02 · How AI Chatbots Work (No Math, Promise) 🧠¶
You don't need to understand engines to drive a car, but knowing a little about how chatbots work makes you dramatically better at using them. In a few friendly pages you'll learn where the AI's knowledge comes from, why it sometimes invents things, why it forgets, why the same question gets different answers, and what that "thinking…" message means. Every tip later in this manual makes more sense once you've read this.
✅ Key Points & Steps
A chatbot is built on a model that was trained on a huge amount of text and then refined to act as a helpful assistant. It writes each answer piece by piece, predicting what should come next based on the patterns it learned.
- Training gives it broad knowledge up to a certain date (its knowledge cutoff); web search fills in recent information.
- It remembers within a conversation, but only remembers across chats if a memory feature is turned on.
- Because it predicts plausible text, it can sometimes state incorrect information confidently, so verify anything important.
📚 Step 1: It read (a lot)¶
Before a chatbot talks to anyone, its "brain" (the model) goes through training. Huge computers feed it a vast amount of text: books, websites, encyclopedias, forums, code, and more. Modern models also learn from pictures, audio and video.
It doesn't memorize all that text like a hard drive. Instead it absorbs patterns: how sentences flow, how ideas connect, how a recipe is structured, what a polite email sounds like, how a math proof goes, how Spanish maps to English.
📌 A useful picture
Think of it less like a library (where you look up the exact book) and more like a very well-read person talking from memory. Usually right, impressively broad, but capable of misremembering details. Keep that picture in your head and most of AI's quirks will make sense.
🔮 Step 2: It predicts the next word¶
Here's the secret at the heart of every chatbot: it writes its answer one small piece at a time, each time predicting what should come next.
You've seen a tiny version of this on your phone keyboard: type "Happy" and it suggests "birthday." A chatbot does the same thing, but it considers your whole question and everything said so far, and it's been trained on so much text that its "suggestions" add up to thoughtful paragraphs.
flowchart LR
Q["You: What should I<br/>cook tonight?"] --> P1["Try"] --> P2["a"] --> P3["quick"] --> P4["veggie"] --> P5["stir-fry…"]Those pieces are called tokens (usually a word or part of a word). You'll see that word around; now you know it just means "a chunk of text." Chatbot limits and prices are often measured in tokens.
Why this matters to you: because it's predicting rather than looking up, the AI is brilliant at anything that follows patterns (writing, explaining, summarizing) and weaker at things where one wrong detail matters (exact figures, quotes, rare facts). You'll learn to lean into the first and double-check the second.
🎓 Step 3: It was coached to be a helpful assistant¶
A model that has only "read the internet" would just continue your text in random directions. So companies do a second stage of training where people (and other carefully designed AI) show it thousands of examples of good assistant behavior: answering helpfully, admitting uncertainty, refusing dangerous requests, following instructions.
That's why each assistant has its own personality. ChatGPT, Gemini, Claude, Grok and Copilot were coached differently, so they have different styles: chattier, more formal, funnier, more cautious. None is "correct." It's partly a matter of taste, which is why Part II helps you find your favorite.
💡 “Why won't it answer?“
Sometimes an assistant declines something harmless because it looked risky. Just add context: "I'm a nurse asking about safe medication doses for a patient handout" or "This is for a mystery novel I'm writing." Most refusals of innocent requests disappear with a sentence of explanation.
📅 What it knows, and what it doesn't¶
Training happens at a certain point in time, so every model has a knowledge cutoff: a date after which it knows nothing on its own. Ask about last night's game and a model without web access will either say it doesn't know or, worse, guess.
That's why most assistants now have web search built in. When it searches, you'll usually see a "Searching the web…" message and little source links in the answer. Those links are gold: click them to check the facts.
| Situation | What happens |
|---|---|
| Timeless question ("How do volcanoes work?") | Answers from its training. Usually great |
| Recent event ("Who won yesterday?") | Needs web search. Look for source links |
| Your personal stuff ("When's my dentist appointment?") | Only works if you've connected your calendar or pasted the info |
| Very niche fact ("my town's recycling rules") | Search helps, but always double-check with the official source |
💭 Memory: what it remembers (and forgets)¶
Chatbots have two kinds of memory, and mixing them up confuses lots of beginners:
- Conversation memory: within one chat, the AI can see everything you and it have said so far. That's why you can say "make it shorter" and it knows what "it" is. But each chat has a size limit (called the context window). In a very long chat, the earliest parts can get fuzzy or dropped.
