ChatGPT is OpenAI’s AI assistant: you type or talk to it, and it writes back. People also use the name for OpenAI’s models and for the company, so it helps to say which one you mean. The most useful thing to understand as a beginner is that each answer is built from more than the message you just typed. Earlier messages, saved memories, your settings, attached files, and tools like web search all feed in, and you can see and control each one.
Say you open a new chat and ask for a quick dinner idea, and ChatGPT suggests three meals with no meat, adding that it kept them vegetarian for you. You never said that in this chat. A few weeks ago you mentioned it in passing, and ChatGPT saved it as a memory. That small moment explains a lot about how ChatGPT works. It is not reading your mind, and it is not only reading your message either. It is answering from a bundle of things, and once you know what is in the bundle, its answers stop being a mystery.
This post opens Learn ChatGPT from scratch. If you want the bigger picture first, meaning what a large language model (LLM, the kind of AI behind chatbots like ChatGPT) is and why it predicts text, start with our plain-English guide to LLMs and generative AI. For habits that work with any AI tool, see the Practical AI series.
What is ChatGPT, exactly?
ChatGPT is an app made by OpenAI, a company based in San Francisco, that lets you have a written or spoken conversation with an AI model. The model is the trained program that reads your words and writes a reply. GPT is OpenAI’s name for its family of models, and the “GPT” in ChatGPT comes from it. The app is what you open on the web at chatgpt.com, on your phone, or on your computer. So when a coworker says “ChatGPT,” they almost always mean the app, and when a developer says “GPT,” they usually mean the models.
The model learned patterns from a huge amount of text during training, and it uses those patterns to write the most likely useful reply to what it is given. It does not look up a stored answer. That is why it can write a fresh email in your tone, and also why it can state something false with total confidence. Everything else in this post is about what it is given, because that part you control.
What goes into a ChatGPT answer
When you press send, ChatGPT gathers several things and passes them to the model together. Not all of them are switched on for everyone, and the exact settings move around as OpenAI updates the app, but these are the usual ingredients:
| Ingredient | What it adds | How you control it |
|---|---|---|
| Your message | The question or task, plus any text you paste | Write who it is for, what it is for, and the format you want |
| This chat so far | Everything said earlier in the same conversation | Start a new chat when you change topic |
| Saved memories | Facts ChatGPT kept from earlier chats, like the vegetarian note | Ask “what do you remember about me?”, and review or delete memories in Settings |
| Custom instructions | Standing preferences you wrote, such as “keep answers short” | Edit them in Settings under personalization |
| Project files and notes | Files and instructions shared by every chat in a Project, a folder that groups related chats | Add or remove files in the Project |
| Attached files | A document, spreadsheet, or image you upload to this chat | Attach only what the task needs |
| Tools | Results from web search, from code it runs on your data, or from apps you connect | Turn tools on for a task, or ask it to search or not |
Under all of that sits what the model learned in training, which stops at a certain date. Anything newer, and anything private to you or your company, reaches the answer only through one of the ingredients above.
This list explains most of the surprises people run into. An answer that mentions something you never said came from memory or an earlier message. An answer that ignores your spreadsheet may mean the file never attached, or that the chat was so long the file got crowded out. An answer with today’s news means web search ran, and an answer with old prices means it did not. When an answer looks odd, ask which ingredient caused it before you decide ChatGPT is broken.
Try it: see the ingredients at work
This short experiment takes about fifteen minutes and makes the list above concrete. Use a free or paid account, and use made-up or harmless details only.
- In a new chat, ask “What do you remember about me?” Read the answer, then open the memory section in Settings and compare. You now know what ChatGPT brings into every chat without being asked.
- Ask for a dinner idea in a normal chat, then ask the same thing in a Temporary Chat, a chat mode that does not use or save memories. Compare the two answers to see what memory added.
- Ask “What is the weather in my city today?” with web search off and then with it on, if your plan has it. Notice how the first answer either refuses or guesses, while the second cites sources you can open.
- Paste a short paragraph you wrote, and ask for a summary in two sentences. Then attach the same text as a file and ask again. The answers should match, which shows that pasting and attaching do the same job for short text.
Write down one thing that surprised you. For many people it is the memory list, which is usually longer than they expected. Delete anything there you would rather it did not keep.
When to start a new chat
The chat so far is one of the strongest ingredients, and it can work against you. A conversation that started as a budget question and drifted into holiday plans carries both topics into every answer. A very long chat has a second problem: the model can only take in so much text at once, so in a long enough conversation the early details, including a file you attached at the start, get less attention or drop out.
Start a new chat when you change topic, when answers start repeating mistakes you already corrected, or when ChatGPT seems to forget a detail from early on. Before you leave the old chat, ask it to “summarize the decisions and facts we settled, as a short list.” Paste that list into the new chat as your first message. You keep what mattered and drop the clutter, and the answers usually get sharper straight away.
