It's hard to understand how a black box works.

You understand a system when your mental model can accurately predict what it will do in response to most inputs.

With a black box you don't know what state it's in at any given point and have to guess.

And as more things happen it drifts away from the previous state you thought it was, making it very hard to "catch up" your understanding of its state to what it's actually in.

A non black box is constantly releasing signals about what state it is, helping you synchronize your mental model to its current state.

Two things that make it hard to synchronize your mental model with how a system actually works:

1) A complex internal state (how many distinct states can it be in?)

You need to distinguish the current state from the other states it might be in, a task that scales in difficulty based on the number of possible states.

This is even harder when there are some states that are very rare, meaning even someone who has interacted with it a lot is unlikely to have experienced it.

2) Very few externally-visible signals of what state it's in.

The extreme of this would be a complex device with hundreds of input sensors, thousands of possible internal states, and only a single LED output.

(You could imagine an instruction booklet for a device as being a part of the signal that comes along with it as a complement.)

If you have lots of external signals then over time with enough experience users can build up a (perhaps intuitive, flawed) understanding of the internal state.

There's basically no way to bootstrap understanding of a system with very few external signals.