What it means
A neural network is built from units that each take several numbers in, combine them using learned weights, apply a simple nonlinear function, and pass one number on. Individually they do almost nothing. Arranged in layers and trained on data, they approximate remarkably complicated relationships.
Training works by comparison and correction: the network makes a prediction, the prediction is compared against the right answer, and the weights are nudged in whatever direction reduces the error. Repeat across enormous numbers of examples and the weights encode something useful.
The biological analogy is loose and worth not over-reading. Artificial neurons are vastly simpler than the real thing, and the learning procedure has no established biological equivalent. The inspiration was real; the resemblance is not.
Why it matters
This is the substrate under nearly all modern AI, and its shape explains the field's characteristic behavior: models are statistical rather than logical, they degrade gradually rather than failing cleanly, and their knowledge lives in millions of numbers rather than in readable rules.
In practice
The useful takeaway for non-builders is what it implies about failure: a neural network has no internal notion of "I do not know." It always produces an output, and confidence in the output is something you have to establish from outside the model.