Yes, I know. But the term used is not the most correct one in my opinion.
Simply put, when we have more bits/sample, this means that we have more levels for the quantizer. Imagine you have the real line part from -10 to 10, and you divide it to 2 equal segments, and each segment is represented by a point. This will have large quantization error, because say the representatives are -5 and 5, then any number between 0 and 10 will be taken as 5, and any number between -10 and 0 will be taken as -5. So, the maximum error here is 5. On the other hand, if we divide this line to 8 equally segments, the maximum error is reduced to 1. Asymptotically, as the number of levels (segments) goes to infinity the quantization error goes to zero.