Right now yes, but that will not always be the case. You can already (with some effort) get high quality results for incongruous juxtapositions like 'hamsters playing golf' or the like, and it's not hard to imagine bridging the different semantic gaps.

So we should discuss it as an ethical problem because it's going to become practical and affordable far sooner than most people think. I would bet money it's already been tried and maybe done successfully, but for obvious reasons not advertised.

Unless that input is newly created, why would using existing data (let's say, b the police, who collected it from various criminals in the past) be any more harmful than not using it?

Suppose the input is hand-drawn.

(This is legal in some countries, at least.)