AWS AI Practitioner Practice Test.
66 free practice questions with answers and explanations.
No signup required. Choose a topic and review each answer.
Start practicing →About these practice questions
These are original study questions written from published exam objectives—not recalled, copied, or confidential live-exam items. Always confirm current coverage with the official sources linked on this page.
Exam format and study resources
The AWS AI Practitioner is administered by Amazon Web Services (AWS), with 65 scored questions, a 1 hour 30 minutes time limit and a 700/1000 result.
This free AWS AI Practitioner practice test has 66 original questions written to Amazon Web Services (AWS)'s official content outline, last checked against it on July 18, 2026. Every question shows a worked explanation, and nothing here requires a signup.
As of 2026, the AWS AI Practitioner exam fee is $100.
Browse all questions & answers
1. A retailer wants a system that can learn patterns from historical sales data without being given explicit rules, and then predict future demand. Which term best describes this general approach?
- A. Machine learning
- B. Robotic process automation
- C. Rule-based expert system
- D. Relational database indexing
Show answer & explanation
Answer: A
Machine learning is the field of building systems that learn statistical patterns from data to make predictions, rather than following hard-coded rules. RPA automates repetitive tasks via scripted rules, not learned patterns. A rule-based expert system relies on explicit if-then logic written by humans. Database indexing is a storage/retrieval optimization technique, unrelated to prediction from patterns.2. A data scientist has a labeled dataset of emails marked 'spam' or 'not spam' and wants to train a model to classify new emails. Which type of machine learning approach fits this scenario?
- A. Unsupervised learning
- B. Supervised learning
- C. Reinforcement learning
- D. Semi-structured clustering
Show answer & explanation
Answer: B
Supervised learning trains on labeled input-output pairs (email text mapped to a spam/not-spam label), which matches this scenario exactly. Unsupervised learning works with unlabeled data to find structure like clusters. Reinforcement learning learns via rewards from an agent interacting with an environment, not static labels. 'Semi-structured clustering' is not a standard ML paradigm.3. In a neural network, what is the primary role of the 'weights' associated with connections between neurons?
- A. They define the number of layers the network will have
- B. They store the raw unprocessed training data permanently
- C. They set the learning rate used by the optimizer
- D. They determine the strength/importance of a given input's contribution to a neuron's output, and are adjusted during training
Show answer & explanation
Answer: D
Weights scale the influence of each input on a neuron's activation and are iteratively updated during training (e.g., via backpropagation) to minimize error. They do not store training data, do not fix the network's layer count (that's architecture), and are distinct from the learning rate, which is a separate hyperparameter controlling how much weights change per update.4. A company is deciding between building a custom ML model from scratch versus using a pre-trained model via an API for a sentiment analysis feature with a tight deadline. What is the main tradeoff of using a pre-trained model via API?
- A. It guarantees higher accuracy than any custom model in all cases
- B. It eliminates the need for any input validation
- C. It speeds time-to-market but offers less control over model behavior and customization compared to training from scratch
- D. It requires the company to label millions of examples before first use
Show answer & explanation
Answer: C
Pre-trained models accessed via API let teams ship quickly without collecting/labeling data or training infrastructure, but the tradeoff is reduced control over the model's internals, biases, and customization versus a model trained in-house. No pre-trained model 'guarantees' superiority in all cases, using an API doesn't remove the need to validate inputs, and using a pre-trained model specifically avoids needing to label large training sets, the opposite of option D.5. A model performs very well on training data but poorly on new, unseen validation data. Which concept best explains this behavior?
- A. Underfitting
- B. Overfitting
- C. Data augmentation
- D. Feature engineering
Show answer & explanation
Answer: B
Overfitting occurs when a model memorizes noise and specific patterns in training data rather than learning generalizable relationships, causing strong training performance but weak generalization to new data. Underfitting is the opposite problem, poor performance on both training and new data because the model is too simple. Data augmentation and feature engineering are techniques used to improve models, not descriptions of this failure mode.6. Which metric would be most appropriate to evaluate a binary classification model when the two classes are heavily imbalanced (e.g., 1% fraud vs 99% legitimate transactions)?
- A. Raw accuracy alone
- B. Mean squared error
- C. R-squared
- D. Precision, recall, and F1 score
Show answer & explanation
Answer: D
With heavy class imbalance, a model can achieve high raw accuracy by simply always predicting the majority class, so precision, recall, and F1 give a more meaningful view of performance on the minority (fraud) class. Mean squared error and R-squared are regression metrics used for continuous-value prediction, not classification.7. What distinguishes a foundation model from a traditional, narrowly-trained ML model built for one specific task?
- A. Foundation models are trained on broad, large-scale data and can be adapted to many downstream tasks
- B. Foundation models can only be used for image classification
- C. Foundation models are always smaller and faster to train
- D. Foundation models never require any fine-tuning or prompting
Show answer & explanation
Answer: A
Foundation models are pretrained on massive, diverse datasets and are designed to be adapted (via prompting, fine-tuning, or other techniques) to a wide variety of downstream tasks, unlike traditional models built and trained narrowly for a single task. They are typically large, not small, are not limited to image classification, and usually still benefit from prompting or adaptation to perform well on specific tasks.8. When a large language model generates text one token at a time, each new token being conditioned on the previously generated tokens, what is this generation process commonly called?
- A. Autoregressive generation
- B. Federated learning
- C. Dimensionality reduction
- D. Batch normalization
Show answer & explanation
Answer: A
Autoregressive generation describes producing output sequentially, where each new token is predicted based on the tokens generated so far, which is how most LLMs generate text. Batch normalization is a training-stabilization technique for neural network layers. Federated learning is a distributed training approach across decentralized data. Dimensionality reduction reduces the number of features in data, unrelated to sequential text generation.9. What is the primary purpose of 'tokens' in the context of large language models?
