Researchers show that AND, OR, and NOT gates can be embedded in neural networks using fixed weights. These gates can be combined into boolean circuits that read model activations, enabling inspection or manipulation of internal computations.