AIF-C01 sample questions — 25 answered and explained

Overfitting: a model fits its training data too closely, including noise, and generalizes poorly to new data.

The deciding signal is the gap: near-perfect on training data, much worse on unseen data from the same distribution. That is overfitting. Data drift is the tempting distractor because performance on new data is poor, but drift happens over time in production as real-world data changes; here the test data is from the same source and period and the gap shows up immediately. Underfitting would be poor on the training data too.

1. Compare training performance with performance on unseen data.
2. A large gap with strong training results points to overfitting.
3. Rule out drift by noting the test data matches the training distribution and time.

Exam tip: Great on training data, poor on new data from the same distribution means overfitting.

Underfitting: a model is too simple to capture the real pattern, so it performs poorly even on its training data.

Poor performance on the training data and similar poor performance on validation data is the signature of underfitting: the model has not captured the pattern at all. Overfitting is the opposite (strong on training, weak on new data), a good fit is strong on both, and data drift describes production data changing over time rather than a model that cannot learn its own training data.

1. Check the training score first: poor training performance rules out overfitting.
2. Poor on both training and validation data points to underfitting.
3. Rule out drift, which concerns production data changing over time.

Exam tip: Bad on training and bad on new data means underfitting.

NLP is the field of AI that understands and works with human language, such as sentiment analysis and entity extraction.

Judging the writer's emotion (sentiment analysis) and pulling out product names (entity extraction) from written email text are natural language processing tasks. Optical character recognition is tempting because it deals with text, but it only turns images of characters into text and does not interpret meaning. Computer vision works on images and speech recognition on audio.

1. Identify the input type: written email text.
2. Identify the tasks: sentiment detection and entity extraction.
3. Match understanding the meaning of text to NLP; rule out image, audio, and character-extraction fields.

Exam tip: Understanding the meaning of text (sentiment, entities) is NLP.

Computer vision interprets images and video.

Detecting objects or spotting defects in images and video is a computer vision task. NLP works on human language and speech recognition on spoken audio, while forecasting predicts future numeric values, so none of them interprets visual input.

1. Identify the input as images or video.
2. Match visual interpretation to computer vision.
3. Rule out language, audio, and numeric-prediction fields.

Exam tip: Interpreting images and video is computer vision.

A large language model (LLM) is a very large neural network trained on huge amounts of text so it can understand and generate language.

Billions of parameters, pre-training on a vast body of text, and generating each reply by predicting the next word are the defining traits of a large language model, the engine behind most generative AI text applications. An embedding model is the closest twin, since it is also trained on text, but it outputs vectors rather than generated language. A diffusion model is generative but produces images, and a speech recognition model transcribes audio instead of writing new text.

1. Note the training data: a vast amount of written text.
2. Note the behavior: it generates new text by predicting what comes next.
3. Separate text models that generate (LLMs) from text models that only encode (embedding models).
4. Match a huge text-trained model that generates language to an LLM.

Exam tip: Huge neural network + vast text + generates language = LLM.

A transformer is the neural-network architecture behind most modern large language models.

Its attention mechanism lets the model weigh how much each token relates to the others, which is why transformer-based LLMs handle long, context-rich text well. Token-based pricing, vector databases and tokenization are real parts of a GenAI application, but none of them is the model architecture.

1. Associate the transformer with model architecture.
2. Recall that its attention mechanism relates tokens to each other.
3. Separate it from pricing, storage and tokenization.

Exam tip: The transformer, built on attention, is the architecture that powers today's LLMs.

Summarization is the GenAI use case of condensing long documents or threads into a short brief.

The need is to turn a long thread into a short brief, which is summarization. Translation changes the language, code generation writes code, and semantic search finds relevant content, so none of them produces a condensed version of the thread.

1. Find the verb in the scenario: condense long content.
2. Map that verb to the summarization use case.
3. Eliminate translating, coding and searching.

Exam tip: Condensing long content into a short brief is summarization.

Code generation is the GenAI use case of producing software from a plain-language request.

Turning a description of behaviour into a working function is code generation. Summarization condenses existing content, semantic search retrieves content by meaning, and a recommendation engine suggests relevant items, so none of them writes the function.

1. Identify the output: new code written from a description.
2. Match it to code generation.
3. Rule out summarizing, searching and recommending.

Exam tip: Writing code from a plain-language request is the code generation use case.

Token-based pricing charges for the number of input and output tokens processed, so cost scales with the amount of text per call.

Both chatbots make the same number of calls, so a per-call price would make their bills equal, and reserved capacity is billed by time rather than by prompt size. The only difference is that one sends a 40-page document as input on every call; with token-based pricing those extra input tokens are billed each time.

1. Note what is equal (request count) and what differs (text sent per request).
2. Recall that on-demand inference bills input plus output tokens.
3. Conclude that token-based pricing explains why a longer prompt costs more.