- A. Tokens are the discrete units (words, subwords, or characters) that text is broken into for the model to process
- B. Tokens are the labels assigned to training data by human annotators
- C. Tokens are a measure of GPU memory bandwidth
- D. Tokens are security credentials used to authenticate API calls
Show answer & explanation
Answer: A
In NLP, tokenization splits text into smaller units called tokens (which can be words, subwords, or characters) that a language model processes numerically. This is distinct from authentication tokens (a security concept), hardware bandwidth measures, or annotation labels, which are unrelated concepts that happen to share terminology in other contexts.10. A user sets a generative model's 'temperature' parameter very high (e.g., close to 1.5-2.0) instead of near 0. What effect should they expect on the model's output?
- A. Output becomes more deterministic and repetitive
- B. Output becomes more random and creative, with more variation between runs
- C. The model will refuse to generate any output
- D. The model's context window size increases
Show answer & explanation
Answer: B
Temperature controls the randomness of token sampling; higher temperature flattens the probability distribution over next tokens, producing more varied and creative but less predictable output, while low temperature makes output more deterministic and focused. Temperature does not cause refusals and has no effect on context window size, which is a separate architectural limit.11. A prompt engineer includes a few example input-output pairs directly in the prompt before asking the model to complete a new, similar task. What is this technique called?
- A. Zero-shot prompting
- B. Few-shot prompting
- C. Fine-tuning
- D. Model distillation
Show answer & explanation
Answer: B
Few-shot prompting provides a handful of example input-output pairs within the prompt itself to guide the model's response format and behavior without updating model weights. Zero-shot prompting gives no examples at all. Fine-tuning actually updates model weights using a training dataset, which is a different, more resource-intensive process. Model distillation compresses a large model into a smaller one, unrelated to prompting technique.12. An LLM confidently states an incorrect fact as though it were true, with no basis in its training data or provided context. What is this phenomenon commonly called?
- A. Hallucination
- B. Quantization
- C. Vectorization
- D. Regularization
Show answer & explanation
Answer: A
Hallucination refers to a generative model producing plausible-sounding but factually incorrect or fabricated content. Regularization is a training technique to reduce overfitting. Vectorization converts data (e.g., text) into numeric vector form. Quantization reduces numeric precision of model weights to shrink model size, none of which describe fabricated outputs.13. A company wants a generative model to produce outputs in a very specific tone and structure unique to their brand, and has thousands of labeled example documents to support this. Which approach is generally best suited to durably encode this behavior into the model itself?
- A. Fine-tuning the model on the labeled brand-specific examples
- B. Increasing the temperature parameter only
- C. Reducing the number of tokens allowed per request
- D. Switching to a smaller, less capable model
Show answer & explanation
Answer: A
Fine-tuning updates model weights using a labeled dataset so the desired tone/structure becomes part of the model's learned behavior, which is well suited to a durable, large-scale example set. Adjusting temperature only changes randomness, not learned style. Reducing token limits per request restricts response length, unrelated to style. Switching to a smaller model does not by itself encode brand-specific tone.14. What best differentiates prompt engineering from fine-tuning as ways to adapt a foundation model's output for a specific use case?
- A. Prompt engineering changes model weights while fine-tuning only changes the input text
- B. Prompt engineering shapes model behavior through crafted input at inference time without changing weights, while fine-tuning updates the model's weights via additional training
- C. They are functionally identical processes with different names
- D. Prompt engineering requires a labeled dataset while fine-tuning does not
Show answer & explanation
Answer: B
Prompt engineering influences output purely through how a request is worded or structured at inference time, leaving the underlying model weights untouched, whereas fine-tuning actually retrains and updates the model's parameters using additional data. Option A reverses the definitions. They are not the same process. Fine-tuning, not prompt engineering, is the one that typically requires a labeled dataset.15. Which scenario best illustrates an appropriate use case for retrieval-augmented generation (RAG) with a foundation model?
- A. Generating random creative fiction with no factual grounding requirements
- B. Reducing the number of GPUs needed to physically host the model
- C. Answering customer questions using up-to-date internal knowledge-base documents the model was never trained on
- D. Compressing a model's weights to run on a smaller device
Show answer & explanation
Answer: C
RAG retrieves relevant external documents at query time and supplies them as context to the model, which is ideal for grounding answers in current or proprietary information not present in the model's training data. Free-form fiction generation has no such grounding need. RAG does not reduce hosting hardware requirements or compress model weights, those relate to infrastructure sizing and quantization respectively.16. In a typical RAG architecture, what is the role of a vector database?
- A. It stores embeddings of documents and enables similarity search to retrieve relevant context for a query
- B. It replaces the foundation model entirely so no inference call is needed
- C. It is used exclusively to store user authentication credentials
- D. It converts text prompts into images for multimodal models
Show answer & explanation
Answer: A
A vector database stores numeric embeddings representing the semantic meaning of documents and supports similarity search, so the most relevant chunks can be retrieved and passed to the model as context, a core piece of RAG pipelines. It does not replace the foundation model's generation step, is not an authentication store, and does not perform text-to-image conversion.17. A company is building a generative AI chatbot and wants to reduce hallucinations when answering questions about its constantly-updated product catalog. Which architectural approach directly addresses this need without retraining the underlying model?
- A. Increase the model's temperature setting
- B. Implement retrieval-augmented generation using the current product catalog as a knowledge source
- C. Reduce the number of tokens allowed in each response
- D. Switch to a model with fewer parameters
Show answer & explanation
Answer: B
RAG grounds responses in retrieved, current product data at query time, directly reducing hallucination risk for frequently changing information without needing to retrain the model. Raising temperature increases randomness/creativity, which would worsen hallucination risk. Limiting response length and using a smaller model do not address factual grounding at all.18. A developer is building an application that needs a foundation model to call external functions (like checking inventory or booking a calendar event) based on user requests. Which capability enables this pattern?
