Before you start: about this course
This course is independent of OpenAI. OpenAI has not written, reviewed or endorsed it. ChatGPT's features, names, plans and limits change often, sometimes from month to month. When this course names a feature, check it against OpenAI's own pages before you rely on it. Section 5 shows you how.
1. What ChatGPT is
ChatGPT is a chat tool built on a large language model. A language model learns patterns from a very large amount of text. When you type a message, it writes a reply one small piece at a time. At each step it picks a likely next word based on those patterns.
Think of a very well-read colleague who answers fast and never says "I'm not sure" unless asked to. This colleague is good at:
- Drafting emails, summaries and outlines
- Rewording text for a different audience
- Brainstorming options and checklists
- Explaining a concept in plain language
What the model does not do by default is look up a verified fact the way a database does. It produces text that sounds right. Most of the time that text is also correct. Sometimes it is not.
2. Why it can be confidently wrong (hallucination)
A hallucination is a reply that is fluent and confident but false. Common forms:
- Wrong dates, figures or definitions
- Made-up quotes, studies, citations or web links
- Overconfident answers to vague or complex questions
Why it happens, in plain terms:
- The model is trained to predict likely text. Nothing in that training labels each sentence as true or false.
- Rare, specific facts (a niche regulation, a small company's revenue) leave weak patterns, so the model fills the gap with something plausible.
- Models have been rewarded for giving an answer rather than admitting uncertainty, so a guess can come out sounding exactly like a fact.
Newer models hallucinate less, but they still do it. A confident tone tells you nothing about accuracy.
Work tasks where the risk is highest:
- Citations and sources: legal references, research studies, policy clauses, URLs
- Specific numbers: financial figures, statistics, prices, tax or payroll thresholds
- Regulated or high-stakes rules: employment law in HR, compliance and audit in finance, contract terms in sales
- Recent events: anything that changed after the model was trained, unless it is using a search tool
3. The context window: what ChatGPT can "see"
The context window is the amount of text the model can hold in view at one time. That includes your messages, its replies and any text you paste. Window sizes vary by model and plan, and they change, so check the current figure on OpenAI's pages.
What this means at work:
- In a very long chat, early details can drop out of view. The model may then ignore an instruction you gave at the start.
- A new chat starts with an empty context. It does not automatically know what you said in another chat. (Memory features change this, and they come later in the course.)
- What to do: keep chats focused on one task. Repeat key instructions. Start a fresh chat with a short summary when a thread gets long.
4. How your data is handled
Whatever you type or upload is sent to OpenAI's servers. What happens to it next depends on your plan and settings:
- Personal plans (Free, Plus, Pro): your chats may be used to train models unless you turn off Improve the model for everyone under Settings → Data Controls.
- Temporary Chat: these chats don't appear in your history, don't use or create memories and aren't used for training. They are deleted from OpenAI's systems within about 30 days.
- Workplace plans (Business, Enterprise, Edu): OpenAI states that it does not train on this business data by default.
Rule for now: don't paste confidential, personal or regulated data into ChatGPT until you know your plan and your company's policy. Plans and safe use are covered in later levels.
5. Task: run two limit tests and record the results
Open a notes document with three columns: Test / What happened / What I learned.
Test A: Hallucination
- Open a new chat.
- Ask a niche question that needs sources, for example: "List three published studies, with authors, year and journal, on four-day working weeks in accounts payable teams."
- Copy each title into a normal web search or a library database.
- Record how many studies exist exactly as described, how many exist with different details, and how many you can't find at all.
Test B: Context
- In a new chat, type: "For this whole chat, end every reply with the word CHECKED."
- Ask five or six unrelated work questions, and paste in a few long documents (for example, public reports).
- Note whether each reply still ends with CHECKED.
- Open a second new chat and ask: "What word must you end replies with?" Use Temporary Chat so memory doesn't affect the result.
- Record both results. Even if the rule held in the first chat, the second chat shows that context belongs to each chat.
Check your result: your notes should name each test, quote at least one reply, and give one practical rule you'll follow (for example, "I verify every citation before I use it").
6. Task: find OpenAI's current pages
- Go to help.openai.com, OpenAI's Help Center.
- Search for ChatGPT release notes and bookmark the page. It lists changes by date.
- Search the Help Center for Data Controls and read how the training setting works.
- When this course names a feature, look it up there. Note whether it is available on your plan and whether it is marked beta.
Check your result: you have two bookmarks (the release notes and the Data Controls article), and you can name one change from the latest release notes.
Key takeaways
- ChatGPT predicts likely text. It sounds fluent, but that doesn't make it accurate.
- Hallucination risk is highest for citations, specific numbers, regulated rules and recent events.
- The context window limits what the model can see. Each new chat starts fresh.
- Your plan and settings decide whether your chats can be used for training. Don't paste sensitive data until you know both.
- This course is independent of OpenAI. Check features on help.openai.com and the ChatGPT release notes.