Exam tip: Token-based pricing charges for input plus output tokens, so the same request count can cost far more when prompts are long.

Context engineering is the practice of giving a model the right information, in the right amount, at the right time so it produces accurate, relevant answers.

A foundation model only knows what it learned in training plus what it receives in the request, so choosing that information is what makes an application work. Too little context gives weak answers; too much wastes tokens and can overflow the context window. A bigger context window only raises the limit, a fixed tone-and-format instruction is prompt wording rather than information selected per request, and fine-tuning changes the model itself.

1. Recall that a model relies on its request for task-specific facts.
2. Match choosing which information to supply, and how much, to context engineering.
3. Reject options that only raise a limit, only set tone and format, or change the model's weights.

Exam tip: Context engineering means giving the model the right information at the right time, without wasting tokens.

The Model Context Protocol (MCP) is an open standard for connecting AI agents to external systems, tools and data sources.

Instead of building a custom integration for every tool, an MCP server exposes a system in a standard way and any MCP-capable agent can use it. Agent-to-agent messaging, model-weight exchange formats and context compression are all plausible-sounding standards, but none of them is what MCP does: it connects agents to outside tools, data and systems.

1. Recall MCP's role: a standard connection between an agent and external systems.
2. Separate it from standards for agent-to-agent messaging or model file exchange.
3. Do not read 'context' in the name as prompt compression.

Exam tip: MCP is the open standard that connects agents to external tools and data.

Model distillation transfers a large model's task behaviour into a smaller, faster, cheaper model.

The deciding requirement is a compact model that keeps most of the quality. Distillation trains a smaller student on the teacher's responses for this task, cutting latency and per-request cost. Prompt caching is the tempting option because it also lowers cost and latency, but the large model still serves every call. RAG and fine-tuning change what the model knows or how it behaves, not its size.

1. Note the problem: good quality, but too slow and costly.
2. Note the requirement: a compact model with similar task quality.
3. Separate approaches that keep the large model from the one that produces a smaller model.

Exam tip: Want a smaller, cheaper model that behaves like a big one? Distil it.

AI agents let a foundation model take actions across systems by calling tools and orchestrating multi-step tasks.

The new requirement is to do things: create a label, update a record, send an email. A RAG assistant, however well informed, only produces text. An agent adds tools the model can call and orchestration that sequences them, which is the business case for agents. More documents or a vector index improve what the model knows, and fine-tuning changes its style, but none of them lets it act.

1. Separate 'answer questions' (RAG) from 'perform tasks' (agent).
2. Notice the request involves several actions in different systems.
3. Choose the design that adds tools plus orchestration.

Exam tip: Answering from documents is RAG; carrying out multi-step tasks in other systems needs an agent.

Latency, the time a model takes to respond, is the selection criterion for real-time interactive use.

The scenario stresses one thing: users waiting for replies. That maps to latency, which usually favours smaller, faster models. Languages, modality and customization are real criteria, but each answers a different requirement.

1. Find the constraint the scenario stresses: reply speed.
2. Map speed to the latency criterion.
3. Set aside criteria that address languages, data types or adaptation.

Exam tip: Users waiting on a live reply: latency is the deciding selection criterion.

Multi-lingual support is a foundation model selection criterion: the model must handle every language its users write in.

The requirement is to read and answer in English, Spanish and Japanese. A model that lacks one of these languages fails regardless of its speed, size or caching. Model size is the tempting choice because bigger models often cover more, but the criterion to check is the model's stated language support.

1. Identify the requirement: several input and output languages.
2. Map it to the multi-lingual criterion.
3. Do not assume size implies language coverage.

Exam tip: Users in several languages: choose a model by its multi-lingual support.

RAG is usually cheaper and less effort to set up than fine-tuning because it needs no training.

Fine-tuning requires preparing labeled data, running a training job and repeating it when the data changes. RAG adds the catalog through retrieval and leaves the model unchanged; its costs are indexing and storing embeddings. The distractors either claim RAG retrains the model or that both change weights, which is the misconception the item tests.

1. Recall which approach changes weights: only fine-tuning.
2. Link training jobs to higher up-front cost and effort.
3. Choose the statement that RAG adds data without training.

Exam tip: RAG sits below fine-tuning on the cost ladder because it adds data without a training job.

RAG reduces hallucination by grounding the model's answer in retrieved source content.

The model is filling gaps in its knowledge with confident fabrications, which is hallucination. Supplying the relevant handbook passages in the prompt gives it real facts to answer from. A temperature of 0 is the tempting option because it makes answers stable, but stable is not the same as correct. Fine-tuning breaks the no-training constraint, and more output tokens only lengthen the answer.

