In a GAN, which network learns to create candidate samples?
A Discriminator
B Generator
C Encoder
D Classifier
Answer: B. Generator. The generator produces candidate data. During training, its outputs are assessed by the discriminator.
What does the discriminator do during GAN training?
A Adds noise to every sample
B Encodes text prompts
C Stores model checkpoints
D Distinguishes real from generated
Answer: D. Distinguishes real from generated. The discriminator learns to distinguish real training examples from samples produced by the generator.
Why are a GAN's generator and discriminator trained against each other?
A Improve samples through feedback
B Remove all training examples
C Guarantee perfect outputs
D Avoid evaluating generated data
Answer: A. Improve samples through feedback. Their opposing objectives provide feedback: the generator tries to fool the discriminator, while the discriminator tries to identify generated samples.
Which GAN training failure describes a generator producing very limited varieties?
A Data leakage
B Tokenization
C Mode collapse
D Underfitting
Answer: C. Mode collapse. Mode collapse occurs when a GAN generator produces only a narrow range of outputs instead of covering the variety in the data.
During VAE training, what does an encoder commonly produce?
A A final class label
B A noise-free photograph
C A discriminator score
D Latent distribution parameters
Answer: D. Latent distribution parameters. A variational autoencoder encoder maps an input to parameters describing a distribution in latent space, from which a latent representation can be sampled.
What role does a VAE decoder perform?
A Labels training examples
B Ranks search results
C Reconstructs from latent codes
D Measures prompt length
Answer: C. Reconstructs from latent codes. The decoder maps latent representations back into data space, supporting reconstruction and generation.
In a generative model, what does latent space represent?
A Public web search results
B Compact internal representations
C Human-written safety rules
D A list of output labels
Answer: B. Compact internal representations. Latent space contains internal representations that encode useful structure from which a model can reconstruct or generate examples.
In a diffusion model's forward process, what is gradually added to data?
A Noise
B Class labels
C Copyright notices
D Text captions
Answer: A. Noise. A common diffusion training process gradually corrupts data with noise, teaching a model to reverse that corruption.
How does reverse diffusion typically form a sample?
A Adds labels in one pass
B Copies a training image
C Removes noise in steps
D Sorts pixels by brightness
Answer: C. Removes noise in steps. During sampling, a diffusion model repeatedly estimates and removes noise to form a coherent sample.
Why condition a generative model on a text prompt?
A Increase the training dataset
B Steer the generated output
C Guarantee factual claims
D Delete model parameters
Answer: B. Steer the generated output. Conditioning supplies information, such as a text prompt, that guides the kind of output the model generates.
What does a text-to-image system return after receiving a description?
A A class label only
B A spreadsheet formula
C A source citation
D Generated image
Answer: D. Generated image. Text-to-image systems use text as conditioning information to generate an image that attempts to match the description.
When editing only a selected region of an image, which technique is commonly used?
A Inpainting
B Outpainting
C Tokenization
D Clustering
Answer: A. Inpainting. Inpainting fills or changes a selected image region while using surrounding context to guide the result.
What does outpainting ask an image model to extend?
A A sentence into tokens
B Labels across a dataset
C Audio into captions
D Beyond original image boundaries
Answer: D. Beyond original image boundaries. Outpainting generates content beyond an image's original edges, extending its canvas while considering the existing scene.
For image-to-image generation, what usually serves as the starting input?
A Only a class label
B A database schema
C An existing image
D A finished audio track
Answer: C. An existing image. Image-to-image workflows use an input image, often with a prompt, to guide a transformed or newly generated result.
Which part of a prompt most directly guides a model's subject and style?
A Descriptive instructions
B Learning rate values
C Training labels
D Database indexes
Answer: A. Descriptive instructions. A prompt can describe a subject, visual style, setting, and constraints to guide the generated output.
Why specify composition and subject details in an image prompt?
A Increase model storage
B Reduce ambiguity for generation
C Prove copyright ownership
D Remove all randomness
Answer: B. Reduce ambiguity for generation. Clear details can narrow interpretation and guide composition, although prompts cannot guarantee exact control.
During prompt iteration, why change one detail at a time?
A Increase context limits
B Prevent every model error
C Identify each change's effect
D Replace evaluation data
Answer: C. Identify each change's effect. Changing one element at a time makes it easier to see which instruction affected the generated result.
Repeated sampling from one prompt can produce different results because generation may use what?
A A fixed database lookup
B Only manual drawing
C A permanent answer key
D Randomized sampling
Answer: D. Randomized sampling. Many generators sample among plausible outputs, so repeated runs can vary even when the prompt is unchanged.
What can a random seed help make more repeatable?
A Training-data ownership
B Sampling choices
C Model architecture
D Internet search results
Answer: B. Sampling choices. A seed can initialize randomness so sampling is more reproducible under matching model and settings, though exact outputs may still vary across systems.
Do identical prompts guarantee identical outputs across model versions?
A No, versions can differ
B Yes, every time
C Only for image prompts
D Only without training
Answer: A. No, versions can differ. Different model versions, settings, and infrastructure can produce different outputs from the same prompt.
In conditional generation, guidance strength usually influences what?
A Number of training labels
B File download speed
C Copyright status
D Adherence to conditioning
Answer: D. Adherence to conditioning. Guidance strength can affect how closely a generated sample follows conditioning such as a prompt, with trade-offs that vary by model.
After producing one token, how does an autoregressive text model usually continue?
A Restarts its training
B Rewrites the dataset
C Predicts another token
D Checks a human answer key
Answer: C. Predicts another token. Autoregressive generation repeatedly predicts a next token based on the preceding context and continues until a stopping condition.
Which setting can limit how much text a model generates?
A Maximum output tokens
B Dataset row order
C Image resolution only
D Number of categories
Answer: A. Maximum output tokens. An output-token limit caps generated text length. Some interfaces also support stop sequences or other ending conditions.
Why might generated text cite a nonexistent book or article?
A It always searches the web
B It can invent plausible details
C It stores every citation
D It only copies databases
Answer: B. It can invent plausible details. Generative models can produce plausible-sounding but unsupported details, including fabricated citations, so references should be checked.
Which result is an example of multimodal generation?
A One numeric class label
B A sorted customer list
C Image with generated caption
D A fixed calculation result
Answer: C. Image with generated caption. Multimodal systems work across more than one data modality; a workflow may generate or combine images and text.
What does text-to-speech generation produce?
A Spoken audio from text
B Text labels from images
C A database table
D A chart from numbers
Answer: A. Spoken audio from text. Text-to-speech systems convert written language into synthesized speech audio.
What output is a music-generation model designed to create?
A Training labels
B Music or audio
C Search indexes
D Camera metadata
Answer: B. Music or audio. Music-generation systems produce musical or audio sequences, often from prompts or learned patterns.
When reviewing generated video, which quality needs attention across frames?
A Spreadsheet cell borders
B File naming rules
C Database normalization
D Temporal consistency
Answer: D. Temporal consistency. Video quality depends on coherent motion and objects across time, not only on how one individual frame looks.
What visual defect can occur in generated images?
A Permanent screen burn
B Missing database keys
C Inconsistent object anatomy
D Wrong audio sample rate
Answer: C. Inconsistent object anatomy. Generated images may contain malformed anatomy, inconsistent objects, or other visual artifacts that need inspection.
Why should lettering inside a generated image be checked carefully?
A Text may be rendered incorrectly
B Images cannot contain letters
C Text always becomes audio
D Every font is copyrighted
Answer: A. Text may be rendered incorrectly. Image models can struggle to render exact spelling and typography, so visible text should be reviewed and corrected.
What does a reference-based generation metric rely on?
A A model's source code
B A user's keyboard
C A website's color palette
D Comparison reference examples
Answer: D. Comparison reference examples. Reference-based metrics compare outputs with selected reference examples; their scores depend on the quality and relevance of those references.
Why include human review when evaluating generated content?
A Replace all objective measures
B Judge context and usefulness
C Avoid checking safety
D Guarantee originality
Answer: B. Judge context and usefulness. Human reviewers can assess context, relevance, and subtle quality issues that automated scores may not capture.
Does a high benchmark score guarantee that generated content fits a real user's task?
A No, task context matters
B Yes, benchmarks prove it
C Only for short prompts
D Only for images
Answer: A. No, task context matters. Benchmarks measure selected tasks and conditions. Real use may involve different users, constraints, and failure costs.
What can happen when a generator learns from unrepresentative data?
A All outputs become identical
B Training becomes unnecessary
C Outputs can reflect bias
D Every claim gains citations
Answer: C. Outputs can reflect bias. Unrepresentative training data can shape the model's learned patterns and may produce skewed or stereotyped outputs.
Why might generated characters reflect social stereotypes?
A Prompts disable all learning
B Training patterns may encode them
C Images contain no information
D Models have no training data
Answer: B. Training patterns may encode them. Patterns in training data and design choices can influence generated representations, including stereotyped depictions.
What information can content provenance records preserve?
A A guaranteed truth score
B Every private prompt
C A copyright transfer
D Origin and editing history
Answer: D. Origin and editing history. Provenance metadata can help record where content came from or how it was edited or generated; it does not prove that content is true.
Does a watermark alone conclusively prove that media was generated by AI?
A No, it can be altered
B Yes, it cannot change
C Only for audio
D Only when printed
Answer: A. No, it can be altered. Watermarks can be removed, altered, or absent, so they should not be treated as conclusive proof of origin.
Why label synthetic media when it could be mistaken for a real event?
A Increase file size
B Improve model accuracy
C Reduce risk of deception
D Hide the source
Answer: C. Reduce risk of deception. Clear labeling can help audiences distinguish synthetic media from authentic recordings, reducing the chance of misleading people.
Which use of generated voice poses a serious misuse risk?
A Reading a public poem
B Narrating a tutorial
C Producing licensed audio
D Impersonation without consent
Answer: D. Impersonation without consent. Cloning or imitating someone's voice without consent can enable deception, fraud, and reputational harm.
Which safeguard can help reduce harmful generated outputs?
A Removing all evaluation
B Content review and filters
C Hiding model limitations
D Disabling human oversight
Answer: B. Content review and filters. Safety measures can combine filters, testing, monitoring, and review; no single safeguard prevents every harmful output.
Why review usage rights before commercializing generated content?
A Terms and rights can vary
B All outputs are public domain
C Prompts transfer every copyright
D Commercial use is always barred
Answer: A. Terms and rights can vary. Rights and terms depend on the service, source material, jurisdiction, and use. Users should check applicable permissions.
Can a generative model sometimes reproduce memorized training text?
A No, by definition
B Only after translation
C Yes, in some cases
D Only from images
Answer: C. Yes, in some cases. Some models may reproduce memorized material, especially when it appeared repeatedly or was unusually distinctive in training data.
Before submitting sensitive material to a generative service, what should a user check?
A Prompt font size
B Data-use and retention terms
C Image color balance
D Output file extension
Answer: B. Data-use and retention terms. Users should understand whether submitted data may be stored, reviewed, or used to improve services before sharing sensitive material.
What risk can arise when generated samples are repeatedly reused for training?
A All bias disappears
B Every output becomes verified
C Models stop needing evaluation
D Errors may compound
Answer: D. Errors may compound. Repeated training on generated data can amplify errors or reduce diversity if the data is low quality or poorly curated.
Why keep evaluation examples separate from generated training material?
A Make prompts shorter
B Increase watermark visibility
C Avoid evaluation leakage
D Reduce image resolution
Answer: C. Avoid evaluation leakage. If evaluation material influences training, reported performance may become misleading because the model has effectively seen the test information.
When real examples are scarce, how can synthetic data help?
A Replace all validation
B Augment selected training data
C Prove regulatory compliance
D Guarantee unbiased outputs
Answer: B. Augment selected training data. Carefully reviewed synthetic examples can supplement limited data, but they need quality checks and should not replace appropriate real-world evaluation.
What concern arises if synthetic examples cover only a narrow range?
A Model accuracy is guaranteed
B Privacy risks disappear
C No review is needed
D Limited diversity persists
Answer: D. Limited diversity persists. Narrow synthetic data can preserve gaps or bias instead of representing the range needed for the intended task.
As the number or size of generated outputs grows, what resource demand may rise?
A Compute and energy use
B Copyright ownership
C Human creativity
D Dataset labels
Answer: A. Compute and energy use. Large generation workloads require computational resources, and the amount depends on model size, hardware, and use.
Why should a professional treat generated material as a draft to review?
A AI outputs are always private
B Errors and omissions remain possible
C Review blocks creativity
D Drafts cannot be edited
Answer: B. Errors and omissions remain possible. Generative output can be useful as a starting point, but people should verify facts, suitability, permissions, and safety before relying on it.
How does unconditional generation differ from prompt-conditioned generation?
A Needs more training examples
B Only returns category labels
C Uses no conditioning input
D Always copies a sample
Answer: C. Uses no conditioning input. Unconditional generation samples without a specific condition such as a text prompt, while conditional generation uses supplied information to steer output.