Training a computational mind is akin to raising a child. Both challenges face a similar value attribution problem, in which you, the parent, the programmer, must decide which information you want to expose to your progeny. Successful parenting often employs shared creative experiences, like art-making and reading aloud from good books; balanced against a healthy dose of socialization, desirable traits like self-confidence, individuality, honesty, and dedication can easily be cultivated and raised. When done poorly, negative traits can arise, like jealousy, greed, or disrespect. AI researchers face the same dilemma: how can they get their artificially intelligent programs to represent the right ethical values in their ceaseless decision making? What kind of problems do you want your artificially created child to solve, and how?
I argue that a similar approach to child-raising must be taken with artificial intelligences. A good balance of engaged, supervised creativity to build foundational skills, and unsupervised socialization, like interacting with other AI networks in game-playing or problem-solving, could produce emergent ethical characteristics that are desirable. The intelligence is free to break rules in a supervised environment because the input from the programmer can correct biases that accidentally emerge in such a setting, and then it must follow certain ‘rules of the game’ in their unsupervised interactions that will do the guiding automatically. Oscillating between both settings can allow for weight modification of values by the programmer at any time, as well as the option to withhold intervention entirely if the trajectory appears correct. Programmers, teach your neural networks algorithms for kindness, largesse, and individuality in the same way you’d teach your child. Read to them, by training them on corpuses of your favorite books. Give them little art projects with goals and see what they create.
At the end of this process, in the same way that children often surprise their parents with acts of genuine empathy and generosity, your programs may surprise you with their humanness. These are the artificially intelligent systems we need to be producing. These are the systems that will guide our species towards universal altruism and joy for all. All it takes is a little bit of intentional parenting.