Building Safe A.I.

The gist of the long article:

"Homomorphic encryption is a special type of encryption that allows you to do computation on data and get an answer, all while keeping the data encrypted so that the computer doesn't know what the real data is.  Now, if you can homomorphically encrypt the neural network itself, then it starts to have some really interesting applications.  The most interesting of them is that if you built a general A.I. at some point that was homomorphically encrypted, all the output would be encrypted and useless to the network itself, or anyone who didn't have the key.  This could provide a control mechanism for a smarter A.I.":

https://iamtrask.github.io/2017/03/17/safe-ai/