Dangerous Liaisons - UW Libraries

December 28, 2017

Musing on Algorithmic Bias

Chloe Horning

At the Critical Pedagogy Summit at the beginning of the month, a thought occurred to me: Understanding ‘ways of knowing’ and modes of discovering information that are not our own, but which belong to people with different cultural and personal experiences from ours, requires subtlety, complexity, and work.  

But it seems to me that, as Barbara Fister puts it, libraries have a “black box” problem, wherein librarians are under increasing outside pressure in the age of Google, to deliver complex information in as simple a manner as possible, all the while disguising the mechanism, so that it looks like we haven’t done anything at all. One problem with this is that search algorithms are not unbiased, and that they can reflect the biases of the human beings who create them. I first engaged with this concept when I encountered the work of Safiya Umoja Noble, which challenges the racist and sexist biases in Google’s search algorithm. Later, I learned about the gender and race gaps in Wikipedia. 

But these examples are in the domain of the open internet. Surely library searches, which are more controlled, don’t have the same problems with algorithmic bias, right?  

Not according to an article by librarian Matthew Reidsma, which appeared on his blog last year.  The article addresses algorithmic bias and the ‘discovery layer.’  

Reidsma examines Summon, and specifically a feature of Summon called “Topic Explorer.” His research project found significant bias in results about women, the LGBT community, race, Islam, and mental illness. Happily, he took his concerns directly to Pro-Quest, who took them seriously and (though I haven’t heard an update) hopefully are taking steps to correct them.  

This made me curious as to whether anyone had undertaken a similar study of algorithmic bias in Primo search results. Topic Explorer is somewhat unique, because it is feature of Summon that displays full-text background information on the topic that it believes you are searching for in Summon, based on your query. We don’t currently employ a comparable tool in Primo. As Reidsma notes: “By returning only a single result and placing the result on the right side on wider screens (mirroring Google’s Knowledge Graph), the Topic Explorer is meant to say, “this is what you are searching for.” Because there is only a single result, there is a confidence inherent in the design that would make this a test subject for algorithm effectiveness.” 

So, to be fair to Primo, they don’t currently employ a feature that is as problematically prescriptive as Topic Explorer. Nevertheless, I’d would be very curious to know whether an analysis of algorithmic bias in the search results returned by Primo is possible. Somebody out there must have the will and the ability to do this research, or maybe someone has already begun.