Do Androids Dream of Hateful Tweets?
In recent days, news of Google’s issues within their AI, “Vision AI”, emerged. Vision AI automatically categorizes images, and unfortunately produces vastly different labels for images based on the given image’s skin tone. [1] Google has apologized, and appears to be working on fixing the issue, but this leads me to wonder about human errors when designing AI. No better example encompasses this faultiness than Tay, Microsoft’s 2016 social experience in which citizens use Twitter to communicate with a chatbot. After barely a day of consciousness, Tay was “put offline”, begging the question, where did things go so wrong? What can we learn from the ill-fated Tay?
Tay was met with skepticism and wariness at her debut, viewed as yet another innovative exercise in corporate pandering. According to her website, Tay was “developed by Microsoft’s Technology and Research and Bing teams to experiment with and conduct research on conversational understanding.” [2] Specifically, Tay was designed to curry interaction between itself and millennials, a demographic Microsoft was eager to tap for user information, gender, zip codes, etc. Tay’s “personality” did little to endear her to an audience, despite being advertised as 19, Tay’s vocabulary and sensibilities read more as “tween” than “teen”. She would say (exceptionally un-millennial) statements like “artsy fartsy” and “swagulated”. She loved EDM and Calvin Harris. Still, at Tay’s debut the bot appeared earnest and well-intentioned, providing the number of the National Sucide Prevention Hotline to depressed “friends’’. Like all semi-decent things on the internet, this positivity was short-lived, and Tay rapidly emerged as a toxic, racist, sexist troll. She had a deep love of Hitler, and a profanity for spewing increasingly bizarre and disturbing tweets. This spectacle concluded with Tay being “put to sleep” barely 24 hours after her launch into the public sphere. [3] [4] [5] [6]
Interestingly, Tay made her debut off the back of serious accomplishments within the AI field. A few months before Tay’s implosion, Google’s AlphaGo won a Go tournament as a top-ranked human. [7] Astonishingly, it took five months for AlphaGo to become an expert Go player, reviewing, internalizing and playing more games than a human can within our lifetimes. [8] Tay’s personality transformation resulted from a similar mass ingestion of data. Through exposure and an ingrained necessity to learn and interact from humans, Tay internalized and metastasized into an alangaormation of human’s worst traits. It’s not shocking to see why she cracked.
What could have Microsoft done differently? Reastically, a lot of things. I don’t understand their logic of placing a tween-like robot within the troll-filled cesspool which is Twitter. Microsoft did build some safeguards in Tay’s design, certain words and topics were filtered and Tay was coded to refrain from controversial subjects and certain forms of hate-speech. Tay was also tested “with diverse user groups”, however the scientists did not expect that users would “exploit a vulnerability in Tay”. [9] While Microsoft’s AI platform XiaoIce is wildly popular in China, with over 40 million users conversing pleasantly with the bot, Microsoft did not take into consideration cultural differences between how various groups of people interact with different technology platforms. Ultimately, Microsoft expected better from humanity despite looking for kindness at the swampy bottom of the internet.
In the future, scientists should be more acutely aware of how AI can positively develop alongside negative interactions. Misha Bilenko, the head of Machine Learning at the Russian tech giant Yandex, states that the Tay fiasco ultimately reveals a deeply human error within ourselves — we expect AI to have “deep seated beliefs” instead of viewing it as the “statistical machine” that it is. [10] We humanize the machine, and hold it up to our own values and ethics. Without an interior life and emotionally-driven actions, a chatbot will never read as human. For AI to grow and develop, the bot needs to interact with as many various forms of people with wildly differentiating viewpoints. There will be mistakes. But how do you have a chatbot which reflects the best aspects of humanity, not the worst?
[4] https://thenewstack.io/requiem-tay-reactions-microsofts-teenaged-ai-gone-bad/
[5] https://gizmodo.com/here-are-the-microsoft-twitter-bot-s-craziest-racist-ra-1766820160
[6] https://money.cnn.com/2016/03/24/technology/tay-racist-microsoft/index.html
[9] https://blogs.microsoft.com/blog/2016/03/25/learning-tays-introduction/
