There's a much similar way to get to the same destination as the author:

> If it can't surprise you, it's not intelligent

> If it can surprise you, it can do so negatively

Ergo, the only way to keep your AI from producing a surprising-in-a-bad-way output is to make it too dumb to surprise you at all. In practice, I feel like this means a lot of domain specific training, e.g. an airline-ticket-booking chatbot that is not a general-purpose chatbot who specializes in booking airline tickets, but rather a bespoke chatbot that can't do anything else.