Behind the scenes of artificial intelligence: Studying how organizations cope with machine learning in practice

From recruitment to health care and from law enforcement to education, artificial intelligence (AI) is increasingly implemented in organizations. Using machine learning, these systems produce insights that potentially go beyond what is humanly possible. However, what is commonly overlooked in existing research is that there is a fundamental difference between the procedures used for machine learning and how human knowledge is produced. This difference makes it challenging to find common ground for sharing and collectively producing knowledge which creates new yet unknown challenges when implