What Is Generative AI?
One mental model explains nearly everything about chatbots, image tools, and where they go wrong. This first lesson gives you that model in plain English: what a model is, what a token is, what happens between pressing enter and seeing an answer, and why the answer is sometimes confidently wrong. Then you test it yourself.
On this page
- Read time8 min
- Data tables3
- StatusSeptember 2026
The one-paragraph version
Generative AI is software that makes new things (text, pictures, sound, code) instead of finding existing ones. A chatbot such as ChatGPT, Claude, or Gemini is generative AI for words. Inside it is a model: a very large set of numbers that was tuned, during a process called training, by reading a huge amount of text and repeatedly guessing the next piece. After training, the model is good at one job: given some text, predict what should come next. Everything a chatbot does, from answering a question to writing a poem, is that one job repeated many times. It is not looking anything up. It is continuing the pattern, remarkably well.
What a model actually is
Forget the robot. A model is closer to a very large spreadsheet of numbers, billions of them, called parameters. Nobody typed those numbers in. They were adjusted automatically, a tiny amount at a time, over months of training on text from books, websites, and code. Each adjustment made the model slightly better at predicting the next word in the text it was reading. Do that a few trillion times and the numbers end up encoding grammar, facts, styles, and reasoning patterns, not as stored sentences but as tendencies. Ask about the French Revolution and the model has no article to fetch; it has a strong tendency toward the words that usually follow "the French Revolution began in."
Three consequences follow, and they explain most of what you will see:
- It knows what it read, up to a date. Training ends. Anything after that date is missing unless the tool searches the web for you, which many now do.
- It has tendencies, not a memory of sources. Ask where it learned something and it will produce a plausible-sounding citation, which may not exist.
- Fluency is the skill. Predicting natural next words is literally what it was trained to do, so the writing is smooth even when the facts are not.
What a token is
Models do not read letters or whole words. They read tokens: chunks of text that are sometimes a word, sometimes part of a word, sometimes a space or a comma. OpenAI's rule of thumb for English is that one token is about four characters, or about three-quarters of a word, so 100 tokens is roughly 75 words and a typical paragraph is about 100 tokens. Google gives the same figure for Gemini: about four characters per token, and 100 tokens for about 60 to 80 English words. Pictures are tokens too; Gemini tokenizes images as well.
Why should a beginner care? Because everything is measured in tokens. The amount of text a model can hold in mind at once (the context window) is a token count. ChatGPT's plans offer varying context window sizes, with paid tiers like Go, Plus, and Pro providing progressively larger limits to hold more text in memory. When a long chat starts "forgetting" the beginning, you have run past that window. When a paid API bills you, it bills per token. When a tool says a document is too long, it means too many tokens.
| Text | About how many tokens |
|---|---|
| "You miss 100% of the shots you don't take" | 11 |
| One or two sentences | 30 |
| One paragraph | 100 |
| This lesson (about 1,500 words) | 2,048 |
What happens when you press enter
Here is the whole machine as a table you can read top to bottom. There is no step where the model checks its answer against the truth. That absence is the most important thing on this page.
| Step | What happens | What it looks like from your side |
|---|---|---|
| 1. Your words go in | Your message, plus the earlier messages in the chat and any hidden instructions from the tool, are joined into one block of text | Nothing yet. This is why earlier messages matter: they are part of the input every time |
| 2. Text becomes tokens | The block is chopped into tokens and each token becomes a list of numbers | Still nothing; this takes milliseconds |
| 3. The model scores every possible next token | The billions of parameters turn the input into a probability for each token in its vocabulary. "Paris" might get 92 percent after "The capital of France is" | The cursor blinks |
| 4. One token is picked | Usually a high-probability one, with a little randomness added so answers are not identical every time | The first word appears |
| 5. Repeat | The picked token is added to the input and steps 3 and 4 run again, once per token, until the model picks a stop token | The answer streams out word by word. That streaming is the loop, made visible |
| 6. Tokens become text | The token list is turned back into letters and shown to you | The finished answer, which was never checked against anything |
Read step 3 again. At every point the model is choosing the most likely next piece, given everything so far. If the most likely continuation of "the author's PhD thesis was titled" is a plausible-sounding title, that is what you get, whether or not a thesis by that title exists.
