How Do AI Chatbots Like ChatGPT and Claude Actually Work?

You type a question, and a few seconds later a fluent, well-organised answer appears. It can write poems, explain tax rules, fix code and plan a trip. It is natural to wonder: Does it understand? Is it searching the internet? Where does the answer come from?

Here is how these systems work, without needing a maths degree.

The AI models behind modern chatbots are called large language models (LLMs). At their core, they do one thing: given some text, predict what comes next.

If you write "The capital of France is", the most likely next word is "Paris". The model has seen many sentences like that, so it assigns a high probability to "Paris".

Now repeat. After "Paris", predict the next piece, and the next. Chain enough predictions and you get paragraphs.

That sounds too simple to produce a useful essay, but the model has learned patterns about grammar, facts, reasoning styles, code and tone from a huge amount of text. Predicting well requires capturing a lot of that structure.

Models don't read letters or whole words directly. They split text into tokens, which are small chunks: a word, part of a word or punctuation. "Unbelievable" may be several tokens. A page of English is roughly 400 to 500 tokens.

Why this matters:

• Models have a context window, a limit on how many tokens they can consider at once, including your prompt and their answer.
• Cost and speed are often measured in tokens.
• Odd behaviour with spelling, counting letters or unusual words often comes from tokenisation.

LLMs use a neural network design called the transformer, introduced in 2017. A key part is attention, a mechanism that lets the model weigh which earlier words matter most for predicting the next one. In "The trophy didn't fit in the suitcase because it was too big", attention helps work out that "it" probably refers to the trophy.

The model is made of billions of numbers called parameters or weights. They are not rules a person wrote. They are values adjusted during training.

The model reads an enormous collection of text (books, websites, code and more) and repeatedly tries to predict the next token. When it guesses wrong, its parameters are nudged to do better next time. After trillions of such adjustments, it becomes good at continuing text.

At this point it can complete sentences but may not follow instructions well.

The model is then trained on examples of good question-and-answer conversations, so it learns to respond helpfully instead of just continuing text.

People (and sometimes other AI systems) rate or compare answers. The model is adjusted to prefer answers that are more helpful, honest and safe. This stage shapes the polite, assistant-like tone and the refusals to do certain harmful things.

Companies also add safety systems, filters and policies on top.

1. Your text, plus any earlier conversation and hidden system instructions, is converted into tokens.
2. The model predicts the next token, using probabilities.
3. It picks one (often with some randomness, controlled by a setting called temperature).
4. It adds that token and repeats until the answer is done.

That is why the same question can give slightly different answers each time.

This is debated. The model clearly captures a great deal of structure about the world, and can reason through many problems. But it doesn't experience anything, and it has no direct sense of truth. It works from patterns in text.

A practical way to think about it: it is a very capable text-producing system that is good at plausible, not guaranteed correct.

By default, no. The model's knowledge comes from its training data, which has a cutoff date. It does not look up things in real time unless it is connected to tools, such as a search engine, a calculator, code execution or your documents.

Many chatbots now combine the model with tools. When you see the assistant searching the web or running code, those are extra systems wired to the model.

The model may produce a confident but false statement, a fake quotation or a nonexistent source. This happens because it generates likely-sounding text. If it doesn't know, it may still fill the gap with something plausible.

Higher risk when:

• The topic is obscure or recent
• You ask for specific numbers, dates or citations
• The question assumes something false
• The conversation is very long

Tax rules, prices and news change. A model may not know about changes after its cutoff. For example, check rates and slabs on official sources, as in old vs new tax regime.

Training text contains human biases, and models can reflect them.

Long arithmetic, counting and strict logic can go wrong unless the model uses a tool such as a calculator or code.

Small changes in a prompt can change the answer. See writing better AI prompts.

• Explaining ideas at different levels
• Drafting and editing text
• Summarising long documents
• Brainstorming
• Writing and explaining code. See using AI at work and how to debug any bug
• Translating and rephrasing
• Acting as a patient study partner

• Verify facts, numbers, laws and medical or financial claims with trusted sources.
• Don't paste secrets: passwords, API keys, customer data or confidential company information, unless your organisation approves the tool.
• Check code by running and testing it.
• Be careful with images, voice and video tools, which can be used to create convincing fakes. See cybersecurity basics and UPI scams.
• Disclose AI use where honesty or rules require it.
• Keep thinking for yourself. Don't hand over decisions you must own.

Understanding the basics helps you use AI better and judge hype. The skills that stay valuable are asking good questions, checking answers and applying judgement in your field. If you are choosing what to learn, see in-demand skills in India and learning to program in the AI era.

There is no evidence that current chatbots have consciousness or feelings. They generate text based on patterns.

It does not change its core parameters during your chat. Some services may use conversations to improve future models depending on settings and policies. Check the privacy settings.

Safety training and policies prevent some harmful or restricted outputs. It can also refuse by mistake.

It depends on the task, cost, speed and privacy needs. Try more than one on your own tasks.

It will change many tasks and some jobs. People who combine domain skill with AI tools are well placed. Keep learning.

This article simplifies a fast-changing field. Details differ by model and company.