- A. Tool use / function calling
- B. Model quantization
- C. Batch normalization
- D. Gradient clipping
Show answer & explanation
Answer: A
Tool use (function calling) lets a foundation model decide when to invoke external APIs or functions, passing structured arguments, so an application can trigger real actions like inventory checks based on model output. Quantization is a model-compression technique, batch normalization and gradient clipping are neural network training techniques, none of which relate to invoking external functions from a model.19. An agentic AI application needs to break a complex user goal into a sequence of steps, call tools, and adapt based on intermediate results. What best describes this capability compared to a simple single-turn prompt-response chatbot?
- A. An agent requires no foundation model at all
- B. An agent can plan, take multi-step actions using tools, and adjust based on feedback, unlike a single-turn chatbot that only returns one response per prompt
- C. An agent is simply a chatbot with a larger context window and nothing else different
- D. An agent can only be used for image generation tasks
Show answer & explanation
Answer: B
Agentic AI systems use a foundation model as a reasoning engine to plan multi-step actions, invoke tools, observe results, and adjust course, which goes beyond a single prompt-in/response-out chatbot interaction. It is not just a larger context window, it still relies on a foundation model for reasoning, and it is not limited to image generation.20. A company wants to evaluate whether a foundation model's output for a summarization task is high quality when there is no single 'correct' reference answer. Which evaluation approach is most appropriate?
- A. Exact string match against one reference summary
- B. Human evaluation or an LLM-as-judge approach scoring criteria like coherence, relevance, and factual consistency
- C. Measuring only the response latency in milliseconds
- D. Counting the number of tokens generated
Show answer & explanation
Answer: B
For open-ended generation tasks like summarization, quality is subjective and multi-dimensional, so human evaluation or LLM-as-judge scoring against criteria such as coherence and factual consistency is far more meaningful than exact string matching, which fails when many valid summaries exist. Latency and token count measure performance/cost, not output quality.21. Why might prompt-based approaches sometimes be preferred over fine-tuning when adapting a foundation model to a new task with limited labeled data and a tight budget?
- A. Prompting always produces strictly better accuracy than fine-tuning in every case
- B. Prompting permanently changes the model's weights for all future users
- C. Prompting requires no labeled examples and no additional training compute, making it faster and cheaper to iterate
- D. Prompting requires access to the model's full training dataset
Show answer & explanation
Answer: C
Prompt-based adaptation (zero-shot or few-shot) needs no additional labeled training set or model retraining, so it is much faster and cheaper to experiment with when data or budget is limited. It does not universally outperform fine-tuning (which can outperform prompting for well-defined tasks with enough data), does not alter model weights at all, and requires no access to the original training dataset.22. A company deploying a generative AI hiring-assistant tool is concerned the model might produce biased recommendations that disadvantage certain demographic groups. Which responsible AI principle does this concern most directly relate to?
- A. Fairness
- B. Latency optimization
- C. Model compression
- D. Autoregressive decoding
Show answer & explanation
Answer: A
Fairness is the responsible-AI principle concerned with ensuring a model does not produce systematically biased or discriminatory outcomes across demographic groups, which is exactly the hiring-bias risk described. Latency optimization and model compression are performance/engineering concerns unrelated to bias. Autoregressive decoding is a generation mechanism, not a fairness consideration.23. What does 'explainability' mean in the context of responsible AI for a model used to approve or deny loan applications?
- A. The ability to describe, in understandable terms, why the model produced a particular decision or output
- B. The raw processing speed of the model measured in tokens per second
- C. The number of parameters contained in the model
- D. The total dollar cost of running the model in production
Show answer & explanation
Answer: A
Explainability refers to the capacity to articulate, in human-understandable terms, the reasoning or factors behind a model's decision, which is critical for high-stakes use cases like loan approvals where affected individuals and regulators may need justification. Processing speed, cost, and parameter count are performance/scale metrics unrelated to interpretability of decisions.24. A healthcare organization wants to deploy a generative AI assistant that summarizes patient records for clinicians. Which responsible AI consideration is especially critical given the potential real-world impact of an incorrect summary?
- A. Minimizing hallucinations and ensuring outputs are verifiable/traceable to source records, given the high stakes of medical decisions
- B. Maximizing the model's temperature to produce more creative summaries
- C. Reducing the number of clinicians who review the output
- D. Avoiding any logging of model outputs for audit purposes
Show answer & explanation
Answer: A
In high-stakes domains like healthcare, minimizing hallucinated or unverifiable content and ensuring traceability to source data is critical because incorrect summaries can lead to patient harm. Increasing creativity via higher temperature would increase, not reduce, hallucination risk. Reducing human review and avoiding audit logging both increase risk rather than support responsible deployment.25. An organization publishes documentation describing a model's intended use cases, known limitations, training data characteristics, and evaluation results. What responsible AI practice does this best represent?
- A. Gradient descent optimization
- B. Load balancing across inference endpoints
- C. Data encryption at rest
- D. Transparency through model documentation (e.g., model cards)
Show answer & explanation
Answer: D
Publishing documentation of intended use, limitations, training data, and evaluation results is a transparency practice, often implemented as 'model cards,' that helps downstream users understand appropriate and inappropriate uses of a model. Data encryption and load balancing are security/infrastructure practices, and gradient descent is a training optimization algorithm, none of which describe documentation practices.26. A generative AI content tool occasionally produces outputs that could be considered toxic or offensive. Which mitigation is most directly aimed at reducing this specific risk before content reaches end users?