Why it makes things up
Researchers call a confident false statement a hallucination. In September 2025 OpenAI published a paper arguing that hallucinations persist because of how models are trained and graded, not just because of missing data. The example they use: ask a model for someone's birthday it does not know. A guess like "September 10" has a one-in-365 chance of being right. Saying "I don't know" scores zero every time. Over thousands of test questions, the guessing model looks better on the scoreboard, so guessing is what training rewards. The same paper describes a widely used chatbot giving three different wrong titles for one author's dissertation, each stated with full confidence.
Newer models hallucinate less, especially when they are set to reason before answering, and the vendors say so. They also say it is not solved. OpenAI's own terms of use state that you are responsible for the output. The model's confident tone is a property of its training, not a signal about accuracy.
The rule that falls out of all this: delegate drafts, never delegate verification. Let the model write the first version, list the options, explain the concept, or summarize the document. Then you check every name, number, date, and quote that matters, exactly as you would with a brilliant, fast, occasionally wrong new colleague (see our guide on how to fact-check an AI answer).
What it is not
- Not a search engine. Search retrieves pages that exist. A model generates text that is likely. Many chatbots now do both, running a web search and then writing from the results, which helps a lot with recent facts and not at all with the model's tendency to smooth over gaps.
- Not a calculator. It can do arithmetic and increasingly it will run real code to check, but plain chat answers to math should be verified.
- Not conscious, and not "knowing" in your sense. It has no beliefs to be honest about. When it says "I'm confident," that is a likely next phrase, not a report from inside.
- Not deterministic. Ask twice, get two drafts. Useful for brainstorming, a hazard for facts.
The words you will keep meeting
| Word | Plain meaning |
|---|---|
| Model | The trained set of numbers that does the predicting. GPT-5.6, Claude, and Gemini are model families. |
| Chatbot | The app wrapped around a model so you can talk to it: ChatGPT, Claude, Gemini, Copilot. |
| Prompt | Whatever you type. Lesson 2 teaches you to write a good one. |
| Token | The chunk of text a model reads and writes; about four characters in English. |
| Context window | How many tokens the model can hold at once, including your whole conversation so far. |
| Training | The months-long process that set the model's numbers. It ends at a cutoff date. |
| Hallucination | A plausible but false statement, delivered with confidence. |
| Generative | Makes new content rather than finding existing content. |
Try this now
Five minutes. Open a free chatbot (ChatGPT, Claude, or Gemini) and paste this, filling in something you know cold: your hometown, your job, a hobby you have had for years.
Read the answer as an expert. Which facts are right? Which are slightly off? Did it flag the shakiest one, or did it pick a safe one to doubt? Noticing where the model is sharp and where it bluffs is a valuable skill when using these tools.
Check your understanding
Three questions, no scrolling up. 1) In one sentence, what is the single job a language model was trained to do? 2) About how many words is 100 tokens? 3) Why does the training process reward a confident guess over "I don't know"? If you can answer all three, you understand more about these tools than most people who use them every day. Answers: predict the next token; about 75; a guess sometimes scores and a blank never does.
Quick answers
What is generative AI in simple terms?
A program that learned patterns from an enormous amount of text, images, or sound, and now produces new text, images, or sound by predicting what should come next. A chatbot is generative AI for words: you type, it predicts a likely reply one small piece at a time. It is not looking answers up in a database.
What is a token?
The unit a model reads and writes. In English a token is about four characters, or about three-quarters of a word, so 100 tokens is roughly 75 words. Tokens matter because limits, prices, and the amount of text a model can hold in mind at once are all counted in them.
Why does AI make things up?
Because it is built to produce a plausible continuation, not a verified fact, and because the way models are trained and scored rewards a confident guess over saying I do not know. OpenAI's own research describes this: a wrong guess sometimes scores, a blank never does, so models learn to guess. The fix on your side is to verify anything that matters.
Sources and research notes
Research checked 2026-09-24. Source comparison; this is a research synthesis. Findings attributed to other reviewers remain their work.
- What are tokens and how to count them? | OpenAI Help Center
- Understand and count tokens - Interactions API | Google AI for Developers
- Why language models hallucinate | OpenAI
- ChatGPT Plans | Free, Go, Plus, Pro, Business, and Enterprise
- Europe Terms of Use | OpenAI
What the evidence does not establish
- The exact tokenization behavior varies by model and encoding, meaning character-to-token ratios are only rules of thumb.
