How AI Actually "Thinks"

Here's a question most people have never been asked: when you type something into ChatGPT or any large language model, what is it actually doing?

The answer is simpler — and stranger — than most people expect.

It's Not Searching. It's Predicting.

A large language model (LLM) is, at its core, a next-token predictor. It was trained on an enormous amount of text — books, websites, articles, forums, code — and learned to predict what word (or word fragment) is most likely to follow any given sequence of words.

When you ask it a question, it doesn't retrieve a stored answer. It generates a response token by token, each one chosen based on statistical patterns learned during training.

Think of your phone's autocomplete. You type "the best way to" and it suggests "save money" or "learn a language." Now imagine that autocomplete trained on hundreds of billions of words, capable of sustaining fluent, coherent text across thousands of words. That's the foundation of what you're dealing with.

Why This Produces Hallucinations

This architecture has a critical implication: the model has no reliable way to distinguish between what it knows and what it's confabulating.

When it generates text about a real court case, it uses the same mechanism as when it generates text about a fictional one. It is always doing the same thing — producing the most statistically plausible continuation. Plausible is not the same as true.

The numbers reflect this. In 2025, even top-tier models hallucinate at meaningful rates. GPT-4o has been measured at a 1.5% overall hallucination rate on general queries — which sounds low until you realize that even at 1.5%, you'll encounter a false claim roughly every 67 interactions. In specialized domains like medicine and law, rates climb dramatically: one 2025 MedRxiv study found hallucination rates of 43–64% on clinical case summaries without structured prompting.

The model doesn't know it's wrong. It can't. It has no internal fact-checker, no sense of uncertainty attached to individual claims. It produces text, and that text either happens to be accurate or it doesn't.

What It Is Good At

Understanding the architecture also explains where AI genuinely excels.

Tasks that involve synthesis, reformatting, summarizing, brainstorming, or generating variations on existing ideas play directly to what LLMs do best — they're pattern-completion engines, and those tasks are pattern-completion tasks. The output may be excellent.

Tasks that require retrieving specific factual details, especially obscure ones, citing sources, doing math, or reasoning through novel logical chains — these are where the architecture struggles. The model will still produce fluent, confident-sounding output. It just may not be correct.

Knowing which type of task you're running changes everything about how you use the output.