15 real AIF-C01 sample questions, each with a worked explanation and a rationale for every option, right and wrong. No account, no card. This is the reasoning the AIF-C01 tests: knowing why the tempting answer is wrong, not just spotting the right one.
The real AIF-C01 is 65 questions in 90 minutes, pass mark 700 / 1000. For a domain-by-domain breakdown and a study plan, read the AIF-C01 study guide. The full bank has 328 questions.
lock_openFree sampleApplications of foundation modelsmedium
A developer is building a customer-support chatbot that must answer questions about a company's product catalogue, which is updated weekly. The catalogue contains roughly 200,000 tokens of text. Which design consideration most directly determines whether a single foundation model call can process the entire catalogue without retrieval augmentation?
- AThe model's training data cutoff date, because any catalogue content newer than the cutoff will be ignored by the model during inference.
- BThe model's output token limit, because generating long catalogue summaries requires a high output ceiling to avoid truncation.
- CThe model's context window size, because the full catalogue must fit within the input limit for a single inference call.check_circle Correct
- DThe model's temperature setting, because a lower temperature ensures the model attends to all catalogue entries rather than sampling selectively.
Understand that a model's context window size is the primary architectural constraint when deciding whether retrieval augmentation is needed. Foundation models have a fixed context window - the maximum number of tokens they can process in a single inference call, covering both input and output. When source material exceeds this window, the application must use retrieval-augmented generation (RAG) or chunking strategies to surface relevant content before the model call. Training cutoff, temperature, and output limits are real model properties but do not govern input capacity.
Why A is wrong: Training cutoff affects factual knowledge baked into weights, not the ability to process input tokens supplied at inference time. Providing the catalogue as context bypasses the cutoff problem entirely, so this is not the limiting factor here.
Why B is wrong: Output token limits affect response length, not input capacity. The question asks about processing the catalogue as input, so output limits are irrelevant to this constraint.
Why C is correct: Context window defines the maximum tokens a model can accept per call. If the catalogue exceeds that limit, the model cannot process it in one shot, making retrieval augmentation necessary.
Why D is wrong: Temperature controls output randomness and has no bearing on how many input tokens the model can accept. It is a tempting distractor because 'attending to all entries' sounds plausible, but attention and context length are separate concepts.
lock_openFree sampleApplications of foundation modelsmedium
A retail company wants to use a foundation model to generate product descriptions. Legal requires that the model never produce outputs containing competitor brand names. Which mechanism is the most direct way to enforce this constraint at the application layer?
- AReduce the model's temperature to zero so it always picks the highest-probability token, making competitor names statistically unlikely.
- BFine-tune the foundation model on internal product data only, so the model's weights no longer associate competitor brand names with product description tasks.
- CUse a larger foundation model with higher reasoning capability, because larger models are better at following nuanced instructions to avoid competitor references.
- DApply output guardrails that scan generated text and block or redact any response containing prohibited brand names before it reaches the user.check_circle Correct
Understand that guardrails are the appropriate mechanism for enforcing deterministic content restrictions on foundation model outputs. Guardrails are application-layer components that inspect model outputs against defined rules - such as blocklists of prohibited terms - before the response is delivered to end users. Unlike temperature adjustments or model-size choices, which influence probability distributions, guardrails provide deterministic enforcement. This makes them the correct tool when a hard constraint (legal compliance) must hold on every inference call.
Why A is wrong: Greedy decoding (temperature 0) selects the most probable token at each step but cannot guarantee exclusion of specific words. Competitor names may still appear if they are highly probable in context.
Why B is wrong: Fine-tuning can shift output distributions but does not reliably eliminate specific outputs. The model retains knowledge of competitor names from pre-training and may still generate them in novel contexts, so this approach does not provide the deterministic enforcement legal requires.
Why C is wrong: Model size improves general instruction-following but is probabilistic. A larger model may comply more often, but compliance is not guaranteed on every call. Guardrails provide the deterministic enforcement the question calls for.
