The reason you cannot just look at one artificial neuron and see what it does. Networks pack far more concepts than they have dimensions by letting concepts share directions, on the reasonable bet that most are absent at any given moment. It is efficient and it is why interpretability is difficult.
Technical view
The representation of more features than dimensions by tolerating small interference between nearly orthogonal directions, viable because features are sparse. It explains polysemantic neurons that activate for unrelated inputs, and it motivates dictionary learning approaches that project into a higher-dimensional sparse space where features separate. Toy models reproduce the phenomenon and show how feature importance and sparsity determine which features get dedicated dimensions.
Related
Part of AI Terminology Index, section: Testing, Safeguarding and Running Models in Practice.