Could a recurrent neural network predict what the redacted text looked like similar to how it can predict sentences from starts by training over newspapers and Wikipedia? Probably not useful like the above work, but it would be entertaining like Google Dream was.