Lessons From Two Artificial Intelligences

Artificial Intelligence has been in the news. Last week Google got a lot of great press with some AI software that learned how to play the ancient game of Go. (Google's AlphaGo scores 4-1 against South Korean Go player)  This week things didn't;t go so well for Microsoft Research when they sent an innocent AI on to social media to learn on it’s own. (Microsoft deletes 'teen girl' AI after it became a Hitler-loving sex robot within 24 hours )

Some differences jump out at me that I think have implications for education. The first is how the two AIs, both of which were designed to learn, were educated. AlphaGo was introduced to the best Go strategy available. It continued to learn on its own but under close supervision and guidance. When it started to interact with people beyond the research team it was with professionals who had nothing but good intentions. Things worked well.

Tay was sent out into the world with very little in the way of training. The idea was for it to learn from interaction with people on the Internet. Unfortunately people on the Internet can have goals other than helping an AI learn good and useful things. Supervision seems to have been less close than would have been ideal. Tay learned from people who seemed, in too many cases, to want to mess things up. Things did not work well and the AI was shut down in less than 24 hours.

As I see it having help selecting what to study and who to study with was a huge advantage for AlphaGo. Tay was set lose with no guidance. It had no way of knowing what was good information, who were reliable sources, or how to use the information it gathered. Poor teachers can do a lot of damage to people or AIs. “Teachers” who are trying to steer people wrong can do even more damage.

I suspect that if most people had communicated with Tay in good ways things would have turned out much better. This experiment turned into a game without rules and that is very unfortunate. Though I suspect the researchers learned quite a bit anyway.

An AI can be turned off, reprogramed, and made harmless pretty easily. People not so much. That is one reason I worry and too little guidance for learners. Experienced, generally older, learners have good filters and better insight into what is going on than younger learners of course. Some people are better at discovering things as well. But most people need some help from teachers in some ways. It is important that we pick good teachers.