A function built from layers of simple weighted units, tuned by training on data rather than programmed by hand.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Nobody writes the rules a neural network follows. Instead, a network is a stack of layers of simple numeric operations whose weights start random and are adjusted, example by example, until its outputs improve, a process called training. The behavior is learned, not coded, which is why nobody can point to the line of code where a model "knows" something.
Modern AI models are neural networks at extreme scale. Their learned, statistical character is the root of both their power (capabilities no one explicitly programmed) and their characteristic failures (confident errors no one explicitly programmed either).