Let's Peek Inside
Welcome back! In Level 1 we sorted out what AI is and what it isn't. Now we're going to do something that sounds harder than it is. We're going to look inside a chatbot.
You don't need any maths or computer skills for this. By the end you'll be able to explain, in a sentence or two, how a chatbot builds its answers. You'll also see why it can sound completely sure of itself and still be wrong.
Ready? Let's go.
A Very Clever Guessing Game
Have you ever typed a text message and noticed your phone suggesting the next word? You type "See you" and it offers "soon" or "tomorrow". Your phone is making a guess about what usually comes next.
A chatbot works on the same basic idea, just far bigger and far more capable.
When you ask a chatbot a question, it doesn't look up a ready-made answer in a filing cabinet. It builds the answer one small piece at a time. At each step it asks itself, in effect: "Given everything so far, what's the most likely next bit?" It picks one, adds it on, and asks again. Over and over, very quickly, until the answer is finished.
Those small pieces are often whole words, but sometimes they're just parts of words. People in the field call them tokens. You can simply think of them as "word pieces".
So here's your one-sentence version: a chatbot writes its answer by repeatedly predicting the most likely next word piece, based on patterns it has learned.
That's really the heart of it.
Where Do the Patterns Come From?
Good question. How does it know what's "likely"?
Before you ever type a word, the chatbot has been through a long learning stage called training. During training it's shown an enormous amount of text, such as books, articles and websites. That collection of text is called its training data.
By working through all that text, it picks up patterns. It learns how sentences usually flow, which facts tend to appear together, and what a recipe or a polite email usually looks like.
Here's an analogy. Picture someone who has read millions of letters but never stepped outside. They'd get very good at writing letters that sound right. But everything they know comes from what they've read.
And that leads us to two important limits.
Gaps and Old News
First, gaps. If something was barely written about, the chatbot saw very few examples of it. Think of the history of a tiny village, a rare medical condition, or a small local business. When the patterns are thin, its predictions get shakier.
Second, old news. Training stops at a certain date. Companies call this the training cutoff or knowledge cutoff, and some publish it. Anthropic's help centre, for example, lists the cutoff for each of its Claude models and warns that they may not know about events after it.
So if you ask about last week's news, a chatbot working only from its training may simply not know. Some chatbots can now search the web to help with this, but the underlying model still learned from a fixed snapshot of the past.
Why It Can Sound So Sure and Still Be Wrong
This is the part that surprises most people, so let's take it slowly.
Remember, the chatbot is predicting words that fit the pattern. It has learned what confident, well-written answers look like. So it produces confident, well-written answers. The smooth tone comes from the writing style it learned, not from the facts being checked.
When it doesn't really know something, it may fill the gap with something that sounds right. That could be a wrong date, a book that doesn't exist, or a made-up quote. People call this a hallucination. It's an odd word, but it just means "a confident-sounding answer that isn't true".
Why doesn't it just say "I don't know"? In September 2025, researchers at OpenAI published a paper that offered one explanation. Think of a multiple-choice test where a blank answer scores zero but a guess might score a point. You'd learn to always guess, right? The researchers argued that the way these systems are trained and scored often rewards guessing over admitting uncertainty in a similar way.
The companies are open about this. ChatGPT shows a small note saying it can make mistakes and that you should check important information. Gemini carries a similar line telling you to double-check it.
So here's the golden rule: a confident tone is not proof. It's just how the chatbot talks. We'll practise spotting mistakes properly in Level 5. For now, simply knowing why they happen puts you ahead of a lot of people.
One Small, Safe Thing to Try
You don't need a chatbot for this one. Open a new text message on your phone, but don't send it. Type "I think" and then keep tapping the middle suggested word, about ten times.
What do you get? Probably a sentence that sounds smooth but means very little. That's prediction without understanding, in miniature. A chatbot is vastly more capable, but the core idea is the same. Delete the message when you're done.
Key Takeaways
- A chatbot builds its answer one word piece at a time by predicting what's most likely to come next.
- It learned those patterns from a huge collection of text called its training data.
- Thin coverage of a topic, or events after the training cutoff, can lead to weak or out-of-date answers.
- A confident, fluent tone comes from the patterns it learned. It doesn't mean the facts were checked.
- Made-up but confident answers are called hallucinations, and even the companies tell you to double-check.