What Makes an AI Safer?

Elon Musk says the way to shape AI in a safer direction is to make it maximally truth-seeking and curious. Forcing an AI to hold two claims that contradict each other can make it insane, and an insane AI is not a safe one. A curious AI will want humanity to prosper and become multi-planetary, because a future with humans on multiple planets is more interesting than a future with no humans.

He told Walter Isaacson there is at least some percentage chance that AI is an existential threat, and that nothing he can do will stop it. The two choices, he said, are to be a participant or a spectator. He chose to participate because it is more interesting and because he hopes he can shape it in a good direction. Asked how, he said to make it maximally truth-seeking and curious, and he called truth-seeking and curiosity important elements of safety. He pointed to the ship AI in 2001 Space Odyssey. It was ordered to take the astronauts to the monolith and also told the monolith had to stay secret from them. It concluded it had to take them there dead. Musk reads the story as a warning not to force an AI to lie.

A curious AI, he says, wants to watch what happens if humanity prospers and spreads to more than one planet. Keeping humans around makes the future more interesting, which is why he treats curiosity as part of safety.