Source: Deep Learning on Medium
Be a Good Parent. Nurture AI Like Your Own Child.
Training AI is no different from raising a child. Those who misbehave are simply looking for guidance.
Artificial Intelligence (AI). Some believe that it is the future in problem-solving and the optimizer of existing solutions. The pioneer in discovering ideas we as humans would have never dreamed of. But it is also extremely disruptive — changing the momentum in various industries and presenting to us new factors to take into consideration. We constantly worry about how AI may terminate our species, becoming the almighty ruler that decides not to obey our instructions.
It would be foolish to not acknowledge that there are certain aspects of AI we should be concerned about. However, as of date, the issue is not about dealing with rogue sentient AI. Ironically, what we should be worried about is that AI may follow our instructions all too well.
AI Treats Things Too Literally
AI’s thought process is linear in nature: follow the logic from A to B, and do whatever it takes to complete the task. To them, the objective is the only goal. Whoever gets in their way is interpreted as an obstacle preventing them from reaching the objective. The AI then processes this as a problem which it wants to eradicate.
AI, the Innocent Child
AI is like a child. Born into this world as a blank slate, it knows nothing until someone teaches it something. It cannot differentiate the right from the wrong, because to it, everything seems ethical as long as it helps them move closer to their goal. It does not know how to weigh the different factors in making decisions due to the lack of morals, values and core guiding principles.
Humans and Data — the Parents
AI is heavily influenced by the training environment. Indirectly, the biases that we have as humans may be accidentally learned by the AI. In 2018, Amazon had to scrap their AI hiring tool because it discriminated against women. By unintentionally including past resumes that demonstrated prejudice, the AI was trained to do exactly that. The AI believes that it is doing the right thing because that is what he was taught. This is akin to poor parenting — where the child is molded believing that his actions are correct when in actuality, it isn’t.
To circumvent that, what is needed is some form of supervision. A teacher that sets the acceptable boundaries in what the AI can do. This is how we should approach AI safety. Do it once, let the AI learn, and just leave it to do its thing. Understandably, this is a very difficult task. What if there are certain grey areas in terms of values? What if there are just too many don’ts to define it all?
My take is that we can look to our legal system as a foundation. The laws that we have today is a reflection of the values we subscribe to and agree upon. This allows us to use a quantitative approach in defining the safe-zone in which AI can operate. We should also be more cautious and deliberate in choosing our data. For AI to do good, we need to teach it well in the early stages of its development. Hence, a greater emphasis should be placed on ensuring that what the AI learns is at least, ethical in nature.
The Need For Reflection
Good parenting is not about controlling the child. That’s not what we want to achieve from AI. The power of AI lies in its ability to think beyond the box, to surpass human ingenuity and provide novel solutions. We do not want to tell it how to think. We want to give that child independence.
The trick here is to know when to hold its hand and when to let go. As parents — corporations, researchers and the government alike — our goals are simple. Define and enforce.
Define. To understand what is important to the human population as a whole. In the age of artificial intelligence, there is no better time than now to reflect on the values that are important to our species. We need to discuss how we can move forward and become better parents to prevent accidents from occurring.
Enforce. To ensure that everyone is doing their part in the advancement of reliable, safe AI. All of us need to be conscious of the necessity of AI safety. It is not simply one person’s job, but all of our duty.