Why D is correct: Guardrails act as a post-generation filter at the application layer, providing a deterministic enforcement mechanism independent of model probabilities. This directly satisfies the legal requirement by intercepting non-compliant outputs before delivery.
lock_openFree sampleApplications of foundation modelsmedium
An organisation is selecting a foundation model for a document summarisation service. The documents are technical engineering reports averaging 15,000 tokens each. The team has shortlisted two models: Model X supports a 4,096-token context window and Model Y supports a 32,000-token context window. Both models have comparable quality benchmarks. Which selection criterion most directly favours Model Y for this use case?
- AModel Y's larger context window allows full-document summarisation without chunking, reducing implementation complexity and potential coherence loss.check_circle Correct
- BModel Y's larger context window means it was trained on more data, so its factual knowledge is more comprehensive for engineering topics.
- CModel Y's larger context window lowers per-token inference cost because fewer API calls are needed when processing large batches of documents.
- DModel Y's larger context window enables the model to generate longer summaries, meeting any output length requirements imposed by the engineering team.
Identify context window size as the decisive model selection criterion when source documents exceed one model's input limit. Context window is the total token budget for a single inference call (input plus output). When documents average 15,000 tokens and one candidate model caps at 4,096 tokens, that model cannot process a full document in one call. Model Y's 32,000-token window accommodates the full document, eliminating chunking complexity and preserving cross-section coherence. Model selection must therefore account for context window relative to expected input length.
Why A is correct: Because the average document (15,000 tokens) exceeds Model X's 4,096-token window, Model X would require chunking and then re-stitching summaries, introducing complexity and risk of losing cross-section context. Model Y can ingest the full document in a single call.
Why B is wrong: Context window size is an architectural inference-time property and is not correlated with training dataset size or domain coverage. Conflating these is a common misunderstanding. The selection advantage here is purely about input capacity at inference time.
Why C is wrong: Larger context windows typically carry higher per-token costs, not lower ones, because they require more memory and compute. Fewer API calls may reduce overhead but per-token pricing generally increases with window size. Cost is not the selection advantage cited in the question.
Why D is wrong: Context window size controls input capacity, not output length directly. Output length is governed by max output token settings. The relevant advantage here is the ability to ingest 15,000-token documents as input, not to produce longer outputs.
lock_openFree sampleGenerative AI fundamentalsmedium
A support team sends the same prompt to a hosted LLM many times a day and gets noticeably different wording each time. They want answers to stay closely on-message and vary far less, without demanding byte-for-byte identical output. Which change best achieves this?
- ALower the temperature toward 0.2 so the model concentrates on the highest-probability tokens.check_circle Correct
- BRaise the temperature toward 1.0 so the model considers a wider spread of next tokens.
- CIncrease the maximum output-token limit so each answer has room to finish completely.
- DRaise top-p toward 1.0 so the model samples from the full probability distribution.
Lowering temperature concentrates output on the most-probable tokens, reducing run-to-run variation. Temperature scales how sharply the model favours its most-probable tokens. Lowering it toward 0 concentrates probability mass on the top tokens, so the same prompt produces closely similar, on-message wording each time without forcing fully identical output.
Why A is correct: Correct. A low temperature sharpens the distribution toward the likeliest tokens, so wording stays consistent run to run while still allowing slight variation.
Why B is wrong: Higher temperature flattens the token distribution and increases variation, which moves away from the on-message consistency the team asked for.
Why C is wrong: The token limit governs how long an answer can be, not how much its wording varies, so consistency is unchanged.
Why D is wrong: A top-p of 1.0 keeps the entire distribution in play, which preserves or increases variation rather than reducing it.
lock_openFree sampleGenerative AI fundamentalsmedium
A developer is tuning a model's top-p (nucleus) sampling value and wants to be sure they understand its effect. Which statement best describes how top-p selects the tokens the model samples from?
- AIt caps the total number of tokens the model is allowed to generate in the response.
- BIt restricts sampling to the smallest set of top tokens whose cumulative probability reaches p.check_circle Correct
- CIt sets a fixed random seed so the same prompt always returns the same response.
- DIt scales the model's confidence by reweighting every token's probability before sampling.