1. Name the symptom: confident answers not in the source, a hallucination.
2. Ask what gives the model the missing facts at answer time.
3. Reject stability and length settings, and the option that needs training.

Exam tip: Hallucination from missing knowledge is fixed by grounding (RAG), not by lowering temperature.

Maximum output tokens caps how long a generated response can be.

Answers that stop mid-sentence hit a length limit. With the context window ruled out by the stem, the limit is the output cap. Temperature, top-p and top-k all shape which words are chosen, not how many, so they cannot cause truncation.

1. Read the symptom: output cut off mid-sentence.
2. Use the stem to rule out the context window.
3. Pick the parameter that limits output length.

Exam tip: Truncated answers point to the maximum output tokens setting; sampling settings change word choice, not length.

Generated content that reproduces copyrighted training material creates intellectual-property infringement risk.

The images closely copy copyrighted artwork, which is the definition of an IP infringement claim risk. Hallucination is fabricated content, biased output is unequal treatment of groups, and end-user risk is harm to the people using the system; none matches reproduced artwork.

1. Identify what went wrong: output copies someone else's protected work.
2. Map protected-work copying to intellectual property.
3. Rule out fabrication, bias and end-user harm.

Exam tip: Output that copies protected work from training data means intellectual-property infringement risk.

A hallucination is confident, plausible output that is false; acting on it creates legal and trust risk.

The assistant invented a refund policy and stated it confidently, which is a hallucination; because the customer relied on it, the business carries the consequence. IP infringement is copying protected work, biased output is unequal treatment of groups, and prompt injection requires a user manipulating the model.

1. Identify the symptom: a confident statement about something that does not exist.
2. Map fabricated-but-plausible content to hallucination.
3. Rule out copying, bias and attacker-driven manipulation.

Exam tip: Confident, plausible, false output is a hallucination, and the business owns what customers do with it.

An unbalanced dataset produces a model that underperforms for underrepresented groups.

Errors that concentrate on particular demographics point to unequal representation in the training data: the model saw too few examples of some groups. Inconsistent labels and a small dataset would degrade results across the board, and vetted licensed sources concern provenance, not balance.

1. Notice the pattern: errors fall on specific groups, not everyone.
2. Ask which data property affects groups unequally.
3. Choose under-representation (an unbalanced dataset).
4. Reject causes that would lower accuracy for all groups.

Exam tip: Errors concentrated on one group usually trace back to that group being underrepresented in the data.

Bias learned from unbalanced data is fixed at the data: curated, inclusive, balanced representation.

The model learned its skew from data dominated by one group, so the effective fix is adding curated examples until every group is represented. Removing the protected attribute is tempting but leaves proxy features and the imbalance in place; a larger model or more dominant-group data reproduces or amplifies the skew.

1. Locate the root cause: the training data is unbalanced.
2. Prefer fixes that change the data's representation.
3. Recognize that dropping an attribute leaves proxies in correlated features.
4. Choose rebalancing with curated, representative examples.

Exam tip: Fix data-driven bias by rebalancing the data; removing the sensitive column alone does not.

Because model behaviour is probabilistic, input checks alone cannot stop prompt injection; output filtering and validation is the defense-in-depth layer that catches what gets through.

The team needs a control that works even when the attack has already succeeded, so it must act on the response. Output filters (for example, Amazon Bedrock Guardrails on outputs or custom validation) do this. Lower temperature and retraining try to change the model's behaviour but give no guarantee, and TLS protects the network path, not the content shown.

1. Note that input checks are already failing.
2. Look for a control that acts after the model, on the response.
3. Output filtering and validation is the last layer before the user.

Exam tip: Prompt injection defense is layered: filter and validate output, not only input.

RAG grounding retrieves relevant passages from a trusted, up-to-date knowledge base at answer time and has the model answer from them.

Two constraints decide it: answers must reflect documents that change weekly, and there must be no retraining. Retrieval at answer time meets both. Fine-tuning is retraining and lags behind, temperature only changes variability, and few-shot examples cannot carry a whole changing policy set.

1. Note the constraints: current documents, no retraining.
2. Eliminate fine-tuning (retraining) and temperature (not about facts).
3. Choose retrieval at answer time: RAG grounding.

Exam tip: Fresh facts, no retraining: ground answers with RAG.

Confidence scoring attaches a certainty measure to each output so low-confidence answers can be flagged or routed to human review.

The routing decision depends on how certain each answer is, which is what confidence scoring provides, compared against a threshold. Schema validation checks format, grounding improves accuracy without producing a routing signal, and denied topics block subjects.

1. Identify the decision: route by certainty.
2. Find the technique that produces a per-answer certainty value.
3. Confidence scoring plus a threshold routes low-confidence answers to a human.

Exam tip: Route uncertain answers to humans with confidence scoring.