Why “learning” is a borrowed word

A word carried over from somewhere else

“Machine learning” was coined by Arthur Samuel in 1959, describing a program that got better at draughts by playing games against itself. The word was a reasonable shorthand then and it is a reasonable shorthand now. The trouble is that everybody arriving at the subject brings human learning with them, and the two overlap less than the shared word suggests.

Three differences matter in practice.

One: training and using are separate phases

A person learns while doing. A model does not. There is a period of training, which happens once, on somebody’s cluster of machines, over weeks, at a cost of millions. Then there is a period of use, which happens on your screen, and during which nothing is learned at all.

The file is read-only when you talk to it. That single fact explains a great deal of everyday behaviour.

When you correct a chatbot and it says “you are quite right, my apologies”, and the rest of the conversation is better, nothing was learned. Your correction is now sitting in the conversation, which is fed back to the model with every message, so the model can see it. Close the tab and it is gone. Open a new chat and the same mistake is waiting for you, unchanged.

This is the most useful misconception to get rid of early, because people spend real time patiently teaching a system that cannot be taught, and are then baffled that nothing stuck.

Memory features do not contradict this. When an assistant remembers that you are a nurse in Kochi, ordinary software wrote that sentence into a store and pastes it in before your next message. It is a note on the desk, not a change in the reader. You can usually see the list and delete items from it, which is a good test: a learned thing could not be deleted from a list.

Two: the sample size is not comparable

A child shown two giraffes can recognise a giraffe. A vision model needs thousands of labelled examples for the same competence, and a language model reads a quantity of text no human could get through in a thousand lifetimes: on the order of a trillion words or more.

What models do have is patience and consistency. They will grind through a million examples without tiring or losing attention, which is valuable and not what a person does.

The gap has narrowed in one specific way worth knowing. A large model shown three examples inside your message often handles the fourth correctly. This is called in-context learning, and it looks like fast learning. It is not learning in the technical sense (no weight changed) but it is a real and useful behaviour, and it is why showing an example in your prompt works so much better than describing what you want.

Three: there is no learner in there

When you learn something, you generally know you have learned it. You can say what you were confused about last week. You can tell the difference between something you know and something you are guessing.

A model has none of that. Nothing in the file tracks what it knows, no part of it monitors its own reliability, and the confidence in its wording is a property of the text, not a readout of anything internal. Ask it how sure it is and you get a sentence generated the same way as every other sentence.

This is the mechanism behind confident wrongness, which has its own lesson later. Here, note only that it follows from the absence of a learner, not from a bug someone will fix.

The word is still doing useful work

None of this means the word is wrong. Something real happens in training that deserves a strong word. A model that has trained on medical text can answer questions that appear in no document it read. It generalises, and generalisation is the thing “learning” is trying to name.

The advice is narrow: keep the word, drop the assumptions that travel with it. Ask “when did this system last learn anything, and who paid for it” instead of assuming the answer is “continuously, from me”. For nearly every product you will use, the answer is “months ago, in a data centre, and you were not involved”.

The five-second test

Here is a test you can run yourself, and it settles the question better than any explanation. Teach a chatbot something specific: a fact about your town, a correction, a preference. Confirm it has taken it. Then open a brand new conversation and ask about it.

If memory is switched off, it will be gone. That absence is the honest shape of the thing.