I85PREP
⌘ K
EXIN BCS Gen AI Award study materials
EXIN BCS Gen AI Award study resources

EXIN BCS Gen AI Award practice questions

One question at a time. Build your EXIN BCS Gen AI Award understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Generative AI fundamentals, capabilities and limitations

Which statement best describes what a generative AI model produces when given a prompt?

  1. A verbatim copy of the closest stored training document
  2. A live web search result returned without any model processing
  3. A deterministic lookup of a pre-written answer keyed to the prompt
  4. New output inferred from patterns learned during training Correct answer.

Explanation

Generative models learn statistical patterns from training data and use them to infer new output at inference time. The output is synthesised rather than retrieved verbatim from a store, which is why it can be fluent yet factually wrong.

Question 2Generative AI fundamentals, capabilities and limitations

What is a 'token' in the context of a large language model?

  1. A unit of text, such as a word fragment, that the model processes Correct answer.
  2. A numeric weight stored inside a neural network layer
  3. A single character that is always exactly one letter long
  4. A security credential used to authenticate API requests

Explanation

Tokenisation splits input text into subword units that the model converts into numeric representations. Tokens can be whole words, word fragments or punctuation, and token counts drive context limits and pricing.

Question 3Generative AI fundamentals, capabilities and limitations

Which capability is characteristic of a text-to-image generative model?

  1. Producing a new image from a written description Correct answer.
  2. Translating an image into a database schema
  3. Compressing an image file without any loss
  4. Measuring the pixel dimensions of an uploaded image

Explanation

Text-to-image models map a natural-language prompt into an image that matches the described content. They synthesise pixels rather than performing file operations or structural analysis.

Question 4Generative AI fundamentals, capabilities and limitations

What does the 'context window' of a large language model limit?

  1. The amount of text the model can consider in a single interaction Correct answer.
  2. The number of users who may access the model per day
  3. The number of languages the model was trained on
  4. The physical memory installed on the host server

Explanation

The context window is the maximum number of tokens the model can attend to at once, covering the prompt plus generated output. Content beyond that limit must be truncated, summarised or retrieved in stages.

Question 5Generative AI fundamentals, capabilities and limitations

A model confidently states a plausible but invented citation. Which term describes this behaviour?

  1. Hallucination Correct answer.
  2. Quantisation
  3. Overfitting
  4. Tokenisation

Explanation

Hallucination is when a generative model produces fluent content that is fabricated or unsupported by evidence. It arises because the model predicts likely text rather than verifying facts against a source.

Question 6Generative AI fundamentals, capabilities and limitations

Why is human review recommended before publishing generative AI output in a business document?

  1. Output is always unreadable without manual rewriting
  2. Output cannot be edited once it has been generated
  3. Output is legally required to be rewritten by a manager
  4. Output can be fluent yet contain errors that need verification Correct answer.

Explanation

Generative output can read convincingly while containing factual mistakes, bias or unsuitable tone. Human review checks accuracy and appropriateness before the content reaches an audience.

Question 7Generative AI fundamentals, capabilities and limitations

What is 'prompt engineering' in generative AI practice?

  1. Encrypting prompts so users cannot read them
  2. Designing and refining instructions to steer model output Correct answer.
  3. Building the GPU hardware that trains the model
  4. Compiling model weights into executable machine code

Explanation

Prompt engineering is the deliberate crafting of instructions, context and examples to obtain more useful and reliable output. It works at the input level and requires no change to model weights.

Question 8Generative AI fundamentals, capabilities and limitations

Which characteristic distinguishes generative AI from a traditional rules-based classifier?

  1. It cannot process natural-language input
  2. It requires no training data of any kind
  3. It always produces numerically exact results
  4. It can create novel content rather than only assign a label Correct answer.

Explanation

A rules-based classifier maps input to a predefined category, whereas a generative model produces new content such as text, images or code. Both may use training data, but only the generative model synthesises output.

Question 9Generative AI fundamentals, capabilities and limitations

A team wants a summary of a 400-page report that exceeds the model's context window. Which approach fits the model's limitation?

