not resolve cryptographic permission errors.

Why A is correct

When an FM on Bedrock cannot read SSE-S3 data, the assumed role lacks permission to decrypt with the correct encryption key. Granting decrypt rights lets Bedrock fetch and index the encrypted objects securely while preserving protection. Making buckets public creates severe exposure, prompt engineering cannot grant cryptographic access and checking for sensitive data avoids the actual permission fault. Fixing role decrypt permission is therefore the correct solution.

Why the others are wrong

  • B. Public S3 access would expose data dangerously and still mishandles key-based decryption.
  • C. Prompt wording cannot grant IAM or key permissions needed for encrypted reads.
  • D. Removing sensitive info dodges governance and does not fix the decrypt failure.

AIF-C01 exam tip

Encrypted S3 plus Bedrock failure means check the assumed role for decrypt rights.

10A company wants to use language models to create an application for inference on edge devices. The inference must have the lowest latency possible. Which solution will meet these requirements?
  • Deploy optimized small language models (SLMs) on edge devices.
  • Deploy optimized large language models (LLMs) on edge devices.
  • Incorporate a centralized small language model (SLM) API for asynchronous communication with edge devices.
  • Incorporate a centralized large language model (LLM) API for asynchronous communication with edge devices.
Answer: A

The short version

A — Optimized small models on edge give the lowest latency. Local SLMs avoid network hops and heavy compute, delivering fast inference where data is created.

Key concepts in this question

  • Edge inference: runs near devices to cut round trips, bandwidth use and dependency on connectivity.
  • Small language models: compact, distilled or quantized models tuned for constrained CPU and memory.
  • Latency drivers: model size, network distance and compute load dominate response time.
  • Centralized APIs: add queuing and transit delays, even when async, hurting lowest-latency goals.
  • Large models at edge: exceed device limits and increase compute time versus optimized small models.

Why A is correct

Lowest possible latency for language inference on edge devices requires both small optimized models and local execution. Deploying optimized SLMs directly on devices removes internet round trips and reduces per-token compute. Large models strain edge resources, while centralized SLM or LLM APIs reintroduce network and queuing delay. Only on-device optimized SLMs satisfy lowest latency plus edge deployment together.

Why the others are wrong

  • B. Large models need more memory and compute, increasing latency on constrained edge hardware.
  • C. Centralized SLM APIs still require remote calls, so latency exceeds local execution.
  • D. Centralized LLM APIs add the largest model plus network delay, worst for latency.

AIF-C01 exam tip

Lowest latency on edge always means small optimized models deployed locally.

Want the full bank of 154 questions for AWS Certified AI Practitioner (AIF-C01)? See all practice exams.