Generative AI¶

Level: Beginner · Time: ~25 min · Exam relevance: CCAO-F D1 (1.4), D4 (4.5), D6 (6.1) · CCAR-P D1 (1.1)

In plain English¶

Most AI before about 2020 answered questions about existing things: "Is this email spam?", "What is in this photo?", "How much will this house sell for?" The output was a label or a number.

Generative AI produces new content: a paragraph, a poem, a slide outline, an image, a piece of code, a summary of a 40-page report. You describe what you want in ordinary language, and the model creates something that did not exist before.

Think of the difference between a librarian and a ghostwriter. A librarian (classic AI) finds and labels what already exists. A ghostwriter (generative AI) writes something new in the style and shape you ask for, drawing on everything they have ever read. A good ghostwriter is fast and versatile, but you still read the draft before putting your name on it.

Why it matters¶

Generative AI changed who can use AI. Before, building an AI system meant collecting labelled data and training a model, which needed specialists. Now a marketing coordinator can type "draft three subject lines for our spring newsletter, friendly tone, under 50 characters" and get a useful result in seconds.

That accessibility is why both certifications exist. CCAO-F tests whether business users can direct generative AI well and judge its output. CCAR-P tests whether architects can build reliable systems around it.

How it works¶

Discriminative versus generative¶

	Discriminative (classic) AI	Generative AI
Question it answers	"Which category does this belong to?"	"What should come next?" or "What would fit this description?"
Output	A label, score or number	New text, images, audio, code
Typical training	Labelled examples for one task	Very large, mostly unlabelled data, then refinement
Flexibility	One task per model	Many tasks from one model, chosen by your instructions
Main risk	Misclassification	Plausible-sounding but wrong content (hallucination)

Generation is repeated prediction¶

Every generative model works by learning the patterns in its training data well enough to produce new samples that fit those patterns. For text, the pattern is "what word-piece tends to come next, given everything so far". The model predicts one piece, adds it to the text, and predicts again, over and over, until the answer is complete. Large language models explains this in more depth.

sequenceDiagram
    participant You
    participant Model
    You->>Model: "Write a one-line welcome for new staff"
    Model->>Model: predict "Welcome"
    Model->>Model: predict " to"
    Model->>Model: predict " the"
    Model->>Model: predict " team"
    Model->>Model: ... continue until the line is complete
    Model-->>You: "Welcome to the team! We're glad you're here."

Because the model is generating rather than retrieving, two important things follow:

The output is new. It is not copied from a stored document. This is why generative AI can adapt to your exact request.
The output is not guaranteed to be true. The model produces what is plausible given its patterns. Usually plausible and true coincide. Sometimes they do not, and that is a hallucination.

Kinds of generative models¶

Modality	What it produces	Example uses
Text (LLMs)	Prose, code, structured data such as JSON	Drafting, summarising, analysis, coding
Images	Pictures from a text description	Illustrations, design mock-ups
Audio	Speech, music	Voice-overs, accessibility
Video	Short clips	Marketing, training content
Multimodal	Reads several kinds of input, such as text plus images	Reading a chart in a PDF and explaining it

Claude is a text-generating model that can also read images and documents (it is multimodal on the input side). It produces text, including code and structured data, and in the Claude apps that text can be turned into documents, diagrams and small interactive tools called Artifacts.

What generative text AI is good at¶

A helpful way to think about tasks is by what you ask the model to do with information:

Task type	You give it	It gives you	Example
Create	A goal and constraints	A first draft	"Draft a job advert for a data analyst"
Transform	Existing content	The same content in a new form	"Turn these meeting notes into an action list"
Condense	Long content	A shorter version	"Summarise this policy in five bullets"
Analyse	Content plus a question	Reasoning and a conclusion	"Which of these three vendor proposals has the weakest security section, and why?"
Brainstorm	A problem	Many options	"Give me ten names for an internal newsletter"

CCAO-F objective 1.4 asks you to adapt your prompting to task types like these. You will practise this in Module 03.

Example¶

Try this in any Claude app (see Your first conversation for how to get started). Paste these two requests separately and compare the results.

Classify this customer comment as positive, negative or neutral:
"The delivery was late but the driver was very polite."

Write a short, warm reply to this customer comment, acknowledging the late delivery
and thanking them for mentioning the driver:
"The delivery was late but the driver was very polite."

The first request uses the model in a discriminative way: one label out. The second is generative: new text that did not exist. The same model handles both; only your instruction changed.

Enterprise use case¶

Contoso Health, a fictional regional healthcare provider, wants to reduce the time nurses spend writing discharge instructions. The team considers two ideas:

Idea A: use generative AI to write discharge instructions from scratch for each patient.
Idea B: use generative AI to transform the clinician's existing notes into plain-language instructions, which the clinician then reviews and signs.

They choose Idea B. Transformation keeps the clinical facts anchored in a source written by a professional, which reduces the chance of invented content, and the mandatory clinician review keeps a qualified human in the loop for a high-risk medical context. This reflects a broader pattern: generative AI is safest when it reshapes trusted input rather than inventing facts.

Common mistakes¶

Treating generated text as retrieved fact. The model writes what fits; it does not look things up unless it is given a tool such as web search or a document to read.
Expecting identical output every time. Generation involves choices at every step, so running the same request twice can produce different wording. If you need consistency, use clear structure, examples and output formats (see Module 03).
Using generative AI for a pure lookup. "What is our office's postcode?" is a database question, not a generation task.

Security considerations¶

Generated content can include confident errors, outdated information or biased framing. Review it in proportion to the stakes.
Generated code can contain vulnerabilities. Treat it like code from an unknown contributor: review and test before use.
Generated content is untrusted output from a software point of view. Never feed it directly into a system that executes it (a database, a shell, a web page) without validation. Module 12 covers this.

Troubleshooting¶

Symptom	Likely cause	What to do
Output is generic and bland	The request gave little context or constraint	Add audience, purpose, tone, length and an example
Output contains a "fact" you cannot find anywhere	The model generated something plausible but unsupported	Ask for sources, provide the source document, or verify independently
Output varies a lot between runs	Open-ended request	Specify structure and format; give an example of the desired output

Check your understanding¶

What is the difference between a discriminative and a generative task? Give one example of each.
Why can generated text be fluent but false?
Which is generally lower-risk: asking a model to invent content from nothing, or to transform a trusted source? Why?
Name the five task types in the table above.

Answers

A discriminative task assigns a label or score to existing input (classifying sentiment); a generative task produces new content (writing a reply).
Because the model generates what is statistically plausible given its training patterns, and plausible text is not always true.
Transforming a trusted source, because the facts come from the source rather than from the model's patterns, and they are easier to check.
Create, transform, condense, analyse, brainstorm.