EML Blog #6

Write about a certain type of machine learning product that has been introduced to the commercial market but has required a large amount of persistence and learning through failure. This could be a current, on-going project. Research and write about its history and what the open or current problems are in using machine learning for this project.

One of the most complex and identifiably human features of humanity is our language. While full of complex patterns that our interlocutors instantly pick up on, it is also full of slight variants on these patterns that people use to instill emphasis, humor or a completely opposite meaning. For these reasons, natural language has long been a goal of machine learning development.

Recently, one such endeavor was Microsoft’s AI chatbot deemed “Tay.” The idea was the Tay would be unveiled on Twitter and through other users having conversations with her, she would learn how to have conversations as well. There weren’t technical difficulties per say, however, it soon became evident that teaching Tay via the general users of Twitter was perhaps a bad idea. Many people began trolling Tay with racist, sexist or just wrong information causing her to be trained on data that Microsoft almost certainly did not intend. Tay quickly started to parse this input into a model of what she was presented with as “conversational English communication.” The following image shows a few of the tweets that were a result of this dataset.

Microsoft turned Tay off after about 24 hours and deleted some of the more offensive comments. However, they do seem to think of Tay as a failure, claiming, “The AI chatbot Tay is a machine learning project, designed for human engagement. As it learns, some of its responses are inappropriate and indicative of the types of interactions some people are having with it. We’re making some adjustments to Tay.”

This situation has raised many questions about the ethics of training bots on human behavior. How can AI be trained on human behavior but still be trained on the “correct” behavior? If bots like Tay are not trained on the general public, how will the subsect of people be chosen?

Parallel to Tay, there is the much more successful Microsoft baby, Xiaoice, that is a mainstay in the Chinese market. It has been integrated into the popular messaging app WeChat and the blogging site Weibo where many people follow it. Unfortunately, the success of Xiaoice which has strong since all the back to 2014, has not been able to help in the English language. Microsoft does not plan to give up however. With rivals such as Facebook, Apple and Google pushing the field further every day, they are focusing not only chat based AI such as Siri, Cortana or Tay but also on background AI that could automatically convert your emails to calendar events or reminders and other functionalities to improve the user experience.