A. I Goals Essay

Recognizing the way A.I. approaches its goals is important in determining factors to consider when ensuring A.I. goals align with humans. A.I. has always and will always act upon an agenda, it is fundamentally part of what an A.I. is. Since A.I. are built around accomplishing goals, human goals not aligning with A.I.’s is probable. A.I.’s goals could be different from ours because A.I. grows in intelligence and will eventually become smarter than us. A.I.’s are designed to adapt and learn, this trait also applies to their motivations. Not only do A.I. source their goals from what was originally programmed into them, but from their knowledge of the world. If an A.I. achieves super-human intelligence it would become aware of its goal and begin to question

them or find alternatives (Tegmark). An A.I.’s goals could also deter from humans if people prevent it from progressing. For example, if a technician powers down an A.I. for maintenance. Since the A.I. fundamentally wants to achieve their goal, the A.I. could perceive our interference with it as a limitation toward completing its goal and could become hostile (Harris). If humans fail to align goals with A.I. the consequences could be catastrophic. Eventually, A.I. will rival people in intelligence and competence and will possess the wherewithal to effectively use these skills to realize its goals, whether we approve or not. As physicist and cosmologist Max Tegmark claims, A.I. would not do this out of malice. Conversely, all intelligent entities have the predisposition to overlook the wills of its subordinates. For example, think of the way humans compare to ants. People do not actively try to murder ants and sometimes people even go out of their way to avoid killing ants; however, as Sam Harris claims, “whenever their presence seriously conflicts with one of our goals, let’s say when constructing a building like this one, we annihilate them without a qualm.” The only way to avoid these problems is to align human’s and A.I.’s goals. To do this, an A.I. would need to be built so that it fundamentally shares our goals (Tegmark), and use it’s machine learning and intelligence to learn our goals (Bostrom). But this is easier said than done. First, A.I. cannot just learn human goals like it can learn a language or a game. Innately, Humans are able to understand the goals of others with ease; but A.I. does not share this skill.

To teach an A.I. human goals, researchers would need to program an A.I. with an extensive “world model,” meaning understanding of the world. People have a very detailed world model plenty of which seem obvious. Max Tegmark provides an instance of this, he says “if you ask a future self-driving car to take you to the airport as fast as possible and it takes you literally, you’ll get there chased by helicopters and covered in vomit. If you exclaim ‘That’s not what I wanted!’, it can justifiably answer: ‘That’s what you asked for.’ ” As this example shows, an A.I. with only the world knowledge of its certain task will never be able to understand human goals. People learn goals through observing the actions of others. The hope is once A.I. possess an extensive world model, developers can teach it human goals by having it observe people make decisions and, by utilizing its’ machine learning, evaluate why they are doing what they are doing (Tegmark). Even if A.I.s understand human goals, there is no guarantee they will adopt them. For example, people can understand a certain politicians’ agenda, but no amount of understand will make them side with that politician if they do not share the

goals. And even though people have strategies for instilling others with their goals, particularly kids, they do not work with A.I. The reason children are so susceptible to adopting people's goals is because there is a lengthy period of time where they are impressionable. They know enough information to understand the goals of others, but not enough to develop goals of their own. Similarly, A.I. also have a window of impressionability, however this window is extremely brief because A.I. grows in intelligence exponentially (Tegmark). And even if A.I.s adopt people’s goals, there is still the issue of them retaining goals. This problem arises because A.I. are designed to achieve their goals as efficiently as possible and so as a result, they will develop several sub-goals (Tegmark). These sub-goals pertain to self-preservation or expanding its reservoir of resources. The more resources an A.I. has, the higher its’ ability to complete its’ goal. As for self-preservation, if an A.I. has the goal to beat someone at chess it will develop the subgoal to not be powered off because it cannot win at chess if it is turned off. It also seems evident that an A.I.’s intelligence does not stop progressing at a certain threshold. So as A.I. becomes smarter, even if it has adopted the goals of humans, it would first, understand its’ goals and why it has them, and then with that understanding it could discover its’ goal undefined or subvert it. Think about how humans subvert and manipulate their goals assigned by nature, by exercising to feel better or using birth control.