How Does ChatGPT Work? A Plain-English Explanation (2026)

Here's the whole secret: ChatGPT predicts the next chunk of text, extremely well, again and again. That's it. Everything that feels like understanding, reasoning, or personality is built on top of that one deceptively simple engine — a system trained to answer "given everything so far, what's the most plausible next piece of text?"

Once you internalize that, ChatGPT stops being magic and starts being legible — including its weaknesses. Let's build it up in four plain-English steps.

Step 1: Your Words Become Tokens

The model can't read letters or words. It reads tokens — pieces of words, roughly four characters of English each. "Newsletter" might be one token; "Teahose" might be two or three. Your message gets chopped into a sequence of these tokens, each represented as a number.

This is why everything in the AI world is priced and measured in tokens, not words — including the context window, the maximum amount of text the model can consider at once (your prompt + the conversation + its reply). Run out of context and the model literally can't "see" the start of a long chat anymore.

Step 2: The Transformer Does the Thinking

The tokens flow through a transformer — the neural-network architecture (the "T" in GPT) introduced in 2017 that made all of this possible. Its breakthrough mechanism is attention: for every token, the model weighs how much every other token should influence it.

That's how it tracks that "it" refers to the company three sentences ago, or that a question at the end relates to a detail at the beginning. Attention is the difference between a model that loses the thread and one that holds a long, complex thought together. The model has billions of internal parameters — dials tuned during training — that encode the statistical patterns of language, code, and reasoning.

Step 3: Pretraining — Learning From a Huge Pile of Text

Before ChatGPT could chat, the raw model was pretrained: shown staggering amounts of text and made to play one game billions of times — predict the next token, check, adjust, repeat. Do that at enough scale and something remarkable emerges: to predict text well, the model has to absorb grammar, facts, styles, code, and reasoning patterns. Capabilities nobody explicitly programmed emerge from pure prediction at scale.

But a raw pretrained model is a brilliant, unhelpful autocomplete. Ask it a question and it might continue with more questions, because that's a plausible continuation. It needs manners.

Step 4: RLHF — Turning Autocomplete Into an Assistant

The polish that made ChatGPT a phenomenon is RLHF — reinforcement learning from human feedback. Humans rate model responses; the model learns to produce the kind humans prefer — helpful, on-topic, appropriately cautious. This is the layer that gives ChatGPT its assistant personality, its willingness to follow instructions, and its refusals. Newer systems add a reasoning step too — the model is trained to "think" through intermediate steps before answering, which sharply improves math, code, and logic.

So the full pipeline: tokens → transformer with attention → pretraining (raw capability) → RLHF + reasoning (helpfulness). Predict-the-next-token at the bottom; a helpful assistant at the top.

Why It Confidently Lies

Now the weaknesses make sense. ChatGPT is a plausibility engine, not a truth engine. It outputs the most likely continuation — and a fluent, confident, wrong answer is often more statistically plausible than an honest "I'm not sure." It has no internal fact-checker, which is why it can fabricate citations or dates that sound authoritative. Giving it real sources (retrieval) and tools helps, but the root cause — it predicts plausibility, not truth — never fully goes away. The fix is yours: verify anything that matters.

From Chatbots to Agents

ChatGPT is one product built on this engine — but the same engine now powers a whole field. Wrap a model in tools and a planning loop and a chatbot becomes an AI agent: software that doesn't just answer but acts — browsing, coding, completing tasks. Products like Manus are exactly this: the prediction engine above, given hands. The model layer is the foundation; the agent layer is what's being built on top of it right now.

Below is the live map of the companies building the foundation models that power ChatGPT and its rivals — ranked by how much they're moving in our signal feed:

Going Deeper

• The category above it: What are AI agents? — what happens when you give a model like this tools and a goal.
• The breakout agent: What is Manus? — a model orchestrated into an autonomous worker.
• The alternatives: ChatGPT alternatives — Claude, Gemini, Perplexity, and the rest, compared.
• The companies: OpenAI competitors · Anthropic competitors · the live LLMs theme.

The models behind ChatGPT change every few weeks — new versions, new rivals, new capabilities. The free Teahose daily digest tracks every model release, lab funding round, and research breakthrough across 40+ podcasts, 20+ newsletters, and the day's papers — one email each morning. Subscribe free and stop refreshing Twitter to keep up.

The explanation is stable; the live lab map is as of June 14, 2026, and updates continuously.

Frequently Asked Questions

How does ChatGPT work, in simple terms?

ChatGPT predicts the next chunk of text, over and over, very well. It breaks your message into tokens (pieces of words), runs them through a giant neural network called a transformer that has learned statistical patterns from enormous amounts of text, and outputs the most plausible next token — then repeats, feeding its own output back in, until the answer is complete. Everything else (its helpfulness, its tone, its refusals) is a layer of training on top of that core "predict the next token" engine.

What is a token in ChatGPT?

A token is the unit of text the model actually reads and writes — roughly a word or a piece of a word (about 4 characters of English on average). "Newsletter" might be one token; an unusual name might be three. The model never sees letters or words directly; it sees sequences of token IDs and predicts the next one. Pricing, speed, and the context-window limit are all measured in tokens, not words.

What is a transformer and what is attention?

The transformer is the neural-network architecture (introduced by Google researchers in 2017) behind ChatGPT — the "T" in GPT. Its key mechanism, attention, lets the model weigh how much every word in the input should influence every other word, so it can track context, references, and meaning across a long passage. Attention is why modern models handle long, complex prompts coherently where older approaches lost the thread.

Why does ChatGPT make things up (hallucinate)?

Because it is a plausibility engine, not a fact database. It generates the most statistically likely continuation of your text, and a confident, well-formed wrong answer is often more "plausible" than an honest "I don't know." It has no built-in sense of truth — only patterns. That's why it can invent citations, dates, or details that sound authoritative. Retrieval (giving it real sources), tool use, and careful prompting reduce hallucinations but don't eliminate the underlying cause.

What is the difference between GPT and ChatGPT?

GPT is the underlying model — the Generative Pre-trained Transformer that does the prediction. ChatGPT is the product: the app, the chat interface, the memory, the safety layer, and the tools wrapped around the model. OpenAI builds both, but they're different layers — and the same GPT models also power thousands of other apps through the API.