  1. Paste the entire report and rely on the model to ignore the overflow
  2. Ask the model to raise its own context window before summarising
  3. Summarise sections in stages and combine the partial summaries Correct answer.
  4. Disable tokenisation so the whole document fits as one token

Explanation

Chunking the document and summarising iteratively keeps each request within the context window. The partial summaries are then combined, which is a standard technique for long documents.

Question 10Generative AI fundamentals, capabilities and limitations

Which prompt is most likely to reduce vague output from a generative model?

  1. Write something useful about the document
  2. Summarise this contract in five bullet points for a non-lawyer reader Correct answer.
  3. Tell me about this contract
  4. Give me everything important here

Explanation

Specifying format, length and audience gives the model concrete constraints to satisfy. Vague prompts leave the model to guess intent, which increases variability and irrelevant content.

Question 11Generative AI fundamentals, capabilities and limitations

Which risk is most directly associated with a model trained on biased historical data?

  1. Reduced electricity consumption during training
  2. Faster inference on consumer hardware
  3. Output that reproduces or amplifies social stereotypes Correct answer.
  4. Guaranteed factual accuracy on current events

Explanation

A model learns the patterns present in its training data, so historical bias can surface in generated content. Mitigation requires data curation, evaluation and human review rather than assuming neutrality.

Question 12Generative AI fundamentals, capabilities and limitations

Why can a large language model give an outdated answer about a recent event?

  1. It deliberately withholds recent facts to save tokens
  2. Its knowledge is fixed at training time and may lack newer information Correct answer.
  3. It can only answer questions phrased in the past tense
  4. It deletes its training data after each conversation

Explanation

A model's parametric knowledge reflects its training data up to a cut-off date. Events after that point are unknown unless the system retrieves fresh information from an external source.

Question 13Generative AI fundamentals, capabilities and limitations

What does 'retrieval-augmented generation' add to a language model system?

  1. An automatic guarantee that every claim is correct
  2. A larger number of trainable parameters in the base model
  3. A compression layer that shrinks the model file size
  4. Relevant documents fetched and supplied as context at query time Correct answer.

Explanation

RAG retrieves passages from an external knowledge source and inserts them into the prompt so the model can ground its answer. This improves currency and traceability but does not guarantee correctness.

Question 14Generative AI fundamentals, capabilities and limitations

A model produces different wording each time the same prompt is submitted. What explains this?

  1. The model is retraining itself between requests
  2. Sampling introduces randomness into token selection Correct answer.
  3. The prompt is being silently rewritten by the user interface
  4. Tokenisation changes the meaning of the prompt each time

Explanation

Generation samples from a probability distribution over next tokens, and settings such as temperature control that randomness. Identical prompts can therefore yield varied but related outputs.

Question 15Generative AI fundamentals, capabilities and limitations

Which task is generative AI generally least suited to performing reliably on its own?

  1. Drafting a first-pass product description
  2. Suggesting alternative phrasings for a sentence
  3. Certifying that a financial statement complies with auditing standards Correct answer.
  4. Producing boilerplate code from a clear specification

Explanation

Formal certification demands accountability, verified evidence and professional judgement that a probabilistic model cannot supply. Drafting, paraphrasing and code generation are assistive tasks where human review remains feasible.

Question 16Generative AI fundamentals, capabilities and limitations

What is the primary purpose of a system prompt in a deployed generative AI assistant?

  1. To store the user's password for later sessions
  2. To set persistent instructions and constraints for the interaction Correct answer.
  3. To replace the need for any evaluation of output
  4. To increase the model's parameter count at runtime

Explanation

A system prompt establishes role, tone and boundaries that apply across the conversation. It guides behaviour but does not remove the need for testing and oversight.

Question 17Generative AI fundamentals, capabilities and limitations

An organisation must process confidential client data with a generative AI service. Which consideration is most relevant?

  1. Whether the interface offers a dark colour theme
  2. Whether the model uses a transformer architecture
  3. Whether the provider retains or trains on submitted data Correct answer.
  4. Whether the model can also generate images

Explanation

Data handling terms determine whether confidential input may be stored, reviewed or used for training. Architecture, theming and modality are not the primary confidentiality controls.

Question 18Generative AI fundamentals, capabilities and limitations

What is 'temperature' controlling in a language model's generation settings?