If you keep coming back to the same subject, such as a monthly report, put it in a Project instead. The Project holds the files and the standing instructions, so each new chat inside it starts with the right background and none of last month’s back and forth.
Why two people get different answers
Two people can type the same question and get different answers, and the same person can get a different answer by asking twice. Four things cause this. First, the model picks its words with some randomness, so wording changes from one try to the next even when the substance stays the same. Second, each person’s ingredients differ: different memories, instructions, and earlier messages. Third, each plan offers different models, and free accounts may use a lighter one. Fourth, tools change the inputs, since a web search today returns different pages than one last month.
This matters when you share results at work. If a teammate says “ChatGPT told me the deadline is Friday,” the useful questions are which chat, with which file, and whether search was on. For anything that will be acted on, the reliable move is to share the source, not the chat. Checking the original document takes a minute and settles the question for everyone.
Where ChatGPT’s knowledge stops
ChatGPT knows only what it learned in training plus what the ingredients bring in. That leaves three gaps worth remembering:
- Recent events. Without web search, it does not know what happened after its training ended, and it may not say so. For news, prices, and opening hours, use search and open the sources it cites.
- Your private facts. It does not know your company’s sales, your bank balance, or your doctor’s advice unless you give them to it. Asked anyway, it may produce numbers that look real and are invented.
- How ChatGPT itself works. It is often unreliable about its own settings, limits, and features. For questions about plans and features, OpenAI’s Help Center is the better source.
That is why ChatGPT is a strong helper for drafts, explanations, and options you can check, and a poor source of record. A source of record is the place your company treats as the official answer, such as the finance system for revenue or the contract for a deadline. Use ChatGPT to write about those facts, and take the facts themselves from the source.
Where you can open ChatGPT
You do not need every version on day one. The web app at chatgpt.com is where nearly everyone starts, and it covers writing, questions, and file work. The phone apps for iPhone and Android are good for quick questions and voice, and they share your chats with the web. The desktop apps for Mac and Windows add more: besides plain Chat, they offer Work, for longer tasks that take several steps on files, and Codex, OpenAI’s coding assistant. Later posts in this series cover the desktop app and each of those modes.
One thing is separate from all of these: the API, the service programmers use to call OpenAI’s models from their own software. It has its own account and its own bill, and a ChatGPT plan does not pay for it. If you are not writing software, you can ignore it for now.
What to keep out of the chat box
Because memories, files, and chats are stored with your account, think before you paste. On a personal account, keep out passwords, customer details, health records, and anything your employer treats as confidential. OpenAI explains how chats may be used to improve its models, and how to turn that off, in its help article on how your data is used. Company plans have their own data rules set by your employer, so if your workplace provides ChatGPT, use that account for work and follow its policy.
A good personal rule is to write down one category of information you will never paste, such as customer phone numbers. Deciding once is easier than deciding in the middle of a busy afternoon.
Practice: thirty minutes to make ChatGPT yours
- Run the four-step experiment above and note what each step showed you.
- Write custom instructions with two lines about yourself and two about how you like answers, for example “I work in operations” and “use plain language and short paragraphs.”
- Paste a non-sensitive paragraph you wrote last week and ask: “Rewrite this for a busy manager in 120 words. Keep all numbers exactly. Flag any claim without evidence.” Compare it with your original, line by line.
- Ask “What might be wrong or missing in your rewrite?” and decide whether you would send the draft unchanged.
- Write your one never-paste rule in your notes.
What to remember about ChatGPT
- ChatGPT is OpenAI’s assistant app, and GPT is the family of models behind it.
- Each answer is built from your message, the chat so far, memories, custom instructions, Project files, attachments, and tools.
- When an answer surprises you, find the ingredient that caused it.
- Answers vary between people and between tries, so share sources, not chats.
- It knows nothing recent or private unless an ingredient brings it in, so check facts at the source.
Your next step
Open ChatGPT today and ask “What do you remember about me?” Then open the memory settings, delete anything you would rather it forgot, and add custom instructions that describe how you like answers. Those two minutes shape every answer you get after them.
Series notes
This post opens Learn ChatGPT from scratch. The next post covers the plans, from Free to Enterprise, and what each one changes. For the same ideas applied to another assistant, see Learn Claude from scratch. For checking AI-written database queries, see how to check AI-written SQL.
Sources
- ChatGPT (the app)
- ChatGPT plans and pricing (which features each plan includes)
- OpenAI Help Center (memory, custom instructions, Projects, Temporary Chat, and data controls)
- OpenAI: How your data is used to improve model performance (the training opt-out)
- OpenAI Help: ChatGPT Work and Codex (the desktop modes)
- Analytics Made Simple: What are LLMs, ChatGPT, and generative AI (background)
Keep going
Same lessons in your feed
Short diagrams, hooks, and weekly tutorials on Substack, Instagram, X, and Facebook.
