This post works on several layers. First, it makes the obvious point that it is very easy to create a racist artificial intelligence (AI). Second, it makes the less obvious, but much more important, point that making a racist AI is the default if you use standard techniques. Using the most popular website crawl data, the most popular sentiment lexicon, and the most popular AI engines, you inevitably get a racist result (for example: Mexican food is rated worse than other foods, typically Black names are rated lower than other names). Third, the author also shows how easy it is to correct for built-in racism (i.e., if you get a racist result, you're not really trying). And fourth, at a meta level, is the use of the notebook format to present the results, so you could work directly with the code yourself if you wanted to. The challenge to learning analytics is, of course, how transparent will LMSs be in showing their analyses, and how can we be sure they didn' simply take the path of least resistance to create racist results? And what other, less obvious, biases are built into our data?
Today: 6 Total: 88 [Direct link] [Share]