  1. The degree of randomness in the next-token selection Correct answer.
  2. The language in which the answer is written
  3. The physical heat of the data-centre processors
  4. The number of tokens reserved for the response

Explanation

Temperature scales the probability distribution used when sampling, so higher values yield more varied output and lower values yield more predictable output. It does not set response length or language.

Question 19Generative AI fundamentals, capabilities and limitations

Which statement about multimodal generative models is accurate?

  1. They can accept or produce more than one data type, such as text and images Correct answer.
  2. They can only process text and never accept images
  3. They require separate models for every single language
  4. They eliminate the need for prompt design entirely

Explanation

Multimodal models handle combinations of modalities, for example accepting an image and text prompt and returning text. This extends use cases but does not remove the need for clear instructions and review.

Question 20Generative AI fundamentals, capabilities and limitations

A developer needs the same deterministic response to a fixed prompt for automated testing. Which setting is most appropriate?

  1. Set temperature to its maximum value
  2. Add several creative examples to the system prompt
  3. Increase the maximum response length substantially
  4. Set temperature to zero and fix the random seed where supported Correct answer.

Explanation

Reducing temperature to zero makes the model favour the highest-probability token, and fixing the seed removes sampling variation where the API allows it. This yields far more reproducible output for tests.

Question 21Generative AI fundamentals, capabilities and limitations

Why does simply adding more retrieved documents to a RAG prompt sometimes reduce answer quality?

  1. Documents are converted into new training examples at query time
  2. Retrieval always corrupts the model's internal weights
  3. The model can only read a single sentence per request
  4. Irrelevant context can distract the model and dilute the useful evidence Correct answer.

Explanation

Stuffing the context with loosely related passages increases the chance the model attends to noise and loses the key evidence. Retrieval quality and relevance ranking matter as much as quantity.

Question 22Generative AI fundamentals, capabilities and limitations

A generative model performs excellently on its training data but poorly on new inputs. Which explanation fits best?

  1. It has overfitted and generalises weakly to unseen examples Correct answer.
  2. Its context window has been set too large
  3. Its tokeniser has been disabled during inference
  4. It has too few parameters to memorise anything

Explanation

Overfitting occurs when a model captures noise specific to the training set instead of general patterns. The symptom is strong training performance combined with weak performance on unseen data.

Question 23Generative AI fundamentals, capabilities and limitations

Which control most directly reduces the risk of prompt injection from untrusted retrieved content?

  1. Allow the model to execute any tool the retrieved text requests
  2. Treat retrieved text as data and restrict the actions the model may take Correct answer.
  3. Increase the model temperature for more creative responses
  4. Shorten the system prompt to a single word

Explanation

Prompt injection exploits instructions hidden in untrusted content, so retrieved text should be handled as data rather than trusted commands. Limiting tool permissions and validating actions contain the impact if injection succeeds.

Question 24Generative AI fundamentals, capabilities and limitations

An organisation wants evidence that a generative AI feature meets quality expectations before release. Which approach is most defensible?

  1. Rely on the vendor's demonstration examples alone
  2. Release to all users and monitor complaints afterwards
  3. Evaluate against a curated test set with defined criteria and human review Correct answer.
  4. Ask the model whether its own answers are correct

Explanation

A curated evaluation set with explicit criteria provides repeatable evidence of quality and surfaces failure modes. Human review is needed because automated self-assessment by the model is unreliable.

Question 25Generative AI fundamentals, capabilities and limitations

Which limitation is inherent to using an LLM as a factual knowledge base without external sources?

  1. It refuses to answer any question about science
  2. It can only respond in a single fixed language
  3. It stores every answer it has ever produced
  4. It cannot verify its claims against authoritative evidence Correct answer.

Explanation

A model generates likely text rather than checking a source of truth, so plausible errors can pass undetected. Grounding in authoritative sources and verification are needed for factual reliability.

Question 26Generative AI fundamentals, capabilities and limitations

In generative AI terminology, what does the term 'foundation model' refer to?

  1. A large model pre-trained on broad data that can be adapted to many downstream tasks Correct answer.
  2. A hardware accelerator optimised for matrix multiplication
  3. A rules engine that maps fixed inputs to fixed outputs
  4. A small model trained only on one company's proprietary spreadsheet data

Explanation

A foundation model is pre-trained at scale on broad, general data and then adapted (via prompting, fine-tuning or RAG) to many downstream tasks. This adaptability is what distinguishes it from narrow, single-purpose systems.

