This is the module that does the most work. Ordinary software follows rules a person wrote. AI is shown examples and works out its own patterns. Almost everything strange about AI — the brilliance and the confident nonsense — falls out of that single swap.
Ordinary software follows rules somebody wrote down, step by step, the same way every time. If this, then that. A calculator does not get better at arithmetic the longer you own it. A spreadsheet will never notice a pattern you did not ask it to look for. That is not a failing — it is the entire design, and it is why you can trust a calculator absolutely.
Crucially, someone can always open that program and read why it did what it did. The logic is right there, in order, in a language a person wrote.
AI is built the other way round. It is never handed the rules. It is shown an enormous number of examples and left to work out its own pattern for what a good answer looks like, then applies that pattern to something it has not seen before.
Here is the difference made concrete. Take the sentence "not bad at all, honestly." A rule-based checker looking for negative words finds "bad" and files it as a complaint. A system that learned from real examples has seen that phrasing before, in that shape, and reads it correctly as praise. Nobody taught it that particular idiom. It simply turned up often enough in the examples.
Machine learning is the name for that method, and the honest analogy is teaching a child what a cat is. You do not hand over a checklist of feline characteristics. You point at cat after cat after cat, and one day they identify a cat they have never seen. Same method, running on a computer instead of a child.
This is where it gets genuinely strange. Nobody can point at a line of logic explaining why a trained model produced a specific answer. Its "reasoning" is spread across an enormous number of internal numbers, nudged during training. Researchers can study it, test it, and probe its behaviour — but no one hand-wrote it, so no one can simply read it back.
That is the deal. The flexibility that lets it handle an idiom nobody anticipated is the same flexibility that lets it be confidently, fluently wrong. You do not get one without the other. Every failure later in this course — the invented facts, the relentless agreement — comes out of this same trade.
It also predicts what these systems are weirdly good and bad at. Good at sarcasm, tone, and finishing a sentence naturally: all things that show up constantly in examples but are miserable to write rules for. Bad at basic arithmetic, unless it has been given an actual calculator to call, because it learned what sums tend to look like rather than how to add.
Ordinary software follows rules a person wrote and can always be read back. AI is shown examples and derives its own pattern, which is why it handles things nobody anticipated and why nobody can point to the line of code behind any given answer. That single swap is the source of both the brilliance and the confident nonsense.