Generative AI and Large Language Models practice questions

Worked examples for Generative AI and Large Language Models

1. Question 1

   Which TWO statements about the characteristic risks of generative AI are accurate?

   1. A. Hallucination is fully eliminated once a model is trained with enough parameters and enough data.

      The 'scale solves hallucination' misconception. Larger models still generate probabilistically and can produce confident fabrications.

   2. B. Because hallucinated output is typically fluent and self-consistent, plausibility is not evidence that the content is correct.Correct answer

      Correct. The defining hazard of hallucination is that fabricated content reads as authoritative, which is why Oracle stresses verifying generated output.

   3. C. Generated output is automatically free of intellectual-property concerns because the content did not previously exist.

      The 'new content is inherently clean' misconception. Output learned from copyrighted training material can resemble it, leaving ownership and licensing questions open.

   4. D. A hallucination manifests as the model returning an error or failing to respond to the prompt.

      Confuses a service failure with a content failure. Hallucination occurs in responses that complete normally but contain false information.

   5. E. A model can reproduce and amplify biases contained in the data it was trained on.Correct answer

      Correct. Oracle lists bias among generative AI's core risks, arising because models learn from large real-world corpora that carry societal bias.

   Explanation

   Two properties define generative AI's characteristic risk profile: fabricated content arrives dressed in fluent, confident prose, so readability is no proxy for truth; and biases present in the training corpus are learned and can be amplified in output. Scaling a model up does not remove probabilistic generation and therefore does not remove hallucination, novelty of output does not settle intellectual-property questions when the model learned from protected works, and a failed or errored request is an availability problem rather than a hallucination.

2. Question 2

   A team summarizes long legal documents with an LLM. For very long documents the request is rejected or the earliest sections appear to be ignored. Which concept explains this behavior, and what does moving to a model with a larger context window actually change?

   1. A. The context window is the number of embeddings the model stores permanently; enlarging it lets the model remember all past conversations indefinitely.

      Confuses the context window with persistent memory and counts embeddings rather than tokens. The window is a per-request input limit, not a store that survives between requests.

   2. B. The context window is the model's parameter count; enlarging it means the model was trained on proportionally more data.

      Conflates context window with model size. Parameter count is fixed by training and describes model capacity; the context window describes how much input a single request may contain.

   3. C. The context window is a throughput cap on tokens generated per second; a larger window mainly makes long documents process faster.

      Confuses a capacity limit with a performance limit. The window bounds how much text can be considered at once, not how quickly tokens are produced.

   4. D. The context window is the maximum number of tokens the model can take into account in one request — prompt plus generated response — so a larger window lets more of the document be considered at once, though nothing is retained after the request unless it is supplied again.Correct answer

      Correctly states the concept as Oracle describes it: the context window bounds the tokens a model can attend to for a single request, and each request is stateless, so a bigger window increases how much input fits, not how much the model permanently remembers.

   Explanation

   The context window is the token budget for one interaction, covering both the supplied prompt and the response the model generates; when a document exceeds it, the input must be truncated or chunked, which is why early sections seem to disappear. A larger window raises how much text the model can weigh in a single pass, but it does not create memory across requests, does not describe parameter count or training-data volume, and is a capacity limit rather than a speed or throughput setting.

3. Question 3

   A medical transcription company needs an LLM to consistently produce output in a highly specialized clinical note structure and to use domain terminology the base model handles poorly. They have tried detailed prompts with examples, but quality is still unacceptable, and they possess tens of thousands of professionally written example notes. Which approach should they adopt?

   1. A. Build a RAG pipeline that retrieves similar past notes and appends them to every prompt

      Applies RAG to a behavior problem rather than a knowledge problem. Retrieval supplies facts the model lacks; here the model has the dictated content and needs to internalize a writing style, which retrieval does not durably teach.

   2. B. Fine-tune the model on the corpus of example notes so the specialized structure and terminology are learned into its weightsCorrect answer

      Correct. Fine-tuning is the appropriate escalation when prompting has been tried and fallen short and a large, high-quality task-specific dataset exists — it teaches durable style, format, and domain behavior.

   3. C. Add more examples to the prompt until the entire corpus of notes has been included in the context

      Assumes few-shot scales without limit. Context windows are finite and per-request cost and latency grow with prompt length, so tens of thousands of examples cannot be prompted — that volume is exactly what training is for.

   4. D. Raise the temperature so the model explores more varied phrasings until it matches the clinical style

      Confuses a randomness setting with task adaptation. Higher temperature produces more varied output, which works against the consistency the company needs and teaches the model nothing about the target format.