Question 27Generative AI fundamentals, capabilities and limitations

Which statement best describes a 'token' in large language model processing?

  1. A complete sentence delimited by punctuation
  2. A permission credential used to authenticate API calls
  3. A single character that always equals one byte of UTF-8 storage
  4. A unit of text, often a word fragment, that the model encodes as a number Correct answer.

Explanation

Tokenisation splits input text into subword units, and each token maps to an integer ID in the model's vocabulary. Token counts drive context limits and pricing, which is why they matter operationally.

Question 28Generative AI fundamentals, capabilities and limitations

What is the 'context window' of a generative language model?

  1. The time period during which the model's training data was collected
  2. The graphical interface through which users submit prompts
  3. The maximum amount of text, in tokens, the model can consider in a single request Correct answer.
  4. The number of users the service can serve concurrently

Explanation

The context window bounds how many tokens (prompt plus generated output) the model can attend to at once. Content beyond that limit cannot influence the response unless it is summarised, chunked or retrieved.

Question 29Generative AI fundamentals, capabilities and limitations

Which capability is the defining feature of a generative model compared with a discriminative classifier?

  1. It always achieves higher accuracy on classification benchmarks
  2. It requires no training data to operate
  3. It cannot be deployed through an API
  4. It produces new content such as text, images or audio rather than only a label Correct answer.

Explanation

Generative models learn the distribution of the training data and sample from it to create new outputs. A discriminative classifier instead learns a decision boundary and returns a class label or score.

Question 30Generative AI fundamentals, capabilities and limitations

What does 'inference' mean when a generative AI service is running in production?

  1. The audit of training data for copyright compliance
  2. The conversion of the model into a compressed file format
  3. The process of updating the model's weights from new labelled examples
  4. The model generating an output from a given input using its trained weights Correct answer.

Explanation

Inference is the forward pass in which a trained model computes an output for a supplied input. Weight updates belong to training, which is a separate lifecycle stage.

Question 31Generative AI fundamentals, capabilities and limitations

Which factor most directly explains why a generative model can produce fluent, plausible text that is nonetheless factually wrong?

  1. It deliberately hides the sources it used
  2. It optimises for statistically likely token sequences rather than verified truth Correct answer.
  3. It has no vocabulary for factual topics
  4. It only outputs text that was copied verbatim from the training set

Explanation

Language models are trained to predict likely continuations, so fluency and plausibility are optimised while factual correctness is not directly guaranteed. This mismatch is the root of hallucination.

Question 32Generative AI fundamentals, capabilities and limitations

What is 'fine-tuning' in the context of generative AI?

  1. Further training of a pre-trained model on task-specific examples to adapt its behaviour Correct answer.
  2. Increasing the model's temperature to make answers more creative
  3. Reducing the number of tokens in the user's prompt
  4. Adding a search index in front of the model at query time

Explanation

Fine-tuning continues training on a smaller, targeted dataset so the model's weights shift toward a desired style, format or domain. It changes the model itself, unlike retrieval which changes the input at runtime.

Question 33Generative AI fundamentals, capabilities and limitations

Which statement correctly describes the relationship between a base model and an instruction-tuned model?

  1. The base model has more parameters because tuning adds none
  2. The base model is always safer to expose directly to end users
  3. The instruction-tuned model has been further trained to follow user directions more reliably Correct answer.
  4. The instruction-tuned model cannot generate free-form text

Explanation

Instruction tuning (often with human feedback) teaches a pre-trained base model to follow requests and adopt helpful response formats. The underlying architecture and parameter count are typically unchanged.

Question 34Generative AI fundamentals, capabilities and limitations

A marketing team asks a generative assistant to draft product descriptions for 500 items and wants consistent structure across all of them. Which approach best supports this?

  1. Provide a template with fixed sections and one worked example in the prompt Correct answer.
  2. Remove all formatting instructions so the model decides the layout
  3. Ask each item's description to be written by a different team member manually
  4. Increase temperature to maximum so each description is unique

Explanation

A template plus a worked example constrains the output structure and demonstrates the expected format, which is a standard few-shot prompting technique. Consistency comes from explicit constraints, not from randomness.

