Jump to a section
A chatbot predicts the most likely next words based on patterns in the vast amount of text it learned from. That's why it's brilliant at writing and explaining, but can sound sure when it's wrong, can miss recent news unless it searches, and can lose track of very long chats. Ask it to search and show sources for anything current, check the things that matter, and start a fresh chat when you change topic.
Idea one: it predicts the next words
The AI inside ChatGPT, Gemini and Claude is called a large language model. OpenAI describes it as something that learns patterns from large amounts of text from many sources, then predicts 'the most likely next piece of language based on context'.
Put simply, it writes an answer one small chunk at a time, each time choosing what's most likely to come next given everything before it. Those chunks are called tokens: common runs of letters that appeared often in the text it learned from.
OpenAI is clear that the model 'doesn't "know" things the way a person does'. IBM puts it the same way, calling these tools highly sophisticated pattern recognition tools. Hold on to that idea. It explains most of what follows.
What is ChatGPT? A plain-English guide for complete beginners
OpenAI Academy: AI fundamentals IBM: what is ChatGPT? Giles Thomas: what AI chatbots are doing under the hood
Idea two: it's trained twice
Training happens in two broad stages. First, pre-training: the model learns general patterns from a huge amount of text. At that point it can continue text, but it isn't much of a helper.
Then post-training, which OpenAI says helps the model follow instructions more reliably. Part of this, IBM explains, involves human trainers ranking the model's answers so it learns which kinds of response people find most helpful. The technique is called reinforcement learning from human feedback.
That second stage is what turns a text predictor into something that feels like a helpful assistant.
OpenAI Academy: AI fundamentals IBM: what is ChatGPT? Hugging Face: illustrating reinforcement learning from human feedback
Idea three: why it sometimes makes things up
If a model is predicting likely words, a likely-sounding answer and a true answer are usually the same thing. Usually. When they aren't, you get what people call a hallucination: something false, stated as fact.
Researchers at OpenAI have argued that the way these systems are tested rewards guessing over admitting uncertainty.
The habit that follows is simple: the more an answer matters, the more you check it. Our guide to checking AI answers shows how in a couple of minutes.
AI chatbots make things up. Here's how to check what they tell you
OpenAI: Why language models hallucinate (5 September 2025) Kalai and colleagues: Why Language Models Hallucinate (research paper, 4 September 2025)
Idea four: why it can miss the news
A model only learns from text gathered up to a certain date, its knowledge cutoff. Anything after that, it doesn't know unless it searches the web.
Most big chatbots can search now, but they don't always choose to. The Columbia Journalism Review described a major news story that ChatGPT dismissed hours afterwards with 'That didn't happen.' A couple of days later, the same question got the right answer, because this time it searched.
So for anything recent or that changes often, such as prices, opening times, rules or news, ask it to search the web and show its sources. Then glance at the sources.
Please search the web for this rather than answering from memory, and list the sources you used with their dates.Columbia Journalism Review: AI chatbots can search the web, so why don't they? (January 2026)
Idea five: why long chats go wobbly
Here's the surprising bit. The model itself doesn't remember your conversation. Each time you send a message, the app sends it the whole chat so far, and it starts again from that text.
There's a limit to how much text it can take in at once, called the context window. Go past it and, as one explainer puts it, earlier parts of the conversation 'effectively stop existing from the model's perspective'. That's why a very long chat can start ignoring something you said at the start.
Memory features, where an assistant remembers facts about you across chats, are a separate add-on built around the model. You control what they keep.
Custom instructions and memory: make AI remember what matters to you
Giles Thomas: what AI chatbots are doing under the hood TDWI: context windows in AI
What this means for you
None of this makes chatbots less useful. It just tells you where to lean on them and where to check.
| Because it… | …it's great for | …so remember to |
|---|---|---|
| Predicts likely words | Drafting, rewording, explaining, brainstorming | Check facts that matter |
| Was trained to be helpful | Friendly, patient answers to any question | Ask 'how sure are you?' and for sources |
| Learned up to a cutoff date | Things that don't change much | Ask it to search for anything recent |
| Rereads the whole chat each time | Building on what you've already said | Start a fresh chat when you change topic |
Is AI bad for the environment? An honest look at the energy and water
The bottom line
Think of a chatbot as a brilliant, well-read assistant who never says 'I'm not sure' unless you ask, and who skipped this morning's paper. Give it good instructions, ask it to look things up when they're current, and keep a human eye on anything important. Do that, and it's one of the most useful tools you'll ever have for free.
How to ask an AI chatbot a good question (the four-part trick)
AI jargon explained: a plain-English A to Z
Sources
These sources support the guide. Tool settings, prices and policies change, so dated source boxes show when those details were checked.
- Columbia Journalism Review, Tow Center: AI chatbots can search the web, so why don't they? (27 January 2026)
- OpenAI Academy: AI fundamentals (10 April 2026)
- IBM: what is ChatGPT? (19 April 2025)
- Giles Thomas: what AI chatbots are actually doing under the hood (29 August 2025)
- TDWI: context windows in AI (May 2026)
- OpenAI: Why language models hallucinate (5 September 2025)
- Kalai and colleagues: Why Language Models Hallucinate (research paper, 4 September 2025)
- Hugging Face: illustrating reinforcement learning from human feedback (9 December 2022)
- The Register: OpenAI says models are programmed to make stuff up instead of admitting ignorance (17 September 2025)
AI helps with research and drafting. Dan checks the sources and is responsible for what is published.
How we check