Essentially nobody uses JSON without a library, but tons of people (maybe even most people) use CSV without a library.

Part of the problem here is standards. There's a TON of encoding variations all using the same .csv extension. Making a library that can accurately detect exactly which one is correct is a big problem once you leave the handful of most common variants. If you are doing subfield encoding, you are almost certainly on your own with decoding at least part of your system.

JSON has just one standard and everyone adheres to that standard which makes fast libraries possible.