본문 바로가기

EDITORIAL/과학 :: Science & Tech

3 Reasons Why Robots Won't (Can't) Take Over the World

<1>

1 Robots, Will They Take Over the Earth?

1.1 Background

       It’s the year 2200, and before we knew it, robots have subjugated us under their rule. Humans are grazed as livestock are, fallen to be entertainment for especially bored robots. As it stands, the future of humankind is bleak and grim. . . or is it?

       The question of a robot takeover, uprising, or rebellion against humans has always been a concern, striking debate about whether the development of artificial intelligence is truly the right direction for humankind. Despite such concerns, robotics and AI have been ever-growing, and people have moved on from trying to stop development of AI to instead trying to create safe AI by subjecting robot behavior under controllable conditions. Perhaps one of the most popular of these rules are Asimov’s Laws, which are rules devised to protect humans from the advancement of AI technology:

  1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
  2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
  3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.

       As hopeful as these laws may seem, they have come under scrutiny for being too general and brittle. For example, imagine the human zoo: the robot deems that keeping all humans controlled in a zoo is beneficial because it removes them from any harm. In addition, all intents to leave the zoo is halted because leaving robot care is potentially harmful for humans. This absurd scenario is may seem perfectly fine under Asimov’s laws, but is clearly not desirable for us humans. Then, does that mean AI is uncontrollable all development of it should be stopped? Not really, and here are some reasons why.

       For robots to want to suppress the will of humans, they need to be able to do a few things: They need to 1) figure out that there exist loopholes in Asimov’s laws (like the zoo idea), 2) create a plan to for an uprising based on these loopholes (model the zoo), and 3) motivate itself to carry out this plan (move people to the zoo). Before these preliminaries are met, no robot uprising will be possible. In the subsequent sections, we will explore why these three preliminaries are hard to accomplish in AI, and persuade ourselves against a robot takeover.

 

1.2 Figuring Out Loopholes In Asimov’s Laws

       How will a robot figure out the loopholes in Asimov’s laws? How does it understand these sentences anyways? Under the hood, robots aren’t like us, and communicate in bits of 2 electricity rather than in well-formed English. What it does instead is associate words with their context. For example, it will know the definition of the word ’love’ based on all the romantic stories that use the word ’love’ in its prose. Therefore, the more a word appears in a specific context, the stronger it ties with it. Now consider a robot thinking of ideas to keep humans safe, as per Asimov’s laws. For each idea, it decides whether or not they are actually harmful to humans before considering them safe to do. Because this decision will be based on the context of the idea, if humans have already written about similar ideas and deemed them harmful, the robot see it as harmful too. Considering the amount of sci-fi novels that exist out there about robot rebellion, most absurd loopholes will probably already be covered and the job of finding a new loophole will be increasingly hard for the robot. This phenomenon is because of the nature of artificial intelligence: it follows the trend of data, and the more there is data, the lesser the chance of digression from it. As long as people write about different ways robots can bend Asimov’s rules, the harder it will be for robots to think of ways to do so.

 

1.3 Creating a Plan for an Uprising

       Even if a robot were to find a loophole that allows it to rule over humanity, to do so it must come up with a plan. This is even harder because let alone plan for the future, robots don’t have a good memory at all. To do future planning, an association between current and previous experiences must be made. Robots can do this in two ways: it can keep a small memory that it can update for on-the-go information, or it can keep a large memory keeping track of all past information. Although keeping a small memory works for short-term tasks, such a short horizon will never realize a plan for robot dominance. Then, the alternative must be taken, where the robot keeps a large memory of all past information. Yet while this is much more performant, for the robot this means that for every new experience it makes, it must compare it with all previous experiences to make a meaningful comparison between the memories. This pairwise comparison is computationally expensive, and scaling it to a very long horizon is a huge burden for a robot. For example, the current state of the art model for language generation would take 355 years to train with a single research-level GPU, and even this model only keeps track of 2048 words. A sophisticated plan for rebellion would probably take a lot more than that to write, and at that point training it would take forever by current standards.

 

1.4 Motivating the Robot for a Rebellion

       While the previous two sections argued against the robot rebellion because of computational intractability, this section will argue whether a robot would or wouldn’t dream of rebellion at all. The argument proceeds as follows: Robots are driven by following an objective, namely a goal. The goal stated by Asimov’s laws can be taken as ’not harming a human’, perhaps in a physical way. While it tries to not injure humans physically, the robot inadvertently harms humans in different ways, such as socially or mentally. However, if the robot could infer that such metaphysical harm existed, it wouldn’t have done what it did. This sort of approach can be formalized in a mathematical manner in what is called Inverse Reward Design (IRD). In IRD, a robot is expected to infer that there are hidden factors to an objective given by 3 a person. When a guess objective (which is limited in information content) is received, the robot understands that there is missing information, and tries not to venture outside of the guess. For example, we can tell a robot that walking in plain grass is good but forget to tell it that walking in lava is bad. The key point from this method is that the robot becomes risk-averse: It does not try to engage in behavior that may potentially lead to an unknown outcome, therefore in the example it avoids walking in lava. Tying this back to the robot rebellion, this means that under the IRD model, if a robot is only told how to do some particular goods for humans, it won’t try anything else because the outcomes are unknown. This does limit the capability of a robot, but at least is a viable way of demotivating robots from harmful behavior, thus potentially relieving us of the fear of a robot takeover.

 

1.5 Conclusion

       Truth be told, Artificial Intelligence still has a long way to go before we can even expect it to be anywhere sentient enough to think of ruling over humans. If it were to jeopardize our lives now, it would probably be because some human gave it a malicious objective rather than because of some mutation in its robotic brain. However, these concerns are very real, and it would be reassuring to know that such a potential robotic hazard is improbable. In any case, robotics and artificial intelligence are bound to become a large part of our world and we should be hopeful for a safe and beneficial human robot interaction rather than live in fear of a destructive and one-sided one.

 

<1> Free Images : cute, romantic, machine, blue, toy, wallpaper, lego, robot, android, wall e 5616x3744 - - 1169009 - Free stock photos - PxHere