Top-p sampling draws from the smallest set of tokens whose cumulative probability reaches p. Top-p, or nucleus sampling, ranks tokens by probability and keeps the smallest set whose cumulative probability reaches the threshold p, discarding the long tail. The model then samples only from that nucleus, so a lower p narrows the candidate set.
Why A is wrong: That describes a maximum-output-length setting, not top-p. Top-p shapes the candidate token set, not the response length.
Why B is correct: Correct. Nucleus sampling keeps the smallest group of most-likely tokens whose probabilities sum to p, then samples only from that group.
Why C is wrong: A seed governs reproducibility of the random draw, not which tokens are eligible, so it is a different control from top-p.
Why D is wrong: Reweighting the whole distribution by a factor describes temperature, not the cumulative-probability cutoff that top-p applies.
lock_openFree sampleGenerative AI fundamentalsmedium
An application caches model responses using the exact prompt text as the cache key, to avoid paying for repeated identical calls. For the cache to be useful, the same prompt must return the same answer. Which setting makes the output deterministic?
- ASet top-p to 1.0 so the model can draw from every available token at each step.
- BUse a larger model with more parameters and a wider context window for the call.
- CSet the temperature to 0 so the model always selects the highest-probability token.check_circle Correct
- DRaise the temperature to 0.8 to make the cached answers more natural and fluent.
Temperature 0 produces greedy, deterministic decoding suitable for caching identical prompts. With temperature 0 the model decodes greedily, always taking the single highest-probability token rather than sampling. The same prompt therefore yields the same output every call, which is exactly what a prompt-keyed response cache depends on.
Why A is wrong: A top-p of 1.0 leaves the full distribution available for sampling, so repeated calls can still diverge and the cache would rarely hit.
Why B is wrong: Model size affects quality and cost, not whether sampling is deterministic, so a bigger model still varies unless sampling is fixed.
Why C is correct: Correct. Temperature 0 forces greedy decoding: the single most-probable token is chosen at each step, so the same prompt yields the same answer.
Why D is wrong: A temperature of 0.8 increases variation, which is the opposite of the reproducibility a prompt-keyed cache needs.
lock_openFree sampleFundamentals of AI and MLeasy
A new analyst is learning how the terms artificial intelligence, machine learning, and deep learning relate to one another as nested fields. Which two statements correctly describe this relationship? Select TWO.
- AArtificial intelligence is one specialised technique that sits entirely inside the broader field of machine learning.
- BMachine learning is a subset of artificial intelligence in which systems learn patterns from data rather than following hand-coded rules.check_circle Correct
- CDeep learning and machine learning are two separate disciplines that share no methods and never overlap in practice.
- DDeep learning is a subset of machine learning that uses neural networks with many stacked layers to learn features.check_circle Correct
- EArtificial intelligence, machine learning, and deep learning are interchangeable synonyms for exactly the same single concept.
Deep learning sits inside machine learning, which in turn sits inside artificial intelligence, forming three nested fields of progressively narrower scope. The three terms form concentric sets: artificial intelligence is the broadest goal of machine intelligence, machine learning is the subset that learns from data, and deep learning is the further subset using many-layered neural networks. Correctly ordering this containment is the core terminology distinction.
Why A is wrong: This reverses the actual nesting; machine learning is a subset of artificial intelligence, not the other way around, so the containment is backwards.
Why B is correct: Machine learning is correctly framed as a branch of the wider artificial intelligence field, defined by learning from data instead of explicit rules.
Why C is wrong: This is tempting because the names differ, but deep learning is a subset of machine learning, so the two clearly overlap rather than being separate.
Why D is correct: Deep learning is accurately placed inside machine learning and characterised by multi-layer neural networks that learn representations directly.
Why E is wrong: Treating the three as identical ignores their nesting; each names a progressively narrower field, so they are related but not interchangeable.
lock_openFree sampleFundamentals of AI and MLeasy
A retail company wants a system that improves its product recommendation accuracy over time by learning from customer purchase history, without being explicitly programmed with new rules each season. Which AI discipline best describes this approach?
