In plain words
You have met chatbots for years: the pop-up on a bank website that offers five buttons, or the airline bot that only understands "check status". Those older bots follow scripts. If your question isn't on the menu, they get stuck.
Modern AI chatbots such as ChatGPT, Claude and Gemini are different. They are apps built around a large language model, which writes a fresh reply to whatever you say, in almost any language, on almost any topic.
The surprising part is how they keep track. The model itself has no memory of the chat. Every time you send a message, the app sends the whole conversation so far back to the model, and it writes the next reply from all of it.
Everyday example
Passing the notebook
Imagine friends writing a story in one notebook, but every new writer has never seen it before. Each person reads every page so far, then adds the next line. The story stays consistent because the whole notebook is handed over every time. A chatbot works like that: each reply comes from rereading the entire chat.
Where the example breaks: Notebooks can be any length. A chatbot can only reread so much, its context window. In very long chats, apps trim or summarise older parts, so early details can get lost.
How it actually works
Under the friendly chat screen, an AI chatbot is a loop around a language model.
The app assembles the conversation
It combines a hidden system prompt, your earlier messages, the model's earlier replies and your new message into one input, split into tokens.
The model writes the next reply
The model predicts the reply one token at a time (next-token prediction). The text streams onto your screen as it is generated, which is why replies appear word by word.
Tuned to be a good conversation partner
A raw model just continues text. Chatbot models are further trained on example conversations and human feedback (RLHF) so they answer questions, follow instructions and decline harmful requests.
- First famous chatbot
- ELIZA, written by Joseph Weizenbaum at MIT in the mid 1960s. It matched keywords and turned your words into questions
- The ELIZA effect
- People's tendency to feel a program understands them, even when it is following simple rules
- Turning point
- OpenAI released ChatGPT as a free research preview on 30 November 2022
A chatbot's fluency is not proof it is right. Even ChatGPT's launch post warned it "sometimes writes plausible-sounding but incorrect or nonsensical answers".
Questions people ask
Is ChatGPT a chatbot?
Yes. ChatGPT is a chatbot app made by OpenAI. The model inside it is GPT, a large language model. The app adds the chat screen, history, memory and tools such as web search on top of the model.
What is the difference between a chatbot and an AI assistant?
The terms overlap. A chatbot is anything you converse with. An AI assistant usually also does things for you, like setting reminders, reading your email or booking a table. Most big AI chatbots now include assistant features.
How does a chatbot remember what you said earlier?
Within a chat, the app resends the whole conversation with each new message, so the model rereads it every time. Across separate chats, it only remembers if the app has a memory feature switched on.
Are chatbots always right?
No. They can state wrong facts with full confidence, which is called hallucination. Use them for drafting, explaining and brainstorming, and check anything important, such as health, legal or money matters, against a reliable source.
What was the first chatbot?
ELIZA, written by Joseph Weizenbaum at MIT between 1964 and 1967, is usually called the first famous chatbot. Its best-known script imitated a therapist by reflecting your words back as questions. It had no understanding at all.