The elusive nature of neural nets is, in many ways, similar to that of the thought process of young children. This was explored in depth in part one of this article. After finding this, I worked to develop a system to manage the juvenile actions of AI.
These cheeky solutions and tantrums all hit that same childish note. It became apparent that I should really start thinking of my AI models simply as peculiar children.
Not having any kids of my own I asked for parenting advice from someone who is impeccably skilled at making sure unruly children follow not just the word but the spirit of the rules. My mom. My ~classic millennial~ move ended up providing a fruitful structure for “parenting” my AI:
Give all the information needed
One of my aunts threw a tantrum right before her second day of preschool after a flawless day one. It turned out that she was upset because nobody told her she would have to go back to that place ever again. After the typical timeline of education was explained to her, there was never another issue with attendance.
An AI’s universe of understanding is only as broad as what it is taught. If its knowledge base is insufficient to cover the subject matter with which it’s trying to work then it will have unpredictable, often bizarre, behaviour.
Give only the information needed
One day when I was young, our dad was taking care of my siblings and me when he stubbed his toe prompting him to blurt out an expletive. This new word delighted my younger sister who proceeded to march around the room chanting that word, pausing briefly to note that this would be the name of her first dog. I still don’t know how my dad successfully deescalated the situation before we woke the baby in the next room and I commend his success.
Often, publicly trained chat bots will have similar issues with learning offensive terms or concepts. AI in general will employ any information at its disposal regardless of absurdity, propriety, or common sense.
The iteration of the cancer detection bot which was trained on images without rulers ultimately had no way of finding the cheap, easy way of completing the given task. Trim the information you provide down to the express essentials. What the computer doesn’t know can’t be used against you.
Provide structure
When I was a kid, we had this container with a bunch of little pieces of paper outlining different chores that needed to get done around the house. It was aptly named the chore jar. Every day when we got home from school, each of the kids would have to blindly pick one task from the jar and get it done before we were allowed to goof off.
This worked pretty well until my mom noticed that some of the chores started to read “eat cookies” and “play video games”. She found out that I had been slowly adding my preferred activities to the jar so I could do whatever I wanted to without explicitly breaking any rules.
I didn’t get in trouble; I didn’t technically do anything wrong. I did, however, cause the end of the chore jar and the beginning of more explicit assignment of responsibilities. Ultimately, that wasn’t expressly a win for me.
Much like fairies, genies, and other mythical deal makers, AI is bound to the word, but not the spirit, of the law. That gameplaying AI couldn’t pause Tetris if that action was explicitly off limits.
Set boundaries and expectations. The more avenues you block, the more likely the machine will follow the type of path you had envisioned.
Empathize with your inner kid
You were young once too. Think back to your own experiences at that age. Therein lies a treasure-trove of knowledge off of which you can base tactics of managing these divergent actors. Your frustrations with the unknown, your brazen tendencies, and your circumvention of the rules were products of an unbridled mind at play.
Many of you have likely never been a machine before so it would be a bit of an honest stretch to “think like AI”.
However, you definitely were (and maybe still are) a creative, disruptive, divergent thinker.