   Explanation

   Fine-tuning is the right choice when the goal is to change how a model behaves — its style, structure, and command of specialized terminology — prompting has already been attempted without success, and a substantial labeled dataset is available to train on. Retrieval-augmented generation targets a different problem, supplying missing facts at query time rather than reshaping generation behavior. Packing an entire corpus into the prompt is infeasible given finite context windows and per-token cost, and adjusting sampling randomness changes variability, not learned task competence.

4. Question 4

   Which statement best describes what prompt engineering is and what it does to the underlying model?

   1. A. It is a lightweight training procedure that adjusts a subset of the model's weights using the text of the prompt

      Represents the misconception that prompting changes the model's weights. Prompt text is consumed as input tokens; no training occurs, which is why prompt changes take effect instantly and cost nothing to deploy.

   2. B. It is the practice of crafting the instructions, context, and examples given to a model at inference time in order to steer its output, without altering the model's parametersCorrect answer

      Correct. Prompt engineering operates entirely on the input side: instructions, role framing, constraints, and demonstrations shape the response for that request while the model itself remains unchanged.

   3. C. It is the process of indexing an organization's documents so the model can search them during generation

      Describes building a retrieval pipeline. Indexing documents for query-time search is RAG; prompt engineering concerns the wording and structure of what is sent to the model.

   4. D. It is the selection of the training data mixture used to pretrain a foundation model

      Confuses prompting with pretraining data curation, which happens long before the model is deployed and has nothing to do with per-request input design.

   Explanation

   Prompt engineering is the discipline of designing what goes into the model's context — task instructions, relevant background, constraints, and optionally worked examples — so the generated output matches the desired intent. It acts purely at inference time and leaves model parameters untouched, which is what makes it the fastest and cheapest customization lever available. Adjusting weights is fine-tuning, indexing documents for query-time lookup is retrieval augmentation, and choosing a training corpus is pretraining data curation.

5. Question 5

   A team is deciding how to adapt an LLM for a new internal use case. Which principle best guides the order in which they should evaluate prompt engineering, retrieval-augmented generation, and fine-tuning?

   1. A. Always fine-tune first, because a fine-tuned model makes prompt engineering and retrieval unnecessary

      Represents the assumption that fine-tuning subsumes the other techniques. A fine-tuned model still needs good prompts and still cannot know facts outside its training data, so retrieval often remains necessary alongside it.

   2. B. Choose based solely on the size of the foundation model, since small models need retrieval and large models need fine-tuning

      Invents a rule based on model scale. The appropriate technique is determined by the nature of the gap between current and desired output, not by parameter count.

   3. C. The techniques are mutually exclusive, so an application must commit to exactly one of them

      Treats the approaches as alternatives rather than complements. In practice a production system commonly combines a carefully engineered prompt, retrieval for grounding, and sometimes a fine-tuned model.

   4. D. Start with the lowest-cost technique that could plausibly work — prompt engineering — then add retrieval when the gap is missing knowledge, and fine-tune when the gap is durable behavior that prompting cannot sustainCorrect answer

      Correct. The techniques form an escalating ladder of cost and complexity, and the right one is determined by whether the deficiency is instruction/format (prompting), access to data (RAG), or ingrained model behavior (fine-tuning).

   Explanation

   Adapting an LLM is a diagnosis problem: identify whether the shortfall is unclear instruction or format, absent knowledge, or unreliable behavior, and then apply the cheapest technique that addresses it — prompt engineering, then retrieval, then fine-tuning. Fine-tuning does not eliminate the need for good prompts or for grounding in current data, model size does not dictate the choice, and the three techniques compose rather than compete, with many production systems using all of them together.

6. Question 6

   A team is explaining to new engineers how an LLM turns a user's sentence into something it can compute on. Which statement correctly distinguishes a token from an embedding?

   1. A. A token and an embedding are two names for the same thing — the model's numeric representation of a piece of text.

      The classic token/embedding conflation. They are different stages: a token is a discrete chunk of text, while an embedding is the dense numeric vector the model derives from it.

   2. B. A token is a chunk of text such as a word or sub-word unit; an embedding is the numeric vector that represents that token's meaning to the model.Correct answer

      Correct: text is first split into tokens (words or sub-word pieces), and each token is mapped to a vector embedding that positions it in a semantic space the model can compute with (Oracle, 'What is a Large Language Model (LLM)?').

   3. C. A token is the vector of numbers the model computes internally; an embedding is the raw sequence of characters the user typed.

