robot reading 
 
 One of the more disturbing trends in robotics is how often some researchers gloss over the moral complexities of AI by suggesting Isaac Asimov’s “Three Laws of Robotics” will be sufficient to handle any ambiguities robots encounter. This only serves to demonstrate how unfamiliar many technologists are with the depth of the issue, for as a closer reading of Asimov’s work reveals, the three laws of robotics leave plenty of room for disastrous outcomes.

There are encouraging signs, however, that at least some other researchers are taking the problem seriously. Two figures leading the charge in this direction are Mark Riedl and Brent Harrison of Georgia Institute of Technology. They are pioneering a system called Quixote, by which an artificial intelligence learns “value alignment” by reading stories from different cultures.

This may seem a strange approach to the problem — until one considers how regional questions of morality are. For example, two counties only a few miles apart in the United States may have vastly different ways of perceiving moral and ethical issues. For instance, spanking your child may be considered a morally unacceptable act among an affluent urban population, whereas immigrants or more rural communities may find this form of punishment more acceptable. This is not to caste aspersions in either direction, but simply to point out that even among humans, there is little consensus on moral and ethical issues. How much more difficult will it be for humans to agree on what constitutes ethical robot behavior?

Quixote system

The Quixote system, by which an AI learns ethical behavior through reinforcement learning and multicultural stories

As it turns out, there may be a way, and it’s not so different from how humans learn the values particular to the region they grow up in: by telling stories to each other. The fables and legends particular to a locale are frequently laced with moral directives that provide clues to a growing child on what constitutes morally acceptable behavior in their culture. The idea behind Quixote is that the same principles can be applied to shaping robot behavior.

Quixote builds on Riedl’s prior research project, which demonstrated that an artificial intelligence can be trained to identify a correct sequence of actions by crowdsourcing story plots from the Internet. This is not at all dissimilar from the RoboWatch project we reported on earlier in the year, which showed how an AI could be trained to learn simple household tasks like the steps involved in making a peanut butter sandwich by watching YouTube videos. The Quixote system takes this a step further, adding a reward signal that reinforces certain behaviors and punishes others during a trial-and-error learning process. In essence, the AI learns to imitate the behavior of the “good guy” in the stories it reads from the Internet and avoid the behavior perpetrated by the ignoble characters.

However, the approach is not without its pitfalls. In some parts of the world, for instance, female genital mutilation is considered de rigueur, though the bulk of humanity regards such practices as objectionable and would certainly not want robots perpetuating them. In other words, in training robots on stories modeled after our own cultural precedents, we may wind up giving new life to behaviors that are better left to wither on the vine. Therefore, formulating a robust and universal ethical system — usually the pastime of ivory tower philosophers — is likely to take on increased significance as we consider the possible repercussions of endowing robots with the same porous and inconsistent moral norms exemplified by our own regional populations.