AI is Still Pretty Stupid

This past week I have been doing some research on artificial intelligence and it has been a lot of fun. I created one program to tell knock-knock jokes. Then I created another to play something similar to the old text-based game called Adventure. Both of those programs I create with the help of AIML and they each only took a few hours.

Next I looked at doing something a bit more advanced and utilized scikit-learn in Python with some machine learning. This time I played with creating a program to converse with me. I downloaded a number of training files that are used to train the program in the English language. The training took less than an hour and soon I could start interacting with my program. I typed in my first statement and it thought about it for a while. After several minutes, it spit out something in English that didn't match my initial comment. I figured my program needed a bit more training and so I taught it a few facts like "2 + 2 = 4". It responded with the correct answer when I asked what "two plus two" was. Then I asked about 2 plus 3 and it returned "2 + 2 = 4".

I played around a bit more and came to the conclusion that simple AI programs with an hour of training are really stupid. So how to do you spend more time teaching an AI program? Microsoft created Tay and let it loose on the Internet. In less than a day, it became incredibly racist. Naturally they quickly stopped the program. Think about it, would you really want your child to learn all its knowledge from the Internet? Of course not and so allowing your AI program to roam free on the Internet is probably not a good idea. The reason we send kids to school is that we want them to learn from a trusted source and for certain knowledge we don't even trust ourselves.