- ARule-based expert systems, which encode domain knowledge as hand-crafted if-then rules.
- BMachine learning, which trains models on historical data so that predictive behaviour improves with experience.check_circle Correct
- CBusiness intelligence reporting, which aggregates and visualises historical sales data for human analysts.
- DRobotic process automation, which executes predefined workflow steps to mimic human repetitive tasks.
Understand that machine learning is the AI discipline in which models improve from data without explicit rule updates. Machine learning algorithms discover patterns in training data and adjust internal parameters so that predictions become more accurate as more data is seen. This contrasts with rule-based systems, which require a human to codify every new condition. The self-improving property from historical purchase data is the hallmark of ML.
Why A is wrong: Tempting because rule-based systems do drive recommendations, but they require explicit programming of every new rule. They do not learn autonomously from data, which is the defining requirement in the stem.
Why B is correct: Machine learning is the discipline in which algorithms learn patterns from data and improve performance on a task without being explicitly re-programmed, matching the scenario precisely.
Why C is wrong: Tempting because it uses historical purchase data, but BI reporting surfaces information for human decision-making rather than training a model that autonomously improves its own predictions.
Why D is wrong: Tempting because RPA is often grouped with AI tooling, but it follows a fixed script and does not learn from data. It cannot improve recommendation accuracy through experience.
lock_openFree sampleFundamentals of AI and MLeasy
An engineer is choosing between two model architectures for an image classification task. Architecture X is a shallow network with two hidden layers. Architecture Y uses many stacked layers of learned feature representations. Which term correctly describes Architecture Y?
- AA shallow neural network, because it relies on manually extracted pixel features passed to a classifier.
- BA decision tree ensemble, because it combines many weak learners into a stronger composite model.
- CA deep learning model, because it uses multiple stacked layers to learn hierarchical feature representations automatically.check_circle Correct
- DA linear regression model, because it maps input features directly to an output without hidden transformations.
Distinguish deep learning from other model families by recognising its defining characteristic of stacked learned feature layers. Deep learning models are neural networks with multiple hidden layers. Each layer learns increasingly abstract representations of the input - early layers detect edges, later layers detect shapes, and so on. This automatic feature hierarchy is what separates deep learning from shallow models or non-neural approaches.
Why A is wrong: Tempting because Architecture X is described as shallow, but Architecture Y stacks many layers and learns hierarchical features automatically. Calling it shallow contradicts the 'many stacked layers' description in the stem.
Why B is wrong: Tempting because ensembles also improve over single models, but a decision tree ensemble does not consist of stacked feature-learning layers. The stem specifically describes a layered network architecture.
Why C is correct: Deep learning is defined by neural networks with many hidden layers that learn increasingly abstract representations, matching the description of Architecture Y in the stem.
Why D is wrong: Linear regression has no hidden layers and learns no hierarchical representations. Choosing this option ignores the central description of stacked layers in the stem.
lock_openFree samplePrompt engineering and evaluationmedium
A developer is building a customer support chatbot that must classify incoming messages into one of five predefined categories. The model occasionally misclassifies edge cases. Which prompt engineering technique is most likely to improve classification accuracy without fine-tuning?
- AIncrease the model's temperature setting to make responses more varied and exploratory.
- BAdd a few labelled examples of each category directly in the prompt before the user message.check_circle Correct
- CAppend a system message instructing the model to respond only in JSON format.
- DBreak the prompt into multiple sequential calls so each call handles one category independently.
Understand how few-shot prompting anchors a model to a correct label space and improves classification accuracy on edge cases. Few-shot prompting supplies the model with labelled input-output pairs that demonstrate the desired mapping. This narrows the model's interpretation of the task and reduces the probability of it choosing an incorrect label, especially for inputs that sit near the boundary between categories.
Why A is wrong: Higher temperature increases randomness, which is likely to worsen classification consistency rather than improve it; it does not address the root cause of misclassification.
Why B is correct: Few-shot prompting provides the model with concrete examples of the task, anchoring its output distribution to the correct label space and reducing ambiguity on edge cases.
