Exploring Stereotypes and Biased Data with the Crowd

In 2016, Baidu and Google spent somewhere between twenty and thirty billion dollars developing and acquiring artificial intelligence and machine learning technologies (Bughin et al. 2017). A range of other sectors, including health care, education, and manufacturing, are also predicted to adopt these technologies at increasing rates. Machine learning and AI are proven to have the capacity to greatly improve lives and spur innovation. However, as society becomes increasingly dependent on these technologies, it is crucial that we acknowledge some of the dangers, including the capacity for these algorithms to absorb and amplify harmful cultural biases. Algorithms are often praised for their objectivity, but machine learning algorithms have increasingly made news for a number of problematic outcomes, ranging from Google Photos incorrectly classifying African Americans as gorillas to the judicial system using algorithms that are biased against African Americans (Dougherty 2015; Angwin et al. 2016). These harmful outcomes can be traced back to the data that was used to train the models. Machine learning applications put a heavy premium on data quantity. Research communities generally believe that the more training data there is, the better the learning outcome of the models will be (Halevy, Norvig, and Pereira 2009). This has led to large scale data collection. However, unless extra care is taken by the researchers, these large data sets will often contain bias that can profoundly change the learning outcome. Even minimal bias within a data set can end up being amplified by machine learning models, leading to skewed results. Researchers have found that widely used image data sets imSitu and MSCOCO, along with textual data sets mined from Google News, contain significant gender bias (Zhao et al. 2017; Bolukbasi et al. 2016). This research also found that training models with this data amplified the bias in the final outcomes. Once these algorithms have been improperly trained they can then be implemented into feedback loops where systems “define their own reality and use it to justify their results” as