2. Conceptual landscape of explicability

2.1 Paradoxes of transparency

Intelligent systems are often referred to as black boxes, which is a pejorative, just as opacity is a metaphor. Transparency seems to be the antidote to this flaw, but it can also mean invisibility, as in "this solution change is transparent to the user", or total visibility, as in access to software source code. Yet even the deep neural networks whose operation seems the most opaque often give full access to all the features (model structure and weight file) that enable their results to be reproduced. Explaining how they work, however, requires a real construction, like understanding uncommented source code. We no longer use the term transparency, which seems to us both ambiguous and irrelevant here, but note however that the argument of decoupling transparency...