I would have disguised as a robot and sent them a picture of me.

Ironically sending a robot picture is the best proof that you're not one, since robots don't have a sense of humour (yet).

Wait; lightbulb moment; someone has to run a NN against a dad joke DB and learn AI to crack some (like the generated face stuff).

Well this is the internets and all, so I suppose it's been done before but no one would admit to owning such a big enough dad joke DB...

> someone has to run a NN against a dad joke DB and learn AI to crack some

A "that's what she said" detector can be implemented as a two class identification problem. This 2011 describes one algorithm that achieves 71% precision: That's What She Said: Double Entendre Identification https://www.aclweb.org/anthology/P11-2016

I believe they could have done an even better job with better training (and test) data. It seems to me there's a big problem with how they selected the negative examples: one of the sources is a set of racy text messages, which would be rife with positives. They claim TWSS positives are rare in general, but I doubt that holds for a sexy text messages corpus.

Their overall strategy is pretty interesting though (not deep learning).