Question 35Generative AI fundamentals, capabilities and limitations

A user needs a generative model to extract dates from invoices into a strict JSON schema. Which technique most reliably improves schema compliance?

  1. Requesting the answer in a long narrative paragraph
  2. Asking politely for a well-formatted answer without showing the schema
  3. Supplying the exact JSON schema in the prompt and requesting only JSON in the response Correct answer.
  4. Setting temperature to its maximum value

Explanation

Showing the exact target schema and constraining output to JSON gives the model an explicit contract to satisfy. Many providers also offer structured-output modes that enforce the schema mechanically.

Question 36Generative AI fundamentals, capabilities and limitations

An organisation wants the model to answer questions about internal policy documents that change weekly. Which architecture is most appropriate?

  1. Retrieval-augmented generation over a regularly refreshed document index Correct answer.
  2. Fine-tuning the model once on last year's policy manual
  3. Relying on the model's pre-trained knowledge of general policy
  4. Increasing the context window and pasting the entire archive every time

Explanation

RAG keeps knowledge external and refreshable, so updated documents take effect as soon as the index is rebuilt. Fine-tuning bakes knowledge into weights and becomes stale quickly.

Question 37Generative AI fundamentals, capabilities and limitations

What is the main function of an embedding vector in a generative AI application?

  1. It encrypts user prompts before transmission
  2. It stores the model's weight matrices on disk
  3. It represents text as numbers so semantic similarity can be computed Correct answer.
  4. It records the billing cost of each API call

Explanation

Embeddings map text into a high-dimensional space where semantically similar items sit close together, enabling search, clustering and retrieval. They are the backbone of vector-based retrieval in RAG.

Question 38Generative AI fundamentals, capabilities and limitations

A team notices that their assistant answers questions about a niche legal regulation incorrectly even when the correct text is supplied in the prompt. Which explanation is most plausible?

  1. The model may be over-weighting its parametric memory over the supplied passage Correct answer.
  2. The model has no decoder component
  3. The prompt length exceeded one token
  4. The embedding index was rebuilt overnight

Explanation

Models can sometimes favour patterns learned in training over the provided context, especially when the context conflicts with memorised associations. Prompt design, ordering and explicit instruction to cite the passage can mitigate this.

Question 39Generative AI fundamentals, capabilities and limitations

Which scenario best illustrates the concept of 'model drift' in a deployed generative AI system?

  1. The API returns a rate-limit error during peak hours
  2. The model file is accidentally deleted from the server
  3. A user submits a prompt containing a spelling mistake
  4. Answer quality degrades as real-world language and user needs shift away from training data Correct answer.

Explanation

Model drift is the gradual mismatch between a static model and a changing environment or data distribution. Monitoring quality over time and refreshing models or retrieval sources are standard mitigations.

Question 40Generative AI fundamentals, capabilities and limitations

Which statement about the knowledge cutoff of a large language model is accurate?

  1. It marks when the model stops accepting new prompts
  2. It reflects the latest point covered by training data and bounds what the model knows natively Correct answer.
  3. It is the date the model licence expires
  4. It is the date the model was last fine-tuned by the user

Explanation

The knowledge cutoff is the temporal boundary of the training corpus; events after it are unknown unless supplied through retrieval or tools. This is why current-events questions often need external sources.

Question 41Generative AI fundamentals, capabilities and limitations

What is 'few-shot prompting'?

  1. Limiting the model to a few output tokens
  2. Running the model on a few GPUs to save cost
  3. Training the model on a small dataset before deployment
  4. Including a small number of worked examples in the prompt to guide the response pattern Correct answer.

Explanation

Few-shot prompting places example input-output pairs inside the prompt so the model infers the desired pattern at inference time. No weights are changed, which distinguishes it from fine-tuning.

Question 42Generative AI fundamentals, capabilities and limitations

Why is a generative model's output sometimes described as 'non-deterministic' by default?

  1. The underlying hardware produces random bit errors
  2. Sampling from a probability distribution introduces variation between runs Correct answer.
  3. The API randomises the user's prompt before sending
  4. The model re-trains itself on every request

Explanation

Decoders sample from the predicted token distribution, so identical prompts can yield different sequences unless sampling is constrained. Setting temperature to zero or using a fixed seed reduces this variation.

