What ChatGPT actually is

• Describe ChatGPT as a product built on a large language model.
• Explain a token and next-token generation without jargon.
• Separate the model from optional tools such as search or files.
• Reject three common myths about how answers are produced.
• State what the model has not been given unless you provided it.

1. 01Say the job in one sentenceBefore you type, finish this: 'I need language that helps me ____.' Draft, rewrite, outline, quiz, compare, or extract are all honest jobs. 'Tell me the truth about ____ with no checking' is not.
2. 02Picture the answer being builtYour words go in. The model continues in the voice and structure your words suggest. It is not opening your files, your inbox, or today's news unless a tool actually did that and the product shows you that it did.
3. 03Ask what it is usingOn any answer that matters, add: 'Say whether you used a live source, a file I gave you, or only your own generation. If you are unsure, say so.' The wording of that control will vary. The question should not.
4. 04Mark what you did not provideIf you did not give numbers, names, dates, or a source, treat those as unconfirmed even when they look tidy.
5. 05Keep the model and the product apartModel names, 'thinking' controls, and modes such as chat versus longer agent work change. Learn the job of each control in your account. Do not memorize a name from this course as if it were permanent.

A new analyst meets the tool

Jordan just started at a small logistics company. A teammate says, 'Ask ChatGPT for our on-time rate last quarter and the industry benchmark.' Jordan has not uploaded anything.

1. Jordan notices the on-time rate is private company data. The model cannot know it.
2. Jordan asks instead: 'Give me a checklist for calculating an on-time rate from a shipment table, and list the definitions I must confirm with my team before I publish a number.'
3. The answer is useful because it is a method, not a fake statistic.
4. Jordan later pastes a redacted sample of column names — not customer addresses — and asks the model to map those columns onto the checklist.

What was our company's revenue last quarter, and how do we compare to competitors? Be accurate.

I will not give you private figures. Teach me a comparison method: which public sources a human should open, which numbers I must export from our own system, and how to label estimates versus actuals in a one-page note for my manager. Do not invent our revenue or anyone's market share.

The first prompt demands facts the model was never given and rewards a confident fiction. The second asks for a method, draws a hard line around invention, and leaves the numbers in the systems that actually have them.

• Treating a smooth paragraph as proof that a lookup happened.
• Assuming the chat can see your email, drive, or last quarter because 'AI knows everything.'
• Memorizing a model name or button from a tutorial that is already out of date.
• Asking it to be 'accurate' instead of telling it what not to invent.

• When you need a definition or a rewrite of your own words, generation is enough. When you need a fact about the world, plan a check.
• If the answer cites a paper, a law, or a quote, assume the citation is unconfirmed until you open it.
• A new chat does not contain the previous chat unless you paste the relevant part or a memory feature surfaces it — and memory can be wrong.
• Say 'method, not a number' whenever the number lives in your business, not in the prompt.

Write one sentence describing a task you actually have this week. Label it 'language job' or 'fact job.' If it is a fact job, rewrite the ask so it requests a method and a list of what you must verify. Paste that rewrite into ChatGPT and see whether it still invents a number. If it does, reply: 'Remove every figure I did not give you.'

Explain ChatGPT to a skeptical colleague in under two minutes, out loud, with no product names except ChatGPT itself. You pass if they can repeat back: it generates language, tools are optional, and I still check facts.

• ChatGPT generates language. That is the feature and the risk.
• Tools may ground an answer. Do not assume they ran.
• It does not know your private world unless you bring that world in — carefully.
• Confidence is a tone, not a source.
• Learn jobs and checks. Let menus change without breaking your skill.

Production notes for diagrams. These are not screenshots of ChatGPT. If a brief asks for the product window, capture it live on your own account.

The player above runs this cut. Captions are the lesson, not a fake recording of ChatGPT.

Opening hook

It writes like it checked. Often, it did not.

Teaching

Tokens, generation, and the difference between the model and optional tools.

Demonstration

Ask for a company's private revenue with no data, then ask for a method instead. Read both answers aloud.

Practice

Learners label one real task as a language job or a fact job and rewrite the fact job.

Challenge

Explain the mental model to someone else in two minutes.

Recap

Fluent is not the same as checked. You own the check.

Next

Next you will see the work it is genuinely good at — and the failures beginners miss.

Instructor script

Pause before the next prompt you type and ask what job you are hiring it for. If you need sentences shaped, summaries tightened, or options laid out, you are in the right place. If you need a number that lives in your company, the model does not have it until you bring it — and even then you check the math. Optional tools can search or read a file. They are not always on, and their buttons will move. Your durable skill is simpler: know when you are looking at a draft, and do not promote a draft into a fact just because it sounds finished.