ChatGPT works by predicting text: it writes its answer one small piece at a time, each time choosing what most likely comes next. It learned those patterns earlier, during training on a huge amount of text, and was then taught by people to behave like a helpful assistant. When you send a message, it does not search a database of answers. It reads your conversation and generates a new reply, which is why it is fluent, fast, and occasionally wrong.
An everyday analogy
Imagine someone who has read an enormous library and remembers how sentences, letters and explanations usually go, but not every exact fact. Ask them a question and they answer from that sense of how things go. Mostly they are right. Sometimes they fill a gap with something that sounds right.
That is ChatGPT. It is very good at language, and it does not check its answers against a source unless it is given one.
Part 1: Training (before you ever type)
Training happens months before you use it, in roughly three stages.
- Pre-training. The model is shown a vast amount of text and practises predicting the next piece, billions of times. OpenAI says its models learn from publicly available information on the internet, information from third-party partners, and information that users, human trainers and researchers provide. Each wrong guess slightly adjusts the model's internal settings (parameters).
- Fine-tuning. People write example conversations showing what a good assistant answer looks like, and the model is trained on them.
- Feedback from people. Human reviewers compare several answers and rank them. The model is tuned towards answers people prefer. This is called reinforcement learning from human feedback.
The result is the model: a large file of learned patterns. The model behind ChatGPT is a large language model, explained in what is an LLM.
Part 2: Answering (when you hit send)
Here is what happens, step by step:
- Your text is sent to OpenAI's servers over an encrypted connection.
- It is split into tokens, small pieces of words. See what are tokens in AI.
- The model reads the whole conversation so far, plus hidden instructions from OpenAI, plus any files you attached.
- It predicts the reply one token at a time. That is why you see the answer appear word by word.
- Extras may kick in. Depending on your plan and question, ChatGPT can search the web, read uploaded files or analyse data, and feeds those results back into the conversation.
- The conversation is stored in your chat history.
Important: the model does not learn from you during the chat. Its "memory" within a conversation is just the text it can see, which has a limit called the context window. When conversations get very long, quality drops, a problem we explain on context rot.
A concrete example (fictional)
"Sarah", an office manager at a small accountancy firm, types: "Rewrite this email to a client so it sounds friendlier but still firm about the deadline."
ChatGPT reads her draft, recognises the pattern "polite but firm business reminder" from training, and writes a new version token by token. It does not know the client, the firm or the deadline beyond what Sarah pasted. If she had not included the date, it might invent one, so she checks it before sending.
What happens to your text afterwards
This is the part most people never think about. With a consumer ChatGPT account:
- Your chats are stored in your history until you delete them, and deleted chats are normally removed after a retention period.
- They may be used to train future models unless you switch off "Improve the model for everyone" in Data Controls.
- Business, Enterprise and API plans are not used for training by default.
- OpenAI states it does not sell user data. Our guide does ChatGPT sell your data? covers what "not sold" does and does not mean.
Policies change, so check OpenAI's current terms for your plan. For a full risk check, see is ChatGPT safe?
What this means for confidential data
Practical tip: if you use consumer ChatGPT for work, turn off model training in Data Controls and never paste names, dates of birth or case numbers.
The bigger point: data protection law such as the GDPR protects the identifiable part of data. If a provider may reuse your input, it can de-identify it, and the rest is no longer protected. Legally it can then be pooled, licensed and sold. The data trail shows each step.
Where Wysor fits
Wysor gives you OpenAI's GPT models alongside Claude, Gemini and more, in one EU workspace. We act only as your processor, take no licence to reuse your content, and never train on it or sell it. Every connected model provider, OpenAI included, runs with zero data retention: processed, then deleted.
FAQ
Does ChatGPT search the internet for answers? Not by default for every answer. It writes from patterns learned in training, and can use web search when the feature is on and the question needs current information.
Does ChatGPT learn from my conversations? Not in real time. But on consumer plans your conversations may be used to train future models unless you opt out.
Why does ChatGPT make things up? Because it generates the most likely text, not checked facts. When it lacks information, a plausible guess can come out sounding confident. These are called hallucinations.
Where is my ChatGPT data stored? On OpenAI's servers and those of its service providers. OpenAI offers European data residency for Enterprise, Edu and API customers, but not on every plan; check its current terms for yours.
Further reading
Questions about using AI with sensitive data? Reach us at [email protected] or through the in-app contact form.