      Inverts the two definitions. The raw characters are what tokenization consumes, and the computed vector is the embedding — not the other way around.

   4. D. A token is produced only during training, while an embedding is produced only at inference time when a user submits a prompt.

      Invents a training-versus-inference split that does not exist. Both tokenization and embedding happen in both phases — every prompt is tokenized and embedded at inference just as training text was.

   Explanation

   Text is processed in two stages: tokenization breaks the input into tokens — whole words or sub-word fragments drawn from a fixed vocabulary — and each token is then mapped to an embedding, a dense vector that encodes meaning in a form the network can operate on. Treating the two as synonyms erases that pipeline, swapping their definitions reverses it, and neither step is exclusive to training or to inference since prompts go through the identical path.

7. Question 7

   A media company plans to publish blog posts and stock artwork produced by a generative AI service, and its engineers routinely paste excerpts of unreleased internal source code into a public chat assistant to get refactoring suggestions. Which risk category of generative AI does this combined practice most directly implicate?

   1. A. Hallucination, since generated blog posts and images may contain factual errors stated with high confidence.

      Hallucination is a real risk but describes factual incorrectness in the output. It says nothing about who owns the generated artwork or what happens to proprietary code sent to a third-party service.

   2. B. Model bias, since generated text and imagery may reflect skewed patterns present in the training corpus.

      Bias concerns unfair or unrepresentative output. It is orthogonal to intellectual-property ownership of generated content and to the confidentiality exposure of pasting internal code into an external system.

   3. C. Intellectual-property and data-privacy exposure — uncertain rights over generated output that may echo copyrighted training material, plus leakage of confidential data submitted to the model.Correct answer

      Oracle groups IP/copyright ambiguity over generated artifacts and the exposure of sensitive or proprietary data supplied to a model among generative AI's principal risks; both halves of the scenario fall under it.

   4. D. No meaningful risk, because content a model generates is by definition original and therefore automatically owned by the party that issued the prompt.

      Misconception that generative output is automatically free of third-party rights and automatically owned by the prompter. Oracle explicitly flags ownership and copyright of generated content as unsettled and a reason for governance.

   Explanation

   Oracle identifies intellectual-property uncertainty and data privacy among the defining risks of generative AI: models trained on large corpora may produce output resembling copyrighted material, ownership of generated artifacts is legally unsettled, and any confidential data placed in a prompt to an external service may be retained or exposed. The scenario combines both halves of that risk — publishing generated artifacts whose rights are unclear, and transmitting proprietary code outside the organization. Hallucination and bias are genuine risks in Oracle's list but describe the accuracy and fairness of output, neither of which is what this practice puts at stake, and treating generated content as automatically original and owned by the prompter is precisely the assumption Oracle warns against.

8. Question 8

   A hiring team notices that a generative AI tool asked to write candidate summaries consistently uses more assertive, leadership-oriented language for some groups of applicants than for others, even when the underlying resumes are comparable. Which characteristic risk of generative AI does this illustrate, and what is its usual origin?

   1. A. Hallucination, because the model invented leadership qualities that the resumes did not state.

      Confuses bias with hallucination. Hallucination is fabricating specific facts; the defect described is a systematic difference in tone applied across groups, which is unfairness rather than invented content.

   2. B. Bias, arising from skewed or unrepresentative patterns in the data the model learned from, which the model then reproduces in its output.Correct answer

      Correct: bias in generative AI is systematically skewed or unfair output that reflects imbalances and historical patterns present in the training data.

   3. C. Intellectual-property leakage, because the model reused phrasing it memorized from copyrighted job postings.

      Substitutes the IP/data risk for the fairness risk. Nothing in the scenario concerns proprietary content being reproduced or exposed; the issue is differential treatment across applicant groups.

   4. D. Bias, arising solely from the temperature setting used at inference time, which can be eliminated by lowering it.

      Names the right risk but the wrong cause. Decoding parameters control randomness and variability of wording, not the learned associations that produce systematically unfair output.

   Explanation

   Bias is one of the characteristic risks of generative AI: because models learn statistical patterns from large training corpora, imbalances and historical prejudices in that data can be reproduced and even amplified in generated content, producing systematically unfair treatment of some groups. Mitigation involves data curation, evaluation for fairness, and human oversight rather than a decoding-parameter change. Fabricating specific unstated facts would be hallucination, reproducing or exposing proprietary content would be the data/IP risk, and inference-time randomness settings affect variability of wording, not learned associations.