ChatGPT works by predicting the next word in a sentence. It does not look things up like a search engine. It generates answers word by word, based on patterns it learned during training. This simple idea powers one of the most useful tools of our time.
Millions of people in the UK use ChatGPT every day. Most of them have no idea what happens behind the chat box. This guide explains it in plain English. No maths, no code, just the concepts that matter.
The Short Answer
ChatGPT is a type of large language model. During training, it read huge amounts of text from books, websites and articles. It learned the patterns of how words follow each other in human language.
When you type a prompt, the model breaks your words into small pieces called tokens. It then predicts the most likely token to come next, then the next one, and so on. It repeats this process until it has built a full answer. That is the whole trick: very good guessing, done at enormous speed.
Step 1: Training on Vast Amounts of Text
Before ChatGPT could answer anyone, it had to learn language. OpenAI fed the model an enormous amount of text. This included books, articles, websites and other public writing. The exact mix stays private, but the scale is what matters.
The model did not memorise the text like a student memorising facts. Instead, it adjusted billions of internal settings called parameters. Each parameter is a tiny number that influences how the model responds. Together, they capture the structure of language: grammar, facts, reasoning patterns, tone and style.
Think of it like learning a language by reading. If you read enough English books, you start to know which sentences sound right and which do not. ChatGPT did that at a scale no human could ever match. It processed more text in training than any person reads in a lifetime.
Training also taught the model facts about the world. It knows that London is the capital of the UK and that water freezes at zero degrees. But it stores facts as patterns, not as entries in a database. That distinction matters, and we will return to it later.
Step 2: Predicting the Next Word
The core job of ChatGPT is simple. Given some text, predict the next word. That is all it does, one word at a time. Your prompt becomes the starting text. The model predicts the first word of the answer, adds it, then predicts the next one, and keeps going.
This is where the transformer architecture comes in. The transformer is the design that made modern AI possible. It lets the model weigh every word in your prompt when choosing the next word. If you ask about Paris, the model gives extra weight to words like France and capital.
A concrete example helps. If you type “The sky is”, the model knows from its training that “blue” is the most likely next word. If you type “The sky is blue because”, it predicts “of” or “the” as likely continuations. Each choice depends on everything that came before it.
Real answers involve far more context. The model considers your entire conversation, not just the last sentence. That is why it can answer follow up questions without you repeating the background. It carries the context with it, up to a limit called the context window.
To get the best from this behaviour, it helps to use ChatGPT effectively with clear, specific prompts. The better your input, the more accurate its predictions become.
Step 3: Human Feedback (RLHF)
Here is how RLHF works. Human reviewers read many of the model’s answers. They rank which responses are helpful, honest and safe. The model is then trained to prefer the answers humans ranked highest. It learns to refuse harmful requests, to say when it is unsure, and to format answers helpfully.
Training on text alone produces a model that writes well but does not behave well. That is why OpenAI added this second stage, called Reinforcement Learning from Human Feedback, after the initial training.
This is why ChatGPT often says “I am not sure” or offers a balanced view. That caution was trained in, deliberately. It is also why the model sometimes refuses certain prompts. The refusals are part of the same feedback system.
RLHF does not make the model perfect. Reviewers disagree with each other. Some biases still slip through. But without this step, ChatGPT would be far less useful and far less safe for everyday users.
Why ChatGPT Sometimes Gets Things Wrong
ChatGPT can be confidently wrong. It invents facts, cites books that do not exist, and gives wrong dates. This happens because the model predicts likely words, not true ones. If a false statement looks like a natural continuation, the model may produce it.
This behaviour is called AI hallucination. It is not the model lying. It is the model guessing with incomplete information. The prediction engine has no built in fact checker. Fluency and truth are separate things inside the model.
Three situations make mistakes more likely. First, obscure topics where training data was thin. Second, recent events beyond the model’s training cutoff, where it fills the gap with guesses. Third, questions that invite speculation, where it continues a pattern instead of admitting it does not know.
You can reduce the risk. Ask for sources. Break hard questions into smaller steps. Check anything important against a reliable source. For high stakes uses, treat ChatGPT as a smart draft writer, not as an authority.
Tokens and Memory, Explained Simply
ChatGPT does not read words the way you do. It splits text into tokens. A token is a chunk of text, often a whole word or part of one. For example, “unbelievable” might become two or three tokens. On average, one token is roughly three quarters of an English word.
Tokens matter for two reasons. First, every model has a token limit for each conversation. This is the context window. If your chat gets very long, the oldest messages fall out of memory. That is why ChatGPT sometimes forgets what you said at the start of a long chat.
Second, tokens matter for cost on paid plans and API use. More tokens mean more work for the model. Short prompts cost less than long ones. Keeping prompts tight is a simple way to stay within limits and get faster answers.
If you want a deeper dive into prompts themselves, our AI prompt engineering guide shows how to structure instructions so the model predicts the answer you actually want.
Frequently Asked Questions
Is ChatGPT just predicting words?
At its core, yes. Every answer is built one token at a time by predicting what comes next. But the predictions draw on a deep model of language, facts and reasoning patterns learned during training. That is why simple prediction produces such capable answers.
Does ChatGPT understand what it says?
It processes language with remarkable skill, but it does not understand in the human sense. It has no experience, no beliefs and no awareness. Treat it as an extremely advanced pattern matcher, not as a mind.
Why does ChatGPT forget things in long chats?
Every conversation has a token limit called the context window. When the chat exceeds it, the oldest messages are dropped from the model’s working memory. Starting a fresh chat or summarising key points keeps the conversation on track.
Can ChatGPT learn from my conversations?
Your chats do not retrain the base model. The model does not absorb new facts from talking to you. Some versions have a memory feature that saves preferences across chats, but that is a separate system, not the same as learning.
Is ChatGPT always up to date?
No. The model’s knowledge has a cutoff date from its training. Newer versions can browse the web for current information. If you need the latest facts, ask it to search the web or verify against a current source yourself.
Conclusion
ChatGPT works through three stages. It trains on vast amounts of text, predicts the next word using that training, and refines its behaviour through human feedback. Understanding this explains both its strengths and its limits.
The model is brilliant at fluent, structured language. It is weaker at guaranteed truth. Once you see it as a prediction engine rather than an oracle, you can use it with far more skill and far less surprise. Start with clear prompts, check important facts, and let the tool do what it does best.