- A. Implementing content filtering/guardrails to detect and block harmful outputs
- B. Increasing the model's context window size
- C. Switching the model to a lower-cost inference tier
- D. Reducing the number of unscored evaluation questions used in testing
Show answer & explanation
Answer: A
Content filtering and guardrail systems are specifically designed to detect and block toxic, offensive, or otherwise harmful generated content before it reaches users, directly targeting this risk. Context window size affects how much input the model can process, not content safety. Switching to a cheaper inference tier is a cost decision, and evaluation question counts relate to testing methodology, neither mitigates toxic output.27. A company handling regulated financial data wants to ensure its generative AI application only allows authorized users to invoke certain sensitive tools/functions exposed to the model. Which security concept applies here?
- A. Access control / authorization enforcement
- B. Model quantization
- C. Data augmentation
- D. Gradient clipping
Show answer & explanation
Answer: A
Access control (authorization) governs which authenticated users or roles are permitted to perform specific actions, such as invoking sensitive tools connected to an AI application, which is exactly the scenario described. Quantization is a model-compression technique, data augmentation expands training datasets, and gradient clipping stabilizes training, none of which relate to restricting who can invoke a function.28. A malicious user crafts input text specifically designed to trick a generative AI application into ignoring its system instructions and revealing restricted information. What is this attack technique called?
- A. Prompt injection
- B. Model distillation
- C. Gradient descent
- D. Data normalization
Show answer & explanation
Answer: A
Prompt injection is an attack where crafted input attempts to override or bypass a model's system instructions or safety constraints, causing it to behave in unintended ways such as leaking restricted data. Model distillation compresses a model into a smaller one, gradient descent is a training optimization method, and data normalization is a data preprocessing step, none of which describe an adversarial manipulation of prompts.29. An organization must ensure that customer data used to fine-tune a generative AI model is not inadvertently exposed in the model's outputs to other customers. Which practice most directly addresses this concern?
- A. Reducing the number of tokens in the context window
- B. Choosing a model with more parameters
- C. Increasing the model's temperature setting
- D. Data governance controls such as data isolation, access boundaries, and testing for training-data leakage
Show answer & explanation
Answer: D
Data governance practices, including isolating tenant data, enforcing access boundaries, and testing whether fine-tuned models leak memorized training examples, directly address the risk of one customer's data surfacing in another's outputs. Temperature controls randomness, context window size limits input length, and parameter count relates to model capacity, none of which prevent data leakage across customers.30. A company operating in a regulated industry needs to demonstrate to auditors that its AI system's decisions can be traced and reviewed after the fact. Which capability is most essential to support this requirement?
- A. Logging and audit trails capturing inputs, outputs, and decision context over time
- B. Maximizing model temperature for creative responses
- C. Minimizing the size of the training dataset
- D. Disabling all human review of AI outputs
Show answer & explanation
Answer: A
Comprehensive logging and audit trails that capture inputs, outputs, and relevant context allow an organization to reconstruct and review AI-driven decisions after the fact, which is essential for regulatory audits. Higher temperature affects creativity/randomness, minimizing training data size would likely hurt model quality, and disabling human review removes oversight rather than supporting auditability.31. Why is data encryption both at rest and in transit considered a foundational security control for an AI application handling sensitive user data?
- A. It eliminates the need for access control policies
- B. It automatically improves the model's prediction accuracy
- C. It is only relevant for image-based, not text-based, AI models
- D. It protects data from being read by unauthorized parties whether it is stored on disk or moving across a network
Show answer & explanation
Answer: D
Encryption at rest protects stored data from unauthorized access if storage is compromised, and encryption in transit protects data moving across networks from interception, together forming a baseline security control regardless of data type or model modality. Encryption has no effect on model accuracy, does not replace the separate need for access control, and applies equally to text, image, or any other data type.32. A company wants to ensure its generative AI vendor contract and internal processes clearly define who is responsible for reviewing outputs, monitoring for misuse, and responding to incidents. Which broader concept does this best represent?
- A. AI governance
- B. Tokenization
- C. Backpropagation
- D. Embedding generation
Show answer & explanation
Answer: A
AI governance encompasses the policies, roles, and processes an organization puts in place to oversee responsible development and use of AI systems, including defining accountability for review, monitoring, and incident response. Tokenization and embedding generation are technical NLP/data-representation steps, and backpropagation is a neural network training algorithm, none of which relate to organizational accountability structures.33. A model trained to predict housing prices performs well in the region it was trained on but produces systematically wrong predictions in a new region with different market dynamics. What underlying issue does this best illustrate?
- A. The model generalizes poorly outside the distribution of its training data (a form of distribution shift)
- B. The model has too few parameters to store any data at all
- C. The model is exhibiting hallucination, a generative-AI-specific failure mode
- D. The model's tokenizer is malfunctioning
Show answer & explanation
Answer: A
When real-world data characteristics differ from the training distribution (distribution shift), a model's learned patterns may no longer apply, causing systematically poor predictions in the new context, exactly as described for a new housing market. Parameter count is unrelated to this issue. Hallucination is a term specific to generative models fabricating content, not a regression model's prediction error, and tokenizers are a text-preprocessing component, not applicable to a numeric housing-price prediction failure.34. A team building a multimodal generative AI feature wants the model to accept both an image and a text question about that image, then produce a text answer. What best describes the type of model needed?
- A. A multimodal foundation model capable of processing more than one data modality (e.g., text and image) as input
- B. A traditional relational database with SQL joins
- C. A single-modality text-only model with no image support whatsoever
- D. A rule-based decision tree with hand-coded if-else logic
Show answer & explanation
Answer: A
Multimodal foundation models are specifically designed to accept and reason across multiple data types, such as combining an image and a text question to produce a text answer, which matches the described use case. A relational database performs structured data queries, not visual question answering. A text-only model cannot process image input at all, and a hand-coded decision tree lacks the learned representations needed to interpret images and free-form text jointly.35. An organization is comparing the cost structure of using an on-demand, pay-per-request foundation model API versus provisioning dedicated capacity for guaranteed throughput. In which scenario would provisioned/dedicated capacity likely be more cost-effective?
