Scientists Want to Teach Robots Morality by Reading Them Bedtime Stories

Thursday, 18 February 2016 - 1:19PM
Thursday, 18 February 2016 - 1:19PM
Scientists Want to Teach Robots Morality by Reading Them Bedtime Stories
When and if robots ever achieve a humanlike intelligence, they will potentially have processing power that far exceeds our own, an idea that has spawned countless robot apocalypse books, movies, TV shows, and conspiracy theories. With artificial intelligence making strides every day, researchers have become concerned with ensuring that AI has the ability to be moral, or at least, to follow Asimov's Three Laws of Robotics. Now, researchers have proposed a way to teach robots culturally accepted moral values--by reading them fables.

This study, led by Mark Riedl, director of the Entertainment Intelligence Lab at the Georgia Institute of Technology, hinges on the theory of "value alignment," which refers to the property of AI that dictates it can only pursue goals that will not be harmful to humans. Value alignment, Riedl and his team assert, is key to safeguarding against unintended consequences of superintelligent AI (or an intentional robot apocalypse). 

Opening quote
Recent advances in artificial intelligence and machine learning have lead many to speculate that artificial general intelligence is increasingly likely. This new, general intelligence may be equal to or greater than human-level intelligence but also may not understand the impact that its behaviors will have on humans... This leads to the possibility of artificial general intelligences causing harm to humans; just as when humans act with disregard for the wellbeing of others.
Closing quote

In order to achieve value alignment, Riedl proposes the Quixote technique, which involves reading them stories that represent "acceptable" behavior in a particular culture, similar to reading children Aesop's fables or William J. Bennett's Book of Virtues. According to Riedl, this technique eliminates several of the difficulties inherent to teaching robots morality, particularly the unfeasibility of teaching the robot all of the values it needs to know and their virtually infinite applications. If you teach a robot specific rules, Riedl asserts, then it may not understand the nuances of the values behind the explicitly stated rules and find "loopholes." But by teaching the robot through storytelling, it can extrapolate patterns for itself (or "reverse-engineer" the tacit values behind the stories) and learn general "good behavior."

In order to illustrate this, the researchers simulated a situation for a computer program, such as obtaining prescription drugs, and showed it hundreds of stories crowdsourced from users on the internet and rated the behaviors in the stories from 0-10. By assigning rewards for "good" behaviors, like getting a prescription from a doctor and paying for the drugs at the pharmacy, and a lower rating for "bad" behaviors, such as stealing the drugs without paying for them, the program was shown to reject the unacceptable behaviors and only simulate the behaviors that were "moral."

Of course, there are several limitations to this study. First, this was demonstrated on a program that is relatively basic compared to the theorized "general intelligence," and the robots aren't necessarily being taught to be moral agents who can engage in ethical reasoning, but are simply being taught which behaviors are socially acceptable based on a specific culture's values. But still, the researchers are confident that this technique would go a long way towards encouraging value-aligned behavior and preventing "psychotic-appearing" behavior in AI.
Opening quote
Our technique is a step forward in achieving artificial agents that can pursue their own goals in a way that limits adverse effects that this may have on humans. Even with value alignment, it may not be possible to prevent all harm to human beings, but we believe that an artificial intelligence that has been encultured-that is, has adopted the values implicit to a particular culture or society-will strive to avoid psychotic-appearing behavior except under the most extreme circumstances.
Closing quote

Via Popular Science.
Science
Artificial Intelligence

Load Comments