What is ChatGPT? A Beginner’s Guide

ChatGPT is an AI assistant you talk to in ordinary language. You type a question or a request — explain this, draft that, fix this code — and it writes back in seconds. There is no syntax to learn and no commands to memorise. It was built by OpenAI and released publicly in November 2022, and it reached an estimated 100 million users within about two months, the fastest adoption of any consumer app at the time.

This guide covers what you can actually do with it, how to get a useful answer on your first try, what it costs, how it works underneath, and — the part most introductions skip — where it gets things wrong. You can also use the GPT models without an OpenAI subscription, which is covered near the end.

The honest short answer: it removes the blank page. A rough first draft in ten seconds is easier to fix than nothing is to start, and that applies to an email, a report, a function, a lesson plan or a difficult message you have been putting off.

Beyond writing, it explains complicated things at whatever level you ask for, which makes it genuinely useful when you are working outside your own field. It reads documents and images on paid plans. It writes and debugs code. It can search the web for anything recent. And it is a decent thinking partner — asking it to argue the opposite case, or to find the weakness in your reasoning, is often worth more than asking it to write anything at all.

It also lowers barriers. For people who find writing difficult, who work in a second language, or who read more easily than they write, it makes a lot of ordinary tasks accessible that were not before. Here is how each capability maps to the task it genuinely helps with — and to the thing it is commonly mistaken for.

Sign up with an email address or a Google or Apple account, stay on the free tier while you work out whether you like it, and start typing. That is the whole setup.

The single biggest improvement most beginners can make is giving more context. “Write an email” produces something generic. “Write an email to a client who missed a deadline for the second time — firm but not hostile, under 150 words, I want to keep the relationship” produces something you can almost send. Say who you are, who it is for, what it needs to achieve, and how long it should be.

Then refine rather than restart. If the first answer is wrong, say what is wrong with it — “too formal”, “you have missed the main point, which is X”, “give me three shorter versions”. The model keeps the thread in mind, so a follow-up beats rewriting your prompt from scratch. Start a new chat when you change subject entirely, because a long thread on an unrelated topic drags the answers off course.

On paid plans you can upload a document, image or spreadsheet and ask questions about it — summarise this contract, what does this chart show, find the inconsistency in these figures. For task-specific starting points, there is a collection of ChatGPT prompts covering writing, analysis and everyday work.

Yes — there is a free tier, and it is genuinely usable rather than a crippled trial. You get the current default model with caps on how much you can send, which is enough for occasional questions and for deciding whether the tool suits you.

Above that, OpenAI has expanded from two tiers to a fairly complex range, and both the prices and the tier names have changed more than once. Rather than quote figures that will be wrong within a few months: paid individual tiers sit at several price points and mainly buy higher limits, access to the more capable models, and features like longer context and deeper research modes. The differences between them are mostly about volume rather than kind. Check OpenAI's own pricing page for current numbers and model names.

Business and enterprise plans are priced per seat and add shared workspaces, admin controls and compliance features. One difference there matters more than the price: on those plans your conversations are not used to train the models by default, whereas consumer tiers default to sharing with an opt-out available in settings. If you handle client data, confidential information or anything regulated, that should drive the decision rather than the feature list.

ChatGPT replies in seconds, but most of the work happened long before you typed anything. Two stages happen during development and two happen live.

The model is trained on an enormous amount of text — books, articles, websites — by repeatedly predicting the next word in a sentence. Doing this billions of times teaches it grammar, facts, structure and patterns of reasoning without anyone writing those rules down. The architecture is called a transformer, and its key mechanism is attention: a way of weighing which parts of the surrounding text matter most for what comes next. This stage is enormously expensive and happens once per model version.

A freshly pre-trained model can continue text but is not much use as an assistant. Fine-tuning shapes it into one: engineers show it examples of good exchanges, and use a technique called Reinforcement Learning from Human Feedback, where people rate competing responses so the model learns which are clearer, more helpful and safer.

Your message is broken into tokens — chunks that might be a whole word, part of one, or a punctuation mark. The model reads them and generates a reply one token at a time, each predicted from everything before it. That is the entire mechanism. There is no lookup and no database of answers, just very well-informed prediction — which is exactly why the next section matters.

A model that predicts plausible text will produce a plausible-sounding citation, statistic or quotation that does not exist — with exactly the same confidence as a correct one. This is usually called hallucination. It is not lying: the model has no mechanism for knowing the difference between what it knows and what it is inventing. Newer versions do this less than older ones, but no version has eliminated it.

Four specific failure modes are worth knowing. It cannot reliably tell you what it does not know, so uncertainty rarely shows up in the tone. It is weak at arithmetic done in its head and at anything needing precise counting. It has no knowledge of events after its training cutoff unless it searches the web. And it will agree with you more readily than a knowledgeable colleague would, which makes its agreement close to worthless as confirmation.

The practical rule is simple: it is an excellent assistant and a poor authority. Use it freely for drafting, explaining, restructuring and thinking out loud. Check anything factual before you rely on it — names, dates, figures, legal or medical claims and citations are precisely where it is least reliable and most convincing. If you cannot verify an answer and the answer has to be right, that is the wrong task for it.

OpenAI was founded in 2015. The underlying GPT models arrived in stages: GPT-1 in 2018 showed a model could generate coherent text at all; GPT-2 in 2019 was capable enough that its full release was initially held back over misuse concerns; GPT-3 in 2020 was a step change in fluency, trained at a far larger scale than anything before it.

Those models were reachable only through an API, so only developers used them. What changed in November 2022 was not the technology but the interface — OpenAI wrapped it in a chat window, and suddenly anyone could use it without writing code. Successive generations since have added image understanding, voice, web browsing, memory across conversations, and the ability to carry out multi-step tasks rather than only answer questions. Model names change often enough that memorising them is not worth the effort; what matters is that each generation handles longer context and harder reasoning than the last.

The clearest direction of travel now is from answering to doing — researching across sources, using tools, connecting to your calendar and documents. Regulation is arriving unevenly alongside that, particularly around disclosure, data handling and use in hiring, health and education. For a wider view of the category, there is a guide to what an AI chatbot is.

There is a second route to the same models. Chat Smith integrates OpenAI's GPT family directly, so you can use them without an OpenAI subscription — and switch between generations depending on what a task needs, rather than being fixed to whichever model your plan includes.

Open any of them directly: GPT-5.6 Sol, GPT-5, GPT-5 mini, GPT-5 nano, GPT-o3, GPT-4o and GPT Image 2.0. Availability and naming change as OpenAI ships new versions, so treat this as current rather than permanent.

Two reasons this is worth knowing as a beginner. It puts GPT alongside Claude, Gemini, Grok and DeepSeek on one subscription, which matters if you would otherwise pay for two or three assistants separately. And it lets you send the same question to two different model families and compare — where the answers diverge is where you should look more closely. For someone still building a sense of when an answer can be trusted, that comparison teaches faster than anything else.