Why C is wrong: Enforcing JSON output controls format, not classification accuracy. Without category examples, the model still has the same ambiguity about which label to assign.
Why D is wrong: Running one call per category would require five API calls and still relies on the model making a binary yes/no decision per category, which does not resolve the original classification ambiguity.
lock_openFree samplePrompt engineering and evaluationmedium
A team needs multiple developers to call the same foundation model for a document summarisation task. They want consistent output structure and tone across all callers, with minimal per-call effort. What is the most appropriate approach?
- ARequire each developer to write a new prompt from scratch and compare outputs in code review.
- BFine-tune the model on summarised documents so it learns the format implicitly, removing the need for detailed prompts.
- CStore a shared prompt template with fixed instructions and placeholders for variable content, and have each caller fill in the placeholders.check_circle Correct
- DUse a retrieval-augmented generation pipeline so the model fetches format instructions dynamically at inference time.
Understand how prompt templates standardise model behaviour across multiple callers by separating fixed instructions from variable inputs. A prompt template defines the constant elements - task description, output format constraints, tone guidance - once, and exposes only the variable slots (such as document text) to callers. This ensures every invocation operates from the same instructional baseline, producing consistent structure and tone without per-caller prompt engineering effort.
Why A is wrong: Writing prompts from scratch per developer introduces inconsistency and defeats the goal of uniform structure and tone; code review does not substitute for a shared baseline.
Why B is wrong: Fine-tuning is expensive and requires a labelled dataset; it also does not solve the per-developer consistency problem if callers still craft ad-hoc prompts alongside the fine-tuned model.
Why C is correct: Prompt templates centralise the fixed instructions and structural constraints, so every caller inherits the same tone and output format while only supplying the variable inputs specific to their document.
Why D is wrong: RAG is designed to supply factual context, not to enforce prompt structure; it adds latency and complexity without addressing the core need for a shared fixed instruction set.
lock_openFree samplePrompt engineering and evaluationmedium
A product team wants a model to generate a short product description in a warm, conversational tone. The model currently produces formal, technical language. No labelled dataset is available for fine-tuning. Which prompt engineering adjustment is most effective?
- AReduce the maximum token limit so the model is forced to use simpler vocabulary.
- BSet the system message to state 'You are a helpful AI assistant' without specifying tone.
- CDeploy the model behind an API proxy that post-processes output to replace formal words with informal synonyms.
- DInclude one or two example product descriptions written in the desired tone within the prompt.check_circle Correct
Apply few-shot prompting to steer model output tone when no fine-tuning dataset is available. Few-shot examples act as in-context demonstrations that shift the model's output distribution toward the demonstrated style. By showing one or two conversational product descriptions inside the prompt, the model infers the expected register and applies it to the new product, without any weight updates.
Why A is wrong: Capping tokens constrains length but does not steer tone; the model can produce formal, technical language in very few tokens if that remains its dominant pattern.
Why B is wrong: A generic assistant persona instruction does not communicate a specific stylistic register; the model is likely to retain its default formal output without explicit tone guidance.
Why C is wrong: Post-processing with a synonym substitution layer is fragile, error-prone, and does not teach the model anything; it also does not adjust sentence structure or phrasing patterns, only individual words.
Why D is correct: One or two few-shot examples written in a warm, conversational tone condition the model on the target style directly, shifting output register without requiring a dataset or model update.
lock_openFree sampleResponsible and secure AIeasy
A data science team notices that a credit-scoring model approves loans at a significantly lower rate for one demographic group compared to all others, even though that group has similar repayment histories. Which practice most directly addresses this disparity during model development?
- AEvaluating disaggregated fairness metrics across demographic subgroups and retraining with corrected data or constraintscheck_circle Correct
- BIncreasing the model's overall accuracy on the full training dataset
- CDeploying the model behind an API so that end users cannot inspect its decision logic
- DRemoving all demographic attributes from the training features before the initial model fit
Understand how disaggregated fairness evaluation and targeted mitigation correct demographic disparity in AI model outputs. Responsible AI development requires measuring model behaviour separately for each subgroup rather than relying on aggregate metrics. Disaggregated fairness metrics (such as equalised odds or demographic parity difference) reveal where a model fails specific groups, enabling corrective actions such as data augmentation, re-weighting, or in-training fairness constraints that reduce the disparity directly.
