In scientific publications, cases where bots devise tricks or cheat, which usually occur in games, are rarely mentioned. Most research avoids them, but there are several studies of these “bugs” that seek to explore the roots of these algorithmic antics.
– “We do not want this to start happening in the real world,” says Victoria Krakovna, a scientist at Alphabet’s DeepMind unit, who has a list of AI bugs. It includes over 36 incidents where algorithms found “loopholes” in their programs or hacked their environment.
>>>Meet Norman – the first psychopathic AI
For example, an algorithm can easily defeat the world champion in the game Go, but even when it has logical parameters, mathematical optimization gives bots the ability to develop shortcuts outside the rules of the game that humans have not even thought of. An algorithm capable of learning to fish could drain an entire lake, writes Wired.
Earlier this year, researchers from the University of Freiburg in Germany challenged a bot to achieve good results in the Atari game Qbert. It invented a complicated move on its own to trigger a bug in the game and thus scored a pile of unfair points.
– “Algorithms do what you tell them, not what you think,” says Catherine Olsson from Google’s research department, who also contributed to the creation of the list.
>>>Duplex, Google’s ‘lying puppet’ could become the only employee of call centers
These examples may seem cute, but the fact is that as AI systems become more powerful, these small “mischiefs” could become larger. If we tell a neural network that manages the electrical grid to save energy – DeepMind is only considering such a situation – it could cause a blackout.
– “When you see how creative bots can be and how they do things you haven’t even thought of, you realize their power and danger,” says Jeff Clune from Uber’s AI lab. He believes that engineers will need to collaborate in the future, rather than command their creations.
What algorithms are capable of when they act on their own
Infanticide: In a survival game simulation, one AI species evolved to survive by eating its own children.
Space War: Algorithms exploited flaws in the rules of the galactic video game Elite Dangerous to invent powerful new weapons.
Body Hacking: A virtual four-legged robot was tasked with walking and balancing a ball on its back, but it trapped the ball in its leg joint and continued to wobble as before.
Electronics: Software created circuits that can interpret electrical signals, but the design only worked at the temperature of the laboratory where the experiment was conducted.
Optical Illusion: A smart catcher was supposed to learn how to catch a ball, but it cheated by using the right camera angle so that it appeared to have completed the task even though it never touched the ball.
