How Do AI Models Learn?

AI models learn by making predictions, measuring how wrong each prediction is, and adjusting their internal settings — called weights — to be slightly less wrong next time. Repeat that cycle an astronomical number of times, over an enormous amount of material, and you get a model that can write, summarize, translate, and reason in text. That is the whole engine; everything else is scale and refinement.

This article walks through the three stages of how a modern language model is made — pretraining, fine-tuning, and training from human feedback — and explains the consequences that matter to anyone who relies on AI answers: why a model’s knowledge stops at a cutoff date, why it can be confidently wrong, and why what it “knows” differs fundamentally from what it can look up.

First, what “learning” means for a model

When a person learns a fact, they can usually point to where it came from: a book, a teacher, an experience. A model learns differently: it does not store facts in a database or file away documents to consult later.

Inside a model are weights — vast numbers of numerical settings that determine how it responds to any input (see What Are Large Language Models?). Learning is the process of setting those numbers. During training, the model predicts, compares its prediction against the correct answer, and nudges the weights slightly in the direction that would have done better. Each nudge is meaningless; billions of them, applied systematically, sculpt the weights into a configuration capturing the patterns of language: grammar, facts, styles, arguments, the shapes of good answers.

A useful analogy is a musician practicing scales. The skill does not live in a notebook of memorized notes — it lives in the hands. A model’s knowledge lives in its weights the same way: as capability, not as a library it can search.

Stage one: pretraining — learning from a mountain of text

The first and largest stage is pretraining. The model is set a simple, relentless task: read text with some words hidden, and predict the missing or next words.

“The committee announced that the new bridge would open in ___” — predict. Be wrong. Adjust. Move to the next sentence. The material is text at a scale that is hard to picture: books, articles, websites, documentation, code — a substantial cross-section of what humanity has published, processed as an ongoing stream of prediction exercises.

Nobody teaches the model grammar rules, world capitals, or how a sonnet is structured. It must discover those regularities itself, because the only way to get better at prediction is to internalize them. Facts are absorbed the same way: a model that has predicted text about photosynthesis, contract law, and sourdough starters millions of times carries those patterns in its weights.

Two things about pretraining matter for everything that follows:

It ends. Pretraining uses material gathered up to a certain point. When training finishes, the weights are frozen — the model stops absorbing the world. That boundary is its knowledge cutoff.
It learns patterns, not truth. The training signal is “what text actually follows,” not “what is correct.” Material common in the corpus is learned strongly — scientific fact, popular misconception, or a phrase repeated by a thousand copied web pages. Frequency is not accuracy, and the model has no separate mechanism for telling them apart.

Stage two: fine-tuning — learning the job

A freshly pretrained model is a prediction engine, not an assistant. Ask it a question and it might simply continue with more questions, because in its training data questions were often followed by other text rather than answers. Fine-tuning turns the engine into a product.

The model is trained further on much smaller, carefully prepared sets of examples: questions paired with good answers, instructions paired with well-executed responses. The same mechanism applies — predict, compare, adjust — but the material is curated to teach a role: answer the question, follow the instruction, decline certain categories of request.

Think of pretraining as a widely read person who has absorbed a library but never held a job. Fine-tuning is the job training: same knowledge, now shaped into a way of working.

Stage three: human feedback — learning what people prefer

The third stage refines behavior using human judgment. In the best-known version, reinforcement learning from human feedback (RLHF), people are shown two responses to the same prompt and asked which is better — clearer, more accurate, more helpful, safer. Those judgments are distilled into a scoring system, and the model is trained further to produce responses that score well.

This stage is why modern assistants feel different from raw text predictors: they have been steered toward what human raters reward. It is also worth being clear-eyed about what it is. RLHF teaches the model what people prefer, which usually tracks quality but is not identical to it — a confident, well-formatted answer can be preferred over a correct but hedged one. Some of the assistant personality users recognize — and some of its failure modes, like unwarranted confidence — are built here.

Variations exist (feedback generated partly by AI systems, preference optimization techniques), but the principle is constant: a final training layer shaped by judgments about output quality, on top of fine-tuning, on top of pretraining.

Why a model can be confidently wrong

Now the consequences. A model trained this way generates the continuation that best matches the patterns in its weights. It has no step in which it checks its answer against reality, no flag that rises when it moves from solid knowledge to thin ice. If your question concerns something well covered in its training material, the patterns are strong and the answer is usually reliable. If it concerns something rare, recent, local, or obscure — a small business’s current hours, an event from last month — the patterns are weak, and the model may fill the gap with a fluent, plausible, incorrect construction.

Crucially, the style does not change when the substance thins. The model does not know what it does not know, because nothing in its training taught it to locate the edge of its own knowledge. Confidence in AI output is a property of the prose, not a measurement of the facts — a direct consequence of how the learning works.

Training-time knowledge vs. retrieval-time information

Everything above describes knowledge baked into weights during training. But many AI products add a second, completely different way of knowing: retrieval. When a system searches the web or a document collection at the moment you ask, pulls in relevant pages, and composes its answer from them, it is working from retrieval-time information — current, external, and in principle checkable against its sources.

The distinction explains a lot of otherwise confusing behavior:

Ask about a historical fact and the model answers from weights: instant, fluent, unverifiable from within the answer itself.
Ask about something that changed last week and a retrieval-equipped system can answer correctly if it finds the right sources — while a model relying on weights alone answers from a world that no longer exists, without knowing anything changed.
Correct a fact on your website and a retrieval system can pick up the correction when it next reads the page. The weights will not change at all — training is over. Whether a corrected fact reaches an AI answer depends entirely on whether the system retrieves, and what it retrieves.

How retrieval works is covered in the next pillar, How AI Systems Retrieve Information. For business owners, the training/retrieval distinction is the most useful idea on this site: you cannot retrain a model, but the information it retrieves is information you can publish, correct, and improve.

Misconceptions and limitations

“The model memorized the internet.” Mostly, no. Training extracts patterns; the weights are not a compressed copy of the corpus, and the model generally cannot reproduce its training material on demand. (Some memorization of frequently repeated passages does occur — memorization is the exception, pattern-learning the rule.)

“It keeps learning from our chats.” A deployed model does not update its weights from your conversation. Your chat may inform the product around it, and providers may use data to train future models under their stated policies — but the model answering you today has fixed weights and learns nothing tonight.

“More training data would fix the errors.” More data helps capability broadly, but the core limits are structural: prediction is not verification, frequency is not truth, and a frozen model cannot know a changing world. That is why retrieval and source quality matter as much as model quality.

“Fine-tuning put the knowledge in.” Knowledge overwhelmingly comes from pretraining. Fine-tuning and feedback shape how the knowledge is used — they teach behavior far more than facts.

The short version

A model learns the way a skill is learned: prediction, error, adjustment, at inhuman scale. Pretraining gives it the patterns of language and the world’s written knowledge up to a cutoff. Fine-tuning gives it a job. Human feedback gives it manners and preferences. None of the three gives it a live connection to reality — that connection, where it exists, comes from retrieval, and retrieval reads what the world has published. Which is why what is published about a business, and how accurately, decides what AI systems can say about it.