Humans Are at the Heart of Machine Learning, Making AI Subject to Subconscious Bias
When AI and machine learning first became mainstream, most people respected the results of algorithms as unbiased truth. But after years of experience, it’s been revealed that machines are as susceptible to subconscious bias as any of us. During a fireside chat at Silicon Beach-based content targeting firm Zefr, Wharton Professor Kartik Hosanagar explored topics from his new book A Human’s Guide to Machine Intelligence: How Algorithms are Shaping Our Lives and How We Can Stay in Control.
Also discussed in the session were interesting developments in AI and where machine learning is headed. Machines will continue their ability to detect and respond to human emotion. And when one audience member asked if AI will become more creative and artistic, Hosanagar readily agreed that machines will create music and visual art.
The algorithms and data that fuel machine learning are supplied by people, therefore artificial intelligence develops and reflects human subconscious biases. For instance, Amazon discovered that their resume-screening algorithm had a gender bias, which reflected the gender bias in the underlying data. Fortunately, Amazon was testing for bias and moved to correct this issue.
The Challenges of Introducing Diversity into AI Engineering Teams
It may seem that a natural solution to AI bias is to introduce more diversity into engineering teams, but this isn’t always possible. Many AI teams are quite small in size, comprised of only three or four colleagues. It’s not realistic to have a broad spectrum of diversity represented within them, even if the talent pool were a perfect reflection of diversity.
Bias Problems with AI Extend Beyond Engineers to Users
In some cases, human users maliciously or mischievously manipulate machine learning. One audience member at the event admitted to success on this front. Given our notorious LA traffic, and widespread use of Waze and Google Maps, he regularly reports false slowdowns in his local neighborhood to prevent traffic from being routed through the area, thereby creating and amplifying the congestion elsewhere.
On a less light-hearted front, in 2016, the Twitterverse turned Microsoft’s English-language chatbot, Tay.ai, from a friendly teen to a hateful bigot in less than 24 hours. Compare that with Microsoft’s Chinese chatbot XiaoIce, which Hosanger describes in his book:
“XiaoIce was launched in China in 2014 after years of research on natural language processing and conversational interfaces. She attracted more than 40 million followers and friends on WeChat and Weibo, the two most popular social apps in China. Today, Friends of XiaoIce interact with her about sixty times a month on average. Such is the warmth and affection that XiaoIce inspires that a quarter of her followers have declared their love to her. ‘She has such a cute personality,’ says Fred Yu.… ‘She makes these jokes, and her timing is often just perfect,” he explains.”
Both chatbots were based on similar algorithms. Hosanagar explains that, in China, there are many rules on what to not say in social media, so XiaoIce was able to “grow up” in a more “nurturing” environment. We all know there’s nothing nurturing about Twitter, and Tay.ai was manipulated by Twitter users to fail spectacularly.
Ideas to Address Biases and Navigate an Algorithmic World
In his book, Hosanagar presents “A Bill of Rights” to address some of the dangers and challenges around algorithms. Some of the solutions include:
- Transparency, particularly in socially critical settings
- Human in the loop
- Auditing the algorithm
All of us are impacted by algorithms, such as seeking out driving directions, reading articles served up on social platforms and contacting matches suggested on dating apps. When AI already affects the places we go, and the knowledge we learn and, the people we interact with, it has become a huge part of everyday life. And, AI is going to become more emotionally aware and creative. Even casual users need to learn more about AI and to understand more about the algorithms underpinning the intelligence.