- A. A workload with consistently high, predictable request volume where guaranteed throughput matters
- B. A prototype used a handful of times during initial experimentation
- C. An application that is rarely used and has highly unpredictable, sporadic traffic
- D. A one-time internal demo run a single time for stakeholders
Show answer & explanation
Answer: A
Provisioned or dedicated capacity typically involves a fixed cost in exchange for guaranteed throughput and often lower per-unit cost at high, steady volumes, making it more economical for consistently heavy workloads. For prototypes, rare/unpredictable usage, or one-off demos, low-commitment pay-per-request pricing is generally more cost-effective since there is no sustained volume to justify a fixed capacity commitment.36. A team must choose between supervised and unsupervised learning for a task. What determines the choice?
- A. Whether labelled examples of the desired output exist, since supervised learning requires them and unsupervised learning finds structure without them
- B. The volume of data available, regardless of labelling
- C. The programming language used to implement the model
- D. Whether the data is stored in a relational database
Show answer & explanation
Answer: A
Labels define the target the model learns to predict, so their presence or absence determines which family of algorithms applies. Clustering, dimensionality reduction and anomaly detection are the common unsupervised tasks, and labelling cost is frequently the binding constraint on whether a supervised approach is feasible at all.37. A model performs excellently on training data but poorly on unseen data. What has occurred?
- A. Overfitting, where the model learned noise and specifics of the training set rather than generalizable patterns
- B. Underfitting, where the model is too simple to capture the pattern
- C. Data leakage from the test set into training
- D. Insufficient computational resources during training
Show answer & explanation
Answer: A
The gap between training and validation performance is the signature of overfitting, addressed by more data, regularization, simpler models or early stopping. Underfitting shows poor performance on both sets, and data leakage produces the opposite symptom of unrealistically good validation results that collapse in production.38. A dataset is split for model development. Why are separate validation and test sets used rather than one holdout set?
- A. Because repeated tuning against the validation set leaks information into the model, so an untouched test set is needed for an honest final estimate
- B. Because two sets provide twice the training data
- C. Because the test set is used to train the final model
- D. Because validation sets must be larger than training sets
Show answer & explanation
Answer: A
Each tuning decision made by looking at validation performance fits the model a little to that set, so validation scores become optimistic after many iterations. The test set is held back and used once, which is what makes its estimate representative of unseen data.39. A classifier for a rare disease achieves 99 percent accuracy on a dataset where 1 percent of cases are positive. Why is accuracy misleading here?
- A. Predicting the negative class every time would also achieve 99 percent, so precision, recall and their balance are the informative metrics
- B. Accuracy cannot be calculated for medical data
- C. 99 percent accuracy is too low to be meaningful
- D. Accuracy only applies to regression problems
Show answer & explanation
Answer: A
With severe class imbalance the majority-class baseline already scores highly, so accuracy carries almost no information about performance on the class that matters. Recall measures how many actual positives were found and precision how many predicted positives were correct, and the relative cost of a missed case versus a false alarm determines which to favour.40. A foundation model is described as pre-trained. What does this mean for a team adopting it?
- A. It has already learned general representations from large-scale data, so the team can adapt it to a task rather than training from scratch
- B. It requires no further configuration or evaluation for any task
- C. It can only be used for the exact task it was trained on
- D. It contains the team's own data by default
Show answer & explanation
Answer: A
Pre-training amortizes the largest cost, letting teams reach useful performance with prompting, retrieval augmentation or fine-tuning on comparatively small datasets. It does not remove the need for task-specific evaluation, since general capability does not guarantee adequacy on any particular task.41. A generative model produces a confident but factually incorrect statement. What is this called and what is a practical mitigation?
- A. A hallucination, mitigated by grounding responses in retrieved source documents and citing them so claims can be checked
- B. Overfitting, mitigated by adding more training epochs
- C. Data drift, mitigated by retraining on older data
- D. Underfitting, mitigated by simplifying the prompt
Show answer & explanation
Answer: A
Language models generate plausible continuations rather than retrieving verified facts, so fluency and correctness are independent properties. Retrieval-augmented generation supplies authoritative source text at inference and citation lets the reader verify, which converts an unverifiable assertion into a checkable one.42. A team must choose between prompt engineering, retrieval augmentation and fine-tuning for adapting a foundation model. What primarily distinguishes retrieval augmentation?
- A. It supplies external knowledge at inference time without changing model weights, so the knowledge can be updated without retraining
- B. It permanently encodes the organization's data into the model's weights
- C. It requires the largest labelled dataset of the three
- D. It removes the need to evaluate model outputs
Show answer & explanation
Answer: A
Retrieval keeps knowledge in an external store the model consults, so updating a document updates the answers immediately and access control can be applied per user at retrieval time. Fine-tuning changes weights and suits teaching format, tone or a specialized task rather than supplying facts that change frequently.43. A retrieval-augmented system embeds documents into vectors. What does the embedding represent?
- A. A numerical representation of semantic meaning, so texts with similar meaning are near one another in the vector space
- B. A compressed copy of the document's exact characters
- C. An encrypted form of the document requiring a key to read
- D. The document's file metadata such as size and author
Show answer & explanation
Answer: A
Embeddings place semantically similar text nearby, which is what allows retrieval by meaning rather than by exact keyword match. Chunking strategy strongly affects retrieval quality, since chunks that are too large dilute the signal while chunks that are too small lose the context needed to answer.44. A prompt includes several worked examples before the actual question. What technique is this?