- Long-term memory: many assistants (ChatGPT, Gemini, Claude, Copilot, Le Chat and others) now have an optional memory feature that remembers facts about you across chats: your name, that you're vegetarian, that you prefer short answers. You can view, edit and delete these memories in settings.
💡 Beginner superpower: start fresh
If a chat goes off the rails or gets super long, start a new chat and paste in a short summary of what matters. Answers often get noticeably better. It's the AI version of "have you tried turning it off and on again?" 🔄
🎲 Why the same question gets different answers¶
Chatbots add a sprinkle of randomness when choosing words, which keeps them creative and natural-sounding. Ask "give me a name for my cat" twice and you'll get different names. That's a feature!
Every assistant has a regenerate (🔄 "try again") button. Use it freely:
- Didn't love the answer? Regenerate.
- Want options? Ask for "5 different versions" in one go.
- Need consistency (like the same format every week)? Give clear instructions and an example, which you'll learn in Prompting 101.
🤔 "Thinking…": fast models vs. reasoning models¶
Most assistants now offer two styles of brain:
| ⚡ Fast mode | 🤔 Thinking / reasoning mode | |
|---|---|---|
| Feels like | Instant reply | "Thinking…" for a few seconds (or minutes) |
| Best for | Everyday chat, quick questions, writing | Math, logic, planning, tricky comparisons, coding |
| Names you'll see | "Instant," "Fast," "Flash" | "Thinking," "Reason," "Expert," "Deep Think," "Extended thinking" |
Many apps now pick automatically. If an answer to a hard question seems shallow, look for a "think" or "reason" toggle and try again. And if you see a little expandable "thought process," you can peek at how it worked it out. It's fascinating. 🧩
🤷 Why it sometimes makes things up¶
Put everything together and you can see why AI hallucinates (confidently invents things):
- It predicts plausible words rather than retrieving verified facts.
- It was trained to be helpful, so it leans toward giving an answer.
- It doesn't know what it doesn't know very well, especially about rare or recent things.
Modern assistants hallucinate much less than early ones, especially when they search the web, but it still happens. The fix is simple habits, which get a whole chapter: When AI Gets It Wrong.
flowchart TD
A[Your question] --> B{Does it need<br/>exact facts?}
B -- No: ideas, writing,<br/>explaining --> C[✅ Enjoy the answer]
B -- Yes: numbers, dates,<br/>health, money, law --> D[🔍 Ask for sources<br/>and double-check]🎯 Key takeaways¶
- A chatbot's model learned patterns from a huge amount of text, then was coached to be a helpful, polite assistant.
- It writes answers by predicting one chunk (token) at a time, which makes it great with words and shaky on exact details.
- It has a knowledge cutoff; web search and source links fill the gap.
- It remembers within a chat (the context window) and, if you turn it on, across chats (memory).
- Randomness means different answers each time: use regenerate and ask for options.
- Thinking modes help with hard problems. Hallucinations happen, so check anything important.
🧠 Check yourself¶
❓ 1. Does a chatbot look up answers in a giant encyclopedia inside its head?
No. It predicts answers word by word based on patterns it learned. That's why it's fluent and flexible, and also why it can be confidently wrong.
❓ 2. You ask about last night's big news and get an outdated answer. Why, and what do you do?
The model has a knowledge cutoff. Ask it to search the web (or turn search on), then check the source links it shows.
❓ 3. Your long chat has started to ignore things you said at the beginning. What's the easy fix?
Start a new chat and paste a short summary of what matters. Long chats can overflow the context window.
❓ 4. When should you switch on a “thinking“ or “reasoning“ mode?
For tricky tasks: math, logic, planning, detailed comparisons, coding. For quick everyday questions, fast mode is fine.
🎮 Try this
In any chatbot, ask the same question three times using regenerate: "Suggest a fun name for a golden retriever puppy and explain why." Notice how the answers differ. Then ask: "What's your knowledge cutoff date, and can you search the web?" Now you know your assistant a little better. 🐶
Progress is saved in this browser only, so no account needed. · 📘 Enjoying the sample? Get the whole book