The phrase 'machine learning' can sound almost mystical, but the underlying idea is a fairly concrete mathematical process. At its core, a machine 'learns' by adjusting internal numeric values until its outputs match a desired pattern as closely as possible.
Learning as Optimisation
Every machine learning model has a set of adjustable parameters, and training is really a search for the specific combination of parameter values that minimises errors across a dataset. This search is guided by an optimisation process, most commonly gradient descent, which nudges the parameters step by step toward better performance based on feedback from the loss function.
Learning From Examples, Not Understanding
It is important to be clear that this kind of learning is statistical pattern matching rather than understanding in the human sense. A model trained to recognise cats has not formed a concept of what a cat is the way a child does; it has instead learned which pixel patterns are statistically associated with images labelled 'cat' in its training data. This is why models can fail in surprising ways on inputs that differ meaningfully from their training data, even when the difference seems trivial to a human.
Despite this limitation, statistical pattern learning at a large enough scale has proven remarkably effective at producing systems that behave usefully across an enormous range of real-world tasks, from translating languages to recommending products, even without anything resembling true human comprehension.