- A. Few-shot prompting, which demonstrates the desired pattern in the context rather than through weight updates
- B. Fine-tuning, which updates the model's parameters
- C. Reinforcement learning from human feedback
- D. Transfer learning across model architectures
Show answer & explanation
Answer: A
Few-shot examples steer output format and approach within the context window without any training, making it the fastest and cheapest adaptation to try first. The cost is that examples consume context on every request, so a pattern needed constantly may eventually justify fine-tuning instead.45. A model's temperature parameter is lowered toward zero. What is the effect on its output?
- A. Output becomes more deterministic and focused on the highest-probability continuations, reducing variety
- B. Output becomes more varied and creative
- C. The model processes requests faster
- D. The model's factual accuracy is guaranteed
Show answer & explanation
Answer: A
Temperature scales the sampling distribution, so low values concentrate probability on the most likely tokens and high values flatten it toward more diverse output. Determinism is not accuracy, since a model can be consistently and confidently wrong, which is why low temperature suits extraction and classification rather than guaranteeing correctness.46. A model's context window limits how much text it can consider at once. What practical consequence follows for a long document?
- A. The document must be chunked, summarized or selectively retrieved, since content beyond the window cannot influence the response
- B. The model automatically remembers earlier documents from previous requests
- C. The window can be extended arbitrarily by the caller
- D. Longer documents always produce more accurate answers
Show answer & explanation
Answer: A
Anything outside the window is simply absent from the model's view, and models are also stateless between requests unless conversation history is explicitly resent. This is why long-document workflows depend on retrieval or summarization strategies rather than on passing everything.47. A user's input to an application is inserted directly into a prompt containing system instructions. What risk does this create?
- A. Prompt injection, where crafted input causes the model to disregard the system instructions or reveal them
- B. SQL injection against the model's training data
- C. A buffer overflow in the model's memory
- D. No risk, since the model cannot act on instructions in user text
Show answer & explanation
Answer: A
Models do not reliably distinguish instructions from data when both arrive as text, so user content can redirect behaviour. Mitigations include separating instruction and data channels where the API supports it, constraining output format, validating outputs, and limiting the model's downstream permissions so a successful injection achieves little.48. An organization plans to send customer data to a hosted model for inference. What governance question must be answered first?
- A. Whether the data may lawfully be sent, whether it is retained or used for training, and which jurisdiction processes it
- B. How many tokens the prompt contains
- C. Whether the model's parameter count exceeds a threshold
- D. Which programming language the client uses
Show answer & explanation
Answer: A
Sending data to an external service is a disclosure, so lawful basis, retention terms, training use and processing location all bear on whether it is permissible. Contractual commitments on retention and training use are the practical control, and they differ substantially between offerings and tiers.49. A model trained on historical hiring decisions recommends candidates in a way that disadvantages a protected group. What is the likely cause?
- A. Bias present in the training data, which the model learned and now reproduces at scale
- B. Insufficient model capacity to fit the data
- C. A software defect in the inference code
- D. An excessively large context window
Show answer & explanation
Answer: A
A model fits the patterns in its data, so historical decisions embedding bias produce a model that reproduces it consistently and at greater scale than the humans did. Removing the protected attribute is insufficient because correlated proxy features carry the same information, which is why fairness evaluation must measure outcomes across groups directly.50. A high-stakes decision is made with model assistance. What does responsible practice require?
- A. Meaningful human review with the authority and information to override, rather than nominal oversight of an output that is accepted by default
- B. Full automation, since models are more consistent than humans
- C. Disclosure to the affected person only after the decision takes effect
- D. Human review of a random sample of decisions only
Show answer & explanation
Answer: A
Oversight is only meaningful when the reviewer has the context and standing to disagree, since rubber-stamping provides the appearance of accountability without its substance. Automation bias makes this harder in practice, as reviewers tend to defer to a confident-seeming machine output unless the process is designed to counteract that tendency.51. A model's decision must be explainable to a regulator. What tension does this create with model selection?
- A. More complex models often perform better but are harder to explain, so interpretability may justify accepting somewhat lower accuracy
- B. Complex models are always more explainable than simple ones
- C. Explainability requires publishing the training data
- D. Regulators only accept models with no accuracy metric
Show answer & explanation
Answer: A
The accuracy-interpretability trade-off is real, and in regulated decisions a model whose reasoning can be articulated may be preferable even at some performance cost. Post-hoc explanation techniques approximate a complex model's reasoning, but they explain the approximation rather than the model itself, which is an important limitation to state honestly.52. A deployed model's performance degrades gradually over months without any code change. What is the likely explanation?
- A. Drift, where the distribution of incoming data or the relationship it models has shifted away from the training conditions
- B. The model's weights decay over time in storage
- C. The inference hardware becomes slower with use
- D. The training data is deleted after deployment
Show answer & explanation
Answer: A
The world changes while the model remains fixed, so data drift in the inputs or concept drift in the input-output relationship erodes performance. Monitoring input distributions and outcome metrics in production is what detects this, and it requires ground-truth labels arriving eventually to measure the outcome side at all.53. An organization compares building a custom model against using a managed AI service for a common task such as text extraction from documents. What favours the managed service?
- A. Guaranteed superior accuracy on any specialized domain
- B. Greater control over the model architecture
- C. Elimination of the need to evaluate output quality
- D. Lower time to value and no need for training data or ML expertise, where the task is well served by a general-purpose capability
Show answer & explanation
Answer: D
Managed services suit commodity capabilities where the organization's data offers no particular advantage, and they remove the training data and expertise barrier entirely. Custom models earn their cost where the domain is specialized enough that a general service underperforms, which should be established by evaluation rather than assumed.54. A team evaluates a generative model's output quality. Why are traditional accuracy metrics often insufficient?
