L04 “In the land of the blind, the one-eyed man is king: Knowledge brokerage in the age of learning algorithms Waardenburg, Huysman, Sergeeva (2021)” Flashcards
According to the evaluation report, the work process should include three steps:
actualizing, interpreting, and explaining.
Actualizing meant adjusting predictions to local changes (e.g., when a burglar was captured).
Interpreting meant adding more information to the crime predictions, such as the most-used crime methods.
Explaining meant deeply analyzing why a crime is predicted (i.e., finding causal explanations for the algorithmic predictions).
“Algorithmic broker acting as messenger”
“The main aim of intelligence officers’ work was to make abstract algorithmic predictions meaningful for local police managers.”
“Algorithmic broker acting as interpreter”
“To translate algorithmic predictions to the police, the intelligence officers realized they lacked a deep under- standing of the machine learning community and the police community and invested in learning more about both”
“Learning about the machine learning community: Consequently, the first step was to interact with the data scientists to find out more about their practices and to see if the causes of predictions could be made transparent.”
“In sum, to translate algorithmic predictions to the po- lice, the intelligence officers realized that they them- selves first had to better understand how these predictions were generated and how police work was performed. In their efforts to find out more about the decision logic of crime predictions, they encountered the opaque nature of learning algorithms, which solidified a knowledge boundary between the machine learning community and the intelligence officers. On the other hand, due to the consistent interactions with the police, the access to the police data, and the police man- agers’ increased belief in the value of crime predictions, the knowledge differences between the intelligence officers and the police community was slowly fading. This allowed the intelligence “officers to contextualize the algorithmic predictions in such a way that they made sense to the police managers (see Table 2). By performing translation practices in the form of “examining” and “domesticating,” the intelligence officers enacted a knowledge brokerage role that can best be described as an “interpreter.” However, even though their contextualizing efforts seemed to work for the police managers, the intelligence officers continued to struggle with understanding the blackboxed machine learning”
“Algorithmic broker acting as curator”
Now that the intelligence officers became used to their ascribed expertise as algorithmic brokers, they searched for ways to deal with the opaque algorithmic predictions and discussed this with their head of department. He suggested that the difference between machine learning and their human interpretation was in fact so large that it could not be overcome and that they should therefore use their own expertise.
“In sum, the intelligence officers eventually realized that the boundary between machine learning and their human interpretation of crime predictions was impassable. As a consequence, they pushed back the learning algorithm and substituted it with explainable alternatives that aligned with their human judgments and that they considered most suitable for the police managers (see Table 2). As such, by performing translation practices in the form of “substituting,” the intelligence officers “enacted a brokerage role that can be best described as a “curator,” in which they grew to become more influential and were eventually considered more as a partner to the police managers.”
Contribution:
“We contribute to the knowledge brokerage literature by providing a more fine-grained and dynamic perspective on how knowledge brokers enact translation practices over time and in relation to opaque algorithmic predictions. Building on Røvik (2016) and based on our empirical findings, we consider “extracting” and “examining” as practices to translate from the machine learning community, and “transferring,” “domesticating,” and “substituting” as practices to translate to the user community, which offers a more refined insight into the complexity of brokerage work.”
Conclusion:
Learning algorithms, because of the black-boxed machine learning, offer an extreme case for understanding how knowledge brokers enact translation practices. In this study, we provided a case of knowledge brokers who aimed to translate algorithmic pre- dictions from a machine learning community to a user community. Translation has always been the core of knowledge brokerage work, yet so far has been mainly taken for granted in organizational literature. It is now, in the age of learning algorithms, of significant importance to question how knowledge brokers are able to translate from a machine learning community, since machine learning has become increasingly difficult to understand. As this study shows, when the outputs of one community are opaque to all actors involved, brokers can become “kings in the land of the blind” and decide to substitute algorithmic pre- dictions with their own judgments. The case of learning algorithms therefore highlights that knowledge brokers should not be considered as merely instrumental in solving knowledge boundaries but even more so as highly influential curators of knowledge.
