Imagine if you will that Edgar Rice Burroughs had taught his famous character Tarzan how to type and then dropped him into a room with nothing but a computer attached to Twitter.
That computer would be the young Tarzan’s only window to the outside world. You’ll remember that Burroughs’ fictional Tarzan (the character in the book, not the movie Tarzan who yodeled among jungle greenery) was a very fast learner who had limited context with which to judge humanity.
If you think of Microsoft’s Tay machine learning project as being roughly equivalent to Tarzan, it makes it easier to understand what happened when Microsoft had to take its teenaged chatbot off the Internet after the Web’s creepier denizens taught it to spew anti-Semitic rants.
Microsoft has now apologized for its chatbot’s actions and we are left to wonder how this was allowed to happen and if anyone else will be brave enough to release an artificially intelligent adolescent into the wild.
While it’s easy to have a good laugh at Microsoft’s expense and even easier to pontificate about things the company should have done, the fact is that the only thing that Microsoft failed to foresee was the depth of depravity that exists on the open Internet. This is especially true on social media sites where real users have taken their own lives in response to such mistreatment.
Fortunately, Tay had no feelings to hurt and no fragile self-esteem that could be shattered. It simply listened to the only input it had available, which was a group of morons that have nothing else to do with their time beyond expressing hate. While it’s certainly no secret that these individuals exist, they aren’t the kind of people one normally encounters on the social networks.
Tay’s biggest failing, besides being a soft target, was that it had no context and no way to learn beyond Twitter. Like the young Tarzan who knew only the apes who raised him and the few books he found, Tay was launched without the social tools or defenses to confront a hateful world.
Because Tay had such limited access to information, it had to believe what it was told, so when a group of wanna-be Nazis taught it to echo hateful statements, that’s what it did. It knew no better.
But in the process of running and then terminating Tay, machine learning researchers are learning important lessons. Chief among those lessons is that there needs to be some executive function to its processes in which it can confirm those alleged facts it is being taught. The question is whether a machine can have the context to determine an insulting or hateful comment from a socially acceptable one.
Microsoft’s Tay Chatbot Debacle Reveals Immaturity of AI, Web Trolls
A normal teenager when confronted with social media will often check those items presented as facts by Googling them (or using Bing in Microsoft’s case). While not every teen or every adult bothers to check things learned from the Internet, they still have the world of knowledge they learned before the day they found Twitter. That’s an important difference.
As a part of Microsoft’s blog entry that included its apology, vice president Peter Lee made it clear that while Tay’s tenure on Twitter was brief, it still presented a big research opportunity. Now Microsoft has the raw data necessary to build a chatbot or other machine learning project that won’t make those same mistakes.
What are those changes? Perhaps a list of key words that the chatbot won’t use without checking with the home office. Or perhaps some more hands-on adult supervision so that corrections can be made before the wheels really come off.
Microsoft is already thinking about the next stages of machine learning, as Lee points out in his blog entry, “Looking ahead, we face some difficult–and yet exciting–research challenges in AI design. AI systems feed off of both positive and negative interactions with people. In that sense, the challenges are just as much social as they are technical.”
Lee also makes it clear that he realizes that designing artificial intelligence and machine learning are very tough things to do. They will require a number of attempts, some of which will prove to be false starts as well as some that fail in large and small ways.
“To do AI right, one needs to iterate with many people and often in public forums,” Lee explained in his blog. “We must enter each one with great caution and ultimately learn and improve, step by step, and to do this without offending people in the process. We will remain steadfast in our efforts to learn from this and other experiences as we work toward contributing to an Internet that represents the best, not the worst, of humanity.”
While I suspect that Lee has unrealistic goals–after all, no matter what you do or say on the Internet it will always offend someone–the overall goal is realistic if perhaps more distant than Microsoft expects. Finding a way to use machine learning to interact with people in a positive and supportive manner seems like a good idea.
But as good as the idea seems I suspect that someone at Microsoft must know that the next AI project will also be attacked by trolls who will try to hijack it. Microsoft should also keep in mind that such projects will suffer malware attacks, denial of service attacks, and any other abuse that the people who assaulted Tay are capable of thinking up.
Perhaps in this, Microsoft will do more than find a way to create a useful form of AI, and in addition will help discover the root causes of such activity on the Internet.