What Is an AI Model? Explained Without the Jargon

A Model Is a File, Not a Program

An AI model is the product of training: a large file containing millions or billions of numbers that were adjusted, bit by bit, until the model got good at a specific task. Those numbers are called weights or parameters. When you send a prompt to ChatGPT, the app isn't running hand-written instructions like “if user says hello, reply with hi” — it's feeding your text through that giant file of numbers, which transforms it, step by step, into a response.

This is a genuinely different kind of software than most people are used to. A calculator app was written by a programmer who typed out the exact rules for addition. A language model's “rules” were never typed out by anyone — they were learned automatically from examples, which is also why nobody, including the companies that build them, can fully explain why a model produced one specific word instead of another. Researchers can inspect the numbers, but reading billions of individual weights doesn't translate into a human-readable explanation the way reading source code does.

Training: Where the Numbers Come From

Training adjusts a model's internal numbers, called weights or parameters, based on how well it predicts the correct answer.

Training starts with the model's numbers set to something close to random. It's then shown an example — say, the beginning of a sentence — and asked to predict what comes next. Its guess is compared to the actual next word, and the difference between the guess and the real answer is used to nudge every one of its billions of numbers very slightly in the direction that would have made the guess more accurate. Repeat that process across trillions of words of text, and the numbers gradually settle into a configuration that's genuinely good at predicting language.

That process, called pretraining, produces what's known as a base model. Most consumer products then run a second stage called fine-tuning, where the model is further adjusted on a smaller, more curated set of examples — conversations, instructions, safety guidelines — to turn a raw text-predictor into something that behaves like a helpful assistant instead of just continuing whatever text you gave it.

Words You'll See Attached to Every Model

“Claude” Isn't Really One Thing

A common source of confusion: “Claude,” “ChatGPT,” and “Gemini” are product names, not model names, and the two don't map one-to-one. ChatGPT the product can run on several different underlying models depending on which one you select in the interface. Claude the product is built around a specific model family (like Claude Sonnet 5), but the chat app around it — memory, file uploads, web search, the interface itself — is separate engineering built on top of the model, not part of the model itself.

This distinction matters practically: two products can use the exact same underlying model and still feel very different to use, because so much of what shapes the experience — system instructions, available tools, safety filters, interface design — sits outside the model file itself, in the application layer wrapped around it.

Not Every Model Does the Same Job

“AI model” covers a lot of ground beyond chatbots. Large language models (LLMs) like GPT and Claude handle text. Diffusion models like Stable Diffusion and Midjourney generate images. Classification models, often smaller and less discussed, power spam filters and fraud detection — predicting a category rather than generating new content. Embedding models convert text or images into the numerical format used for search and recommendation systems. Most of the AI products you interact with day to day are actually combinations of several different model types working together behind one interface, even when the product presents itself as a single “AI.” A photo app's “AI search” might use one small model to detect faces, another to read captions, and a third to rank results — all invisible to you, all bundled under one search bar. Even a single chatbot reply sometimes involves more than one model working together: a small, fast model checking your message for safety issues before the main model ever sees it, for instance.

A Model Isn't a Complete Product

A bare model file, on its own, can't browse the web, remember your last conversation, or run code — those are capabilities added by the software wrapped around it, often called tool use or agentic behavior. When ChatGPT searches the web for you, the model isn't reaching out to the internet directly; it's generating a request that a separate piece of software executes, then feeding the result back into the model to continue generating a response. Understanding this separation helps explain why the same model can behave very differently depending on which app or API it's plugged into — the model provides the reasoning, but the surrounding system decides what it's actually allowed and able to do.

This is also why the same underlying model can seem smarter in one product than another. A model connected to live web search and a code execution tool can verify facts and check its own math before answering; the same model with neither of those tools has to rely purely on what it learned during training, which is more likely to produce a confidently wrong answer on anything requiring current information.

Frequently Asked Questions