How AI Actually Works: A Simple Explanation Anyone Can Understand

Almost every confusing thing about AI dissolves once you have one idea, and the idea takes about a minute to state.

These systems predict what comes next, based on patterns found in enormous amounts of data.

That is it. Not thinking, not understanding, not reasoning in any sense you would recognise. Prediction, at a scale that makes it look like the other things.

Everything below is that idea applied. And the reason it is worth ten minutes is not curiosity — it is that it tells you precisely where these tools are reliable and where they fail in ways that are easy to miss.

Start With Spam

Spam filtering is the least glamorous AI in existence and the clearest illustration of the whole mechanism.

Nobody writes rules for it any more. Instead you take an enormous pile of emails already labelled spam or not spam, and let a system work out what distinguishes them. It finds patterns — certain words, certain sender behaviours, certain structural quirks — that correlate with the label.

Then a new email arrives and it asks: given everything I have seen, how likely is this one to be spam?

Notice what is absent. The filter does not know what spam is. It has no concept of deception or commerce or annoyance. It has correlations between features and labels, and that is enough to work extremely well.

Hold onto that gap. Every AI system has a version of it.

The Same Thing, Scaled Up

Language models do the same operation on a harder problem: given all the words so far, what word comes next?

Start with "The capital of France is" and the training data makes "Paris" overwhelmingly likely. Output it, add it to the context, and predict again. And again. That loop, running one word at a time, produces every essay, every explanation, every piece of code you have ever seen a chatbot generate.

The part that surprised everyone — including the researchers — is what doing this well requires. To predict the next word accurately across billions of examples, a system has to implicitly learn grammar, facts about the world, the shape of arguments, tone, register, and the structure of code. Those capabilities were not designed in. They emerged because they are all instrumentally useful for prediction.

That is genuinely the whole architecture. A mechanism simple enough to explain in one sentence, producing behaviour complex enough that people argue about whether it is thinking.

What Training Actually Is

Three stages, and knowing them explains several things people find mysterious.

Pre-training. The model reads a vast quantity of text, repeatedly predicting the next word and adjusting itself when wrong. This is where nearly all the capability comes from, and it takes months and enormous compute. What emerges is capable and unfocused — good at continuing text, with no particular sense of what you want.

Fine-tuning. Additional training on examples of good responses, teaching it to behave like an assistant rather than a text-continuation engine.

Human feedback. People rate outputs, and the model is adjusted toward what they preferred. This is where helpfulness and tone come from — and, worth knowing, where the tendency toward agreeableness comes from too. It was optimised to produce responses people rated well, and people rate agreement well. That is why you should explicitly ask it to argue against its own answer.

Everything after pre-training is shaping, not knowledge. The knowledge was set when training data was fixed, which is why a model has a cutoff date and cannot learn from your conversation.

Why It Invents Things

This is the most consequential consequence, and it follows directly.

The model generates text that is statistically likely given the context. It has no separate mechanism checking whether that text is true. Producing a well-supported fact and producing a convincing fabrication are, internally, the same operation.

So when it does not have solid information — an obscure topic, a specific figure, a citation — it does not stop. It generates what such an answer would plausibly look like. The output is fluent, well-formed, correctly structured, and wrong, and it arrives in exactly the same confident register as everything else.

This is not a bug being patched. It is what next-token prediction does. The practical rule: anything you would need to cite a source for should not come from here. Why AI fails goes into the failure modes in detail.

Three Misconceptions Worth Dropping

"It learns from our conversation." It does not. The weights are frozen after training. Within a chat it can refer back to what you said because that text sits in its context window, but close the conversation and it is gone. Memory features work by storing text and re-inserting it, not by changing the model.

"It looks things up." By default, no. It generates from patterns learned in training. Some products add search — which is a genuinely different mechanism and why citation-first tools like Perplexity behave differently from a bare chatbot.

"Bigger is always better." Scale drove enormous gains and the returns have flattened. Smaller models with better training now match much larger ones on most real tasks, which has done more to change what is economically possible than any frontier release.

What This Buys You

One distinction, and it is worth the whole article.

Transformation is reliable. Summarising a document you provided, restructuring your notes, translating your text, rewriting your draft, explaining a concept it saw ten thousand times in training. Everything needed is present. Nothing has to be recalled or invented.

Recall is not. Specific figures, dates, citations, what a named company did last quarter, anything niche. Here it is generating what a plausible answer looks like, and you cannot tell the difference from the output.

Applied consistently, that single rule prevents nearly every serious mistake people make with these tools. It is why using ChatGPT for work comes down to a division of labour rather than a list of tricks.

Where To Go From Here

If you want the terminology sorted out, AI vs machine learning vs deep learning untangles three words people use interchangeably and should not. For what is under the hood of a chatbot specifically, what a large language model is goes a level deeper, and deep learning and neural networks a level deeper still.

For the same story in other domains, natural language processing and computer vision both describe fields that spent decades stuck and got past it the same way.

And if you want to actually get good at using this, learning AI from scratch lays out a realistic path — which starts, usefully, with exactly the idea at the top of this page.