When A.I. does what you say, not what you mean
Note: this was originally published elsewhere.
Artificial intelligence can drive a car, buy/sell a stock, and perform healthcare tasks, amongst numerous other tasks as the technology grows and changes each day. At the end of 2016, The White House predicted that 3.1 million individuals who drive for work could see their jobs replaced by autonomous vehicles.
But A.I. isn’t going to take every job in the world just yet. In order to better understand that fact and the nature of what constitutes Artificial Intelligence, it’s worth taking a temporary to detour to examine not just its successes, but its failures, curiosities, and hiccups as well.
A list has been making its way around Science Twitter featuring — as one individual put it — “instances of AI doing what creators specify, not what they mean.” It is a delight. Here are some of our favorites.
1. A robotic arm trained to move a block to a certain point on the table opted in to move the table to the block instead.
2. AI was charged with evolving a creature built for speed — and so grew a really tall creature who generated high velocities by … falling over.
3. In one simulation where creatures required energy and giving birth cost no energy, one species evolved themselves into a lifestyle that consisted of them sitting around, mating, and then eating their children.
4. A four-legged creature trained to carry a ball on its back decided to instead drop the ball between its legs and wiggle across a floor.
5. A pancake-making robot decided that the best way to make a pancake would be to throw it as high in the air as possible to keep it as far away as possible from the ground.
6. A creature learns to bait an opponent into following it off a cliff, which gives the creature enough points to get an extra life — which is a strategy the creature then pursues in an infinite loop.
7. A robot awarded for touching a soccer ball learned to get to the ball whereupon it would touch the ball by vibrating as fast as possible.