- A. Because many valid outputs exist for an open-ended prompt, so evaluation requires human judgment, rubrics or model-based grading against defined criteria
- B. Because generative models cannot be evaluated at all
- C. Because accuracy applies only to image models
- D. Because output quality is determined solely by response length
Show answer & explanation
Answer: A
There is no single correct string for a summary or an explanation, so exact-match metrics penalize good outputs that differ in wording. Practical evaluation defines criteria such as faithfulness to source, completeness and format compliance, then scores against them consistently, with human review anchoring any automated grader.55. A generative application's cost scales with usage. What is the primary cost driver to manage?
- A. Tokens processed, covering both input context and generated output, so prompt size and response length both matter
- B. The number of users registered in the application
- C. The storage consumed by the application's database
- D. The number of programming languages used
Show answer & explanation
Answer: A
Both the context sent and the text generated are billed, which is why large retrieved contexts and verbose responses drive cost as much as request volume. Caching repeated responses, trimming retrieved context to what is relevant and constraining output length are the practical levers.56. An AI system will be used in a domain where errors carry serious consequences. What should the deployment approach be?
- A. The same process used for a low-stakes internal tool
- B. Deployment without a disable mechanism to prevent misuse of the off switch
- C. Staged rollout with defined performance thresholds, monitoring and the ability to disable the system, matching oversight to the consequence of error
- D. Full deployment immediately, with monitoring added later
Show answer & explanation
Answer: C
Risk-proportionate deployment scales the rigour of evaluation, monitoring and human oversight to what failure would cost, which is the organizing principle behind emerging AI regulation. The ability to disable quickly is essential, since discovering a systematic error is only useful if the system can be stopped.57. A model is fine-tuned on proprietary data. What governance consideration arises about the resulting model?
- A. The tuned model may reflect or reveal characteristics of the training data, so access to it must be controlled as the underlying data would be
- B. The tuned model contains no information about its training data
- C. Access controls become unnecessary once data is in weights
- D. The tuned model may be shared publicly without further review
Show answer & explanation
Answer: A
Weights encode information from the tuning data and models can reproduce memorized content, particularly distinctive or repeated examples. Treating the tuned model as a derivative of the data it was trained on, subject to the same access restrictions, is the conservative and defensible position.58. An organization must decide what to log about AI system usage. What supports both operations and accountability?
- A. Only the total number of requests per day
- B. Only errors returned by the model API
- C. Inputs, outputs, model and version identifiers, and the decision taken, retained subject to privacy requirements
- D. Nothing, to minimize data retention
Show answer & explanation
Answer: C
Reconstructing why a system produced a particular result requires knowing what it was given, which model version answered and what was done with the response. This has to be balanced against the privacy implications of retaining prompts that may contain personal data, which is why retention periods and access controls are part of the logging design.59. A model is given the ability to call external tools and take actions. What control is essential?
- A. Constraining the available actions and their permissions so a manipulated or mistaken model cannot exceed what the task requires
- B. Granting broad permissions so the model can handle unforeseen cases
- C. Relying on the model's instructions to prevent misuse
- D. Removing all logging to reduce data retention
Show answer & explanation
Answer: A
Instructions are not a security boundary because prompt injection can override them, so the boundary must be the permissions actually granted to the tools. Least privilege applied to the action surface, plus confirmation steps for irreversible operations, is what limits the consequence of a model behaving unexpectedly.60. A team considers whether a problem needs machine learning at all. What indicates a conventional rules-based approach is preferable?
- A. The rules are known, stable and few enough to express explicitly, since a deterministic implementation is then cheaper, testable and explainable
- B. The organization has already purchased machine learning tooling
- C. The dataset is very large
- D. Competitors have announced machine learning features
Show answer & explanation
Answer: A
Machine learning earns its complexity where the rules are unknown, numerous or shifting, and applying it to a problem with known logic adds cost, opacity and failure modes for no benefit. Establishing a rules-based or simple statistical baseline first also gives the honest comparison point that determines whether a model adds anything.61. A classification threshold is adjusted from 0.5 to 0.8. What is the expected effect?
- A. Fewer positive predictions, raising precision while lowering recall, since only higher-confidence cases are classified positive
- B. More positive predictions with both precision and recall rising
- C. No effect, since the threshold does not change the model
- D. The model must be retrained for the change to take effect
Show answer & explanation
Answer: A
Raising the threshold makes the classifier more conservative about predicting positive, so it makes fewer positive calls but is right more often when it does. The threshold is chosen from the relative cost of a false positive against a false negative, and it is a deployment decision requiring no retraining.62. A team notices their model performs far better in evaluation than in production. What should be suspected?
- A. Data leakage, where information unavailable at prediction time was present in the training or evaluation data
- B. Underfitting of the training data
- C. Insufficient inference hardware in production
- D. An excessively small model architecture
Show answer & explanation
Answer: A
Leakage most often comes from features derived after the outcome was known, or from splitting data randomly when it should have been split by time or by entity. It produces evaluation results that look excellent and cannot be reproduced in production, which is why the split strategy deserves as much scrutiny as the model.63. A generative model is asked to work through a problem step by step before giving an answer. What is this technique and why does it help?
- A. Chain-of-thought prompting, which improves performance on multi-step reasoning by allocating intermediate generation to the problem
- B. Fine-tuning, which adjusts the model's weights toward reasoning
- C. Quantization, which reduces the model's precision
- D. Distillation, which transfers knowledge to a smaller model
Show answer & explanation
Answer: A
Generating intermediate steps gives the model more computation applied to the problem and makes the reasoning inspectable, which helps most on tasks requiring several dependent inferences. The stated reasoning is not a guaranteed account of how the answer was produced, so it aids verification without being proof of correctness.64. An organization wants model outputs restricted to safe and on-topic content. What layered approach is appropriate?
- A. Relying solely on the model provider's built-in safety training
- B. Input filtering, model-level safety configuration and output validation before the response reaches the user
- C. Reviewing outputs manually after users have received them
- D. Filtering only the user's input
Show answer & explanation
Answer: B
Each layer catches what the others miss, and provider safety training is a general-purpose baseline that cannot know an organization's specific policy or topic boundaries. Post-hoc review has value for improvement but does not prevent the user seeing the output, which is the harm the controls exist to avoid.65. A generated image or text will be published. What transparency practice is increasingly expected?
- A. Disclosing that the content was generated or assisted by AI, so recipients can weigh it accordingly
- B. Presenting the content as human-authored to avoid distraction
- C. Disclosure only if the content is later found to be inaccurate
- D. No disclosure, since the tool used is an internal matter
Show answer & explanation
Answer: A
Disclosure lets audiences calibrate their trust and is increasingly required by regulation and platform policy, particularly for synthetic media depicting people or events. Watermarking and provenance metadata are the technical complements, though both can be stripped, which is why the disclosure practice matters independently.66. A team must estimate whether a machine learning project is feasible before committing. What is the most decisive early question?
- A. Whether sufficient representative data with reliable labels exists or can be obtained, since no modelling choice compensates for its absence
- B. Which framework the team prefers
- C. How many GPUs are available for training
- D. Whether the results will be presented in a dashboard
Show answer & explanation
Answer: A
Data availability and label quality bound what any model can achieve, and projects fail on this constraint far more often than on algorithm selection or compute. Establishing how labels will be obtained, how much they cost and how reliable they are is therefore the first feasibility question rather than a later implementation detail.
More in this family
Explore more Technology & IT Certifications
In the same family
More in this category
- Microsoft Certified: Azure Administrator Associate (Exam AZ-104)Practice questions →
- Microsoft Azure AI FundamentalsPractice questions →
- Microsoft Certified: Azure Fundamentals (AZ-900)Practice questions →
- Cisco Certified Network Associate (200-301 CCNA)Practice questions →
- Cisco Certified Support Technician (CCST) NetworkingPractice questions →
- Certified Information Systems Security Professional (CISSP)Practice questions →
- Certified Kubernetes Administrator (CKA)Practice questions →
- Claude Certified Associate – FoundationsPractice questions →
- CompTIA A+ (Core 1: 220-1201 and Core 2: 220-1202)Practice questions →
2026 statistics
Key facts: AWS AI Practitioner exam
Every free resource for this exam
Get a free AWS AI Practitioner study plan
A week-by-week plan plus new practice questions, straight to your inbox.
Official sources
Primary documents used to verify the exam details shown on this page.
- AWS Certified AI Practitioner (AIF-C01) Exam GuideAmazon Web Services (AWS)docs.aws.amazon.com
- AWS Certified AI Practitioner — Certification OverviewAmazon Web Services (AWS)aws.amazon.com
- AWS Training and Certification FAQsAmazon Web Services (AWS)aws.amazon.com
- Schedule an AWS Certification ExamAmazon Web Services (AWS)aws.amazon.com
- AWS Certification — Pearson VUE Test DeliveryPearson VUEpearsonvue.com
Last verified against the official exam content outline:
Frequently asked questions
How many questions are on the AWS Certified AI Practitioner exam, and what score do I need to pass?
The exam has 65 questions delivered in a 90-minute testing window, which works out to just under 90 seconds per question on average — a useful pacing target for timed practice sessions. Only 50 of the 65 questions actually affect your score; the other 15 are unscored questions that AWS uses for evaluation, and you won't know which is which, so treat every question as if it counts. Results are reported on a scaled score of 100 to 1,000, and the minimum passing score is 700. The exam uses a compensatory scoring model, meaning you don't need to pass each section individually — a strong showing in one domain can offset a weaker one. When you take full-length practice exams, aim to score comfortably above the 700 threshold so normal test-day variance doesn't put you at risk.
Which topics should I prioritize when practicing for the AWS Certified AI Practitioner exam?
The exam is organized into 5 content domains, and their weightings tell you exactly where to spend your practice time. Domain 3, Applications of Foundation Models, is the largest at 28% of scored content, followed by Domain 2, Fundamentals of GenAI, at 24%. Together those two generative-AI-focused domains account for over half the scored exam, so they should anchor your practice plan. Domain 1, Fundamentals of AI and ML, covers 20%, while Domain 4, Guidelines for Responsible AI, and Domain 5, Security, Compliance, and Governance for AI Solutions, each make up 14%. Don't skip the two 14% domains entirely, though — because scoring is compensatory across the whole exam, reliable points from responsible AI and security questions can offset misses in the heavier domains.
How much does the AWS Certified AI Practitioner exam cost, and where can I take it?
The exam costs 100 USD, which is the standard fee for AWS Foundational-level exams. All AWS Certification exams, including this one, are delivered through Pearson VUE — you schedule from your AWS Certification Account. You have two delivery options: an in-person appointment at a Pearson VUE testing center, which is a facility operated independently from AWS that offers in-person proctoring, or an online proctored exam through OnVUE, which lets you test from any private space such as your home or office. If your plans change, you can cancel or reschedule up to 24 hours before your appointment without additional fees — a helpful buffer if your practice scores suggest you need another week or two of preparation before test day.
When will I get my exam results, and how long does the certification last?
Exam results are available within five business days after you complete the exam, so you won't get an official pass/fail the moment you finish. Once you pass, the certification is valid for 3 years, and you'll need to recertify every three years to keep it active. That timeline is worth factoring into your study plan: because results can take up to a week, schedule your exam with enough lead time if you're targeting a deadline like a job application or performance review. And since the credential expires, keeping your practice materials and study notes organized now will make recertification prep noticeably faster three years from now.