What Is Generative AI, Really?

Ask a search engine "what were our Q3 support trends" and you get links to open and read. Put the same question to Copilot and it writes a paragraph that didn't exist before you asked. One finds existing material. The other generates a response, which makes the wording of your request part of the work.

Generative AI writes answers

Generative AI creates content (text, tables, summaries, images) from an instruction called a prompt. That's the useful definition. Earlier AI systems were commonly used to sort things, such as routing a support ticket or flagging an unusual transaction. Generative AI produces a draft email, a five-row comparison, or a rewritten paragraph.

Three words get used as if they mean the same thing. They don't:

Every one of these tasks follows the same arc. You give an input (your prompt and any material you supply), the model processes it, and an output comes back that you can use, edit, or reject. The systems behind Microsoft Copilot are large language models (LLMs): models that learned patterns from an enormous amount of text and use them to generate language. You do not need to know how one works inside to use it well. You need to know what it is good at, and where it slips.

Is Copilot a search engine or a reasoning engine?

Two very different jobs hide inside the words "answer this question," and telling them apart changes how you ask:

• Retrieval locates information that already exists somewhere. "What is the current mileage reimbursement rate?" You need an authoritative, up-to-date source.
• Reasoning connects facts, constraints, and priorities into a conclusion. "Given these three venue quotes, which one fits our budget and date?" You need the answer worked out from the supplied facts.

A search engine is built for retrieval. A reasoning engine is built for analysis. Copilot can do both, but it does them best when you know which one you're asking for. "Find the current policy and recommend an option" combines both jobs: retrieve the policy first, then reason over it.

Microsoft Copilot offers conversation modes that lean toward one job or the other. Search is grounded in the web and returns citations you can open, so use it when the answer depends on current public information. Think Deeper uses reasoning models, which suits tasks where you have supplied the facts and need them compared, sequenced, or explained. Neither mode guarantees a correct answer. Each changes how Copilot approaches the request, which is why identifying the job matters more than clever phrasing.

A fluent answer can still be wrong

An LLM can write smoothly while getting a name, date, or number wrong. The incorrect answer arrives in the same assured voice as the correct one, so tone gives you no warning.

A made-up detail is called a fabrication (you'll also hear "hallucination"). It could be a meeting date missing from your notes, a statistic rounded incorrectly, or a next step the source only implied. These details are easy to miss because they fit the paragraph. The useful question is "where exactly is this supported?" If you can't point to the evidence, the claim still needs checking.

Two assumptions are best dropped early. A bigger or newer model isn't automatically more accurate. Capacity affects how much text a model can consider at once, while accuracy still depends on the source, the task, and your checks. The same prompt can also return different answers on separate runs. Repetition doesn't prove a claim. Check each answer on its own evidence.

Facts, guesses, and the line between them

When Copilot writes an answer, it mixes four different kinds of statement, and telling them apart is a skill you'll use forever:

• A fact is something your source directly supplies: "The venue holds 120 people." It's written down.
• An inference is a conclusion Copilot worked out from facts: "So it's too small for our 150 guests." Sound, but it's a calculation, not a quote.
• An assumption is something treated as true without confirmation: "The room will be free that evening." Nobody checked.
• A recommendation is a proposed action built on all of the above. "Book the larger hall instead."

Copilot rarely labels these for you. It blends them into one smooth paragraph, so ask of each important sentence: which kind is this? Trace a fact to its source, rerun an inference, flag an assumption, and take responsibility for a recommendation. This is the verification skill you'll build through the rest of the module.

Where generative AI is most useful

Keep it on work you can judge. Most useful tasks in this course fall into four groups:

• Create: produce a first draft, outline, agenda, or list of ideas from scratch.
• Summarize: compress a long document, thread, or meeting into decisions, risks, and next steps.
• Transform: shorten text, change its tone, or put the same material into a table.
• Analyze: compare options or apply a stated rule to information you've supplied.

In each case, you can check the result against something concrete: the source document, the numbers you pasted, or your judgment of the tone. Generative AI is less dependable when verification is difficult, including obscure facts, current prices, and information that changed last week. A quick first draft can still be valuable when checking it takes less time than starting from a blank page. In the next lesson, you'll try these four patterns in Microsoft Copilot and see how its conversation modes affect the result.