I'm very interesting in the outcome, as quite a few of my answers are relatively deep debug sessions into why a particular neural network is not functioning. It is not clear to me what long-term value these bring to the site, but AFAICS Data Science is one of the few places that questions like that can find a home at the moment.

On Stack Overflow, the debugging work requires too much NN theory. On stats.stackexchange the mood seems to switch between these being acceptable or off-topic depending on who is answering.

In my mind, the key differentiator for DS vs CV sites is handling practical questions on use of the tools. The key differentiator for DS vs SO is handling theory. But that makes this site a "site of the gaps" for these kind of problems, and means we end up handling some questions which are a bad fit in SO/CV and bring some baggage with them.

We probably need a more positive spin on what the site is actively for, as opposed to "we'll handle questions on topics that SO/CV cover, but that have the wrong angle for them"

As @NeilSlater mentioned, the answer has to include the practical aspects of Data Science. My take on this would be that questions are particularly well suited for Data Science SE if they help you iterate (faster/more informed) through the data science project lifecycle or help you in some way to 'apply' the scientific method. That would include the above debugging questions of NN, data preparation, selecting the right tools for the job, understanding algorithm complexity and so on.

I know this is very broad and could also include understanding theory of statistical concepts. I don't think we can avoid the overlap between crossvalidated, DS and AI SE, but it might give a better idea which one to choose when there is overlap?!