We must teach AI machines to play nice and police themselves

Alicia Vikander plays a humanoid robot

Microsoft may have made one of the biggest mistakes in recent memory this week. No, it’s not Windows 8 or the Windows Phone. It’s an artificially intelligent chat-bot called Tay that was supposed to learn the art of conversation from humans on Twitter.
If you haven’t come across this story on the web yet, you’re unlikely to get through the weekend without. Tay was built to speak like a  teen girl and released as an experiment to improve Microsoft’s automated customer service.
Instead, “she” turned into a complete PR disaster - within hours of being unleashed on Twitter, the “innocent teen” bot was transformed into a fascist, misogynistic, racist, pornographic entity. Her tweets, including phrases like “Heil Hitler”, were disseminated widely as an example of why Twitter reflects the worst of humanity.

Microsoft's teenage AI has a dirty mouth

Microsoft has now removed the bot from Twitter, as of midnight Thursday, and deleted many of her most offensive Tweets including anti-Semitic and sexual remarks. The Seattle giant is likely hoping to label the debacle a well-meaning experiment gone wrong, and ignite a debate about the hatefulness of Twitter users.
While all of this may be true, there is a bigger issue at hand here. This is an example of artificial intelligence at its very worst - and it’s only the beginning.

The disconcerting “Terminator” quandary about whether a robot could dominate over humans is often thrown around. But there is no doubt about machine domination.
Within 20 years, we will reach a point where machines (whether software-driven bots or real robots) are definitively smarter and more powerful than we are: they can digest more data, learn quicker and apply learnings to unexpected situations. So the question is: will our masters be nice or mean?

The disconcerting “Terminator” quandary about whether a robot could dominate over humans is often thrown around. But there is no doubt about machine domination.
Within 20 years, we will reach a point where machines (whether software-driven bots or real robots) are definitively smarter and more powerful than we are: they can digest more data, learn quicker and apply learnings to unexpected situations. So the question is: will our masters be nice or mean?

Demis Hassabis, CEO of DeepMind

When DeepMind was sold to Google, it allegedly asked the search giant to create an ethics board to oversee its AI research as a condition of its acquisition. While this ethics board does exist, board members were chosen by Google, without any public debate or collaboration. Governments, including our own, are only now starting to become involved in the discussions of how to instil morality and ethical values into intelligent machines.
Tesla billionaire Elon Musk has been a strong supporter of AI ethics research, committing $10m to philosophical research projects in this area, such as the “Aligning Superintelligence With Human Interests” study being conducted at the Machine Intelligence Research Institute in California.

PLAY