Ask ten people for an example of machine learning and you will get ten answers about chatbots. It is a shame, because the clearest examples are older, duller and far easier to reason about — the spam filter that quietly learned your inbox, the map that reroutes you around traffic it has never seen before, the bank alert that fires on the one transaction in ten thousand that does not fit.
Those systems make a better teaching example precisely because nothing about them feels magical. They show the shape of the thing: a model that gets better as it sees more of the world, and worse the moment the world changes underneath it.
The spam filter is still the best example
Spam filtering is the canonical case for a reason. Nobody can write the rules by hand — the moment you ban a word, the senders misspell it. What works is showing a model a very large number of messages that people marked as junk, and a very large number they did not, and letting it work out which combinations of signals separate the two.
Notice what that requires. Not intelligence, but labels: millions of small human judgements, collected almost for free every time someone clicks a button. Most successful machine learning projects are, underneath, a story about where the labels came from.
What good examples have in common
Look across the systems that genuinely work in production and the same handful of traits keep appearing:
- The question is narrow. “Is this message spam?” is answerable. “Is this a good email?” is not.
- Being wrong is survivable. A misfiled message costs a click. That tolerance is what makes the problem suitable in the first place.
- Feedback arrives quickly. The system finds out it was wrong in hours, not quarters, so it can improve.
- The data is a by-product. Nobody set out to label a training set; it accumulated because people were using the product.
Where the examples stop working
The same traits explain the failures. A model that decides who gets a loan is narrow enough to build and nowhere near tolerant enough to be wrong casually. Feedback is slow and one-sided: you learn about the borrowers you approved and almost nothing about the ones you turned away. The data is not a neutral by-product but a record of past decisions, complete with whatever those decisions got wrong.
A model does not learn the world. It learns the data you happened to collect about the world, which is not the same thing.
A more useful question
Rather than asking whether a problem could be solved with machine learning — almost anything could, badly — it is worth asking what happens on the day the model is confidently wrong. If the answer is a shrug and a correction, you probably have a good candidate. If the answer involves a lawyer, you are looking at a decision that still wants a human attached to it.