Question 43Generative AI fundamentals, capabilities and limitations

A bank wants to use a generative assistant to draft replies to customer complaints. Which control is most directly relevant to managing regulatory risk?

  1. Removing customer names from the prompt to speed processing
  2. Logging prompts and outputs and requiring human sign-off before sending Correct answer.
  3. Allowing the model to send replies automatically to save time
  4. Increasing the model's temperature for more empathetic wording

Explanation

Auditability and human oversight create the evidence trail regulators expect and prevent unreviewed automated commitments. Logging supports investigation, while sign-off keeps accountability with a person.

Question 44Generative AI fundamentals, capabilities and limitations

Which limitation is most characteristic of using a generative model to perform arithmetic on large numbers without tool access?

  1. It can only handle numbers below ten
  2. It may produce plausible but incorrect results because it predicts tokens rather than computing Correct answer.
  3. It always returns the correct answer but slowly
  4. It refuses all numeric questions by design

Explanation

Arithmetic requires exact symbolic computation, whereas token prediction approximates. This is why reliable systems route calculations to a calculator or code tool.

Question 45Generative AI fundamentals, capabilities and limitations

What does the term 'multimodal' mean for a generative model?

  1. It serves multiple tenants from one endpoint
  2. It supports multiple programming languages
  3. It can accept or produce more than one data type, such as text and images Correct answer.
  4. It can be deployed across multiple cloud regions

Explanation

Modality refers to the type of data, such as text, image or audio. A multimodal model processes or generates across several such types within one system.

Question 46Generative AI fundamentals, capabilities and limitations

Which statement best captures the difference between a hallucination and a deliberate simplification?

  1. A hallucination only occurs in image models
  2. A hallucination is always longer than a simplification
  3. A simplification is always factually wrong
  4. A hallucination is an unsupported fabrication presented as fact, whereas a simplification conveys a true idea less precisely Correct answer.

Explanation

Hallucination is the generation of content with no grounding in fact or source, presented confidently. Simplification preserves truth while reducing detail, so it is a communication choice rather than an error.

Question 47Generative AI fundamentals, capabilities and limitations

An engineer wants to reduce the cost of a RAG system that sends ten long documents with every query. Which change is most directly effective?

  1. Send the documents in a random order
  2. Increase the model's temperature
  3. Retrieve and send only the top-ranked passages that fit a token budget Correct answer.
  4. Duplicate each document to reinforce the signal

Explanation

Token cost scales with prompt length, so selecting fewer, more relevant passages lowers cost while often improving answer focus. Ranking and budgeting retrieved chunks is standard practice.

Question 48Generative AI fundamentals, capabilities and limitations

Which risk arises when a generative model is used to summarise user-supplied documents that may contain hidden instructions?

  1. Licence violation in the model's weights
  2. Token overflow in the tokeniser vocabulary
  3. Embedding drift caused by vector normalisation
  4. Indirect prompt injection that redirects the model's behaviour Correct answer.

Explanation

Untrusted content can carry instructions the model may follow, which is indirect prompt injection. Defences include treating retrieved text strictly as data and applying output filtering.

Question 49Generative AI fundamentals, capabilities and limitations

A team must demonstrate that a generative feature meets quality expectations before launch. Which approach is most appropriate?

  1. Deploy to all users and watch for complaints
  2. Build a labelled evaluation set and measure accuracy and groundedness against it Correct answer.
  3. Ask the developers whether the outputs look good
  4. Compare the number of parameters with a competitor's model

Explanation

A representative labelled evaluation set gives repeatable, quantitative evidence of quality, including whether answers are grounded in sources. Informal impressions and post-launch complaints are not pre-release evidence.

Question 50Generative AI fundamentals, capabilities and limitations

Which statement about the relationship between model size and capability is most defensible?

  1. Capability is determined solely by the number of GPUs used
  2. Size alone guarantees factual accuracy
  3. Smaller models cannot generate coherent text
  4. Larger models often perform better on complex tasks, but gains vary and cost rises Correct answer.

Explanation

Scaling generally improves capability on many benchmarks, but returns diminish and inference cost grows. Task fit, data quality and tuning also matter, so size is not a guarantee of correctness.