Why A is correct: Disaggregated evaluation exposes differential performance per subgroup, allowing targeted mitigation such as re-sampling, re-weighting, or fairness constraints during retraining.
Why B is wrong: Tempting because higher accuracy sounds like a better model, but aggregate accuracy can improve while bias against a subgroup worsens if that group is a small fraction of the dataset.
Why C is wrong: Hiding decision logic via an API does nothing to correct the underlying bias; it only reduces transparency, which compounds responsible AI concerns.
Why D is wrong: Removing protected attributes is a common first step but does not eliminate bias because correlated proxy features (postcode, occupation) can still encode demographic information.
lock_openFree sampleResponsible and secure AIeasy
An organisation is preparing to publish an open-source machine learning model. To help downstream users understand the model's intended uses, limitations, and known biases, which artefact should they produce?
- AA data dictionary listing every feature in the training dataset
- BA model card documenting intended use cases, performance metrics, and known limitationscheck_circle Correct
- CA deployment runbook describing the infrastructure required to serve the model
- DA README file listing the Python package dependencies and installation steps
Identify the model card as the standard responsible AI artefact for communicating intended use, limitations, and known biases to downstream users. Model cards, introduced to formalise responsible disclosure, are structured documents that accompany a model and cover its intended applications, out-of-scope uses, evaluation across relevant demographic subgroups, and known limitations. They allow downstream developers and organisations to assess fit-for-purpose and potential harms before integrating the model into their own systems.
Why A is wrong: A data dictionary describes features but does not communicate intended use, limitations, performance across subgroups, or ethical considerations - the core content of a model card.
Why B is correct: A model card is the standardised responsible AI artefact specifically designed to disclose a model's purpose, evaluation results, fairness considerations, and constraints so users can make informed deployment decisions.
Why C is wrong: A deployment runbook covers operational setup, not model behaviour or ethical considerations. It addresses how to run the model, not whether or how it should be used.
Why D is wrong: Installation documentation helps users set up the software but contains none of the bias, fairness, or intended-use information that responsible AI disclosure requires.
lock_openFree sampleResponsible and secure AIeasy
A company wants to incorporate human review into its AI content moderation pipeline to catch errors before decisions affect users. Which responsible AI practice does this reflect?
- ATransfer learning, where the model is pre-trained on a large corpus before fine-tuning
- BHyperparameter tuning, where model configuration is optimised on a held-out validation set
- CHuman-in-the-loop oversight, where a person reviews or can override model decisions at defined checkpointscheck_circle Correct
- DBatch inference, where model predictions are computed on large groups of inputs simultaneously
Identify human-in-the-loop oversight as the responsible AI practice that places human review at critical decision points in an AI pipeline. Human-in-the-loop (HITL) oversight keeps a person in the decision chain for consequential outcomes. In content moderation, this means reviewers can examine borderline cases, override incorrect model decisions, and feed corrections back into retraining data, reducing both false positives (incorrect removals) and false negatives (missed violations) while maintaining accountability.
Why A is wrong: Transfer learning is a training technique for improving model performance, not a governance mechanism for reviewing individual decisions before they affect users.
Why B is wrong: Hyperparameter tuning improves model accuracy offline; it does not place a human reviewer in the live decision path to catch runtime errors.
Why C is correct: Human-in-the-loop is the responsible AI practice of retaining human judgement at critical decision points, reducing risk from model errors, edge cases, or unexpected inputs.
Why D is wrong: Batch inference is an efficiency pattern for generating predictions at scale; it does not involve human review of individual decisions before they take effect.
Examworthy is not affiliated with or endorsed by Amazon Web Services. All questions are original, blueprint-aligned practice material. We never reproduce live exam items. AIF-C01 and related marks belong to their respective owners.
