Generative AI — Module Assessment

Question 1

What most clearly distinguishes a generative model from a discriminative one?

Accepted Answer

A generative model produces new content, while a discriminative model assigns inputs to categories

Answer

A generative model is always more accurate than a discriminative model

Answer

A generative model needs no training data, while a discriminative model does

Answer

A generative model understands meaning, while a discriminative model does not

Question 2

Which best describes how a generative model produces a novel output?

Accepted Answer

It recombines statistical patterns learned from training data into a new sample

Answer

It copies the closest matching example from its training set

Answer

It invents content from nothing, independent of any data

Answer

It searches a database of stored answers and returns the best match

Question 3

What is a large language model fundamentally trained to do?

Accepted Answer

Predict the most likely next token in a stretch of text

Answer

Retrieve verified facts from an internal database

Answer

Reason about meaning the way a human reader does

Answer

Store and recall the exact documents it was trained on

Question 4

Why can a large language model produce fluent output that is completely false (a hallucination)?

Accepted Answer

It optimises for plausible-sounding text, not for checking what is true

Answer

It deliberately lies when it lacks information

Answer

It malfunctions and stops predicting tokens correctly

Answer

It copied the false statement verbatim from its training data

Question 5

Why does a vague prompt like "Write about climate change" tend to produce a generic answer?

Accepted Answer

With nothing to narrow it, a next-token predictor drifts toward the most average, likely wording.

Answer

The model refuses to answer prompts it considers too short.

Answer

The model looks the topic up online and returns the top result.

Answer

The model remembers your earlier chats and reuses them.

Question 6

Which statement best captures what 'prompting' is?

Accepted Answer

Steering a probabilistic system with instruction and context, usually refined over several tries.

Answer

Writing exact commands the model must execute the same way every time.

Answer

Permanently teaching the model new facts about you.

Answer

Selecting which website the model will search.

Question 7

What does it mean when a generative model 'hallucinates'?

Accepted Answer

It generates plausible-looking but fabricated content, such as a fake citation, with full confidence.

Answer

It stops responding because it lacks the information.

Answer

It deliberately lies to mislead the user.

Answer

It searches the web and returns an outdated page.

Question 8

Which task is generative AI most reliably suited to?

Accepted Answer

Summarising a document you paste in and suggesting alternative headlines.

Answer

Reporting your company's revenue from memory with no data provided.

Answer

Computing a long multiplication exactly.

Answer

Listing verified citations for a legal brief.

Question 9

In plain terms, how does a diffusion-based image generator produce a picture from a prompt?

Accepted Answer

It starts from random noise and refines it step by step, steered by the prompt, until a coherent picture emerges.

Answer

It searches a database of photos and pastes matching pieces together.

Answer

It photographs a real scene and edits it to match the words.

Answer

It copies the single closest training image and relabels it.

Question 10

Why is checking provenance a better defense against deepfakes than trying to spot fakes by eye?

Accepted Answer

Realistic fakes are designed to pass visual inspection, so tracing where a file came from and who vouches for it is more reliable.

Answer

Provenance software can always detect any synthetic file automatically.

Answer

Deepfakes always contain obvious visual glitches that anyone can see.

Answer

Only experts can generate deepfakes, so ordinary media is safe.

Question 11

An AI gives you three well-formatted book citations for an essay. What is the responsible next step?

Accepted Answer

Independently confirm each source exists and says what the AI claims before citing it.

Answer

Cite all three, since the formatting shows they are real.

Answer

Cite only the first one to be safe.

Answer

Assume they are fabricated and never use AI for research again.

Question 12

Which situation best fits keeping a human responsible rather than letting a generative tool decide?

Accepted Answer

Making a medical or legal decision with real consequences for a person.

Answer

Brainstorming a list of essay topic ideas.

Answer

Rephrasing your own clumsy sentence.

Answer

Drafting routine boilerplate you will review.

Question 13

What best captures a calibrated view of generative AI's recent progress?

Accepted Answer

Progress is real but uneven — strong at language, weak and unpredictable at reliability and reasoning.

Answer

Progress has been fast and smooth, so capability will keep improving along the same line.

Answer

Progress is an illusion created by marketing; the tools do nothing useful.

Answer

Progress means the systems now understand and reason the way people do.

Question 14

Why are the earlier 'AI winters' relevant to how you read today's AI claims?

Accepted Answer

They show a repeating pattern of overpromising, disappointment, and then under-rating the genuine progress that remained.

Answer

They prove AI never works and should be ignored.

Answer

They show that once hype fades, all real progress is lost.

Answer

They show that government funding is the only thing that matters for AI.