Deception by Default

investigate what it means for an algorithm to lie by omission. We describe what we take to be the central mechanism for successfully deceiving one's reader/listener, the abuse of defaults. As we shall see, no matter what safeguards are put in place against outright lies, AGIs will have to be perfectly capable of circumventing these. We consider the design of algorithms capable of effective deception and show that the 'negative' skills such an algorithm relies on are not any different from the 'positive' skills needed for effective communication with humans in the first place, lending considerable support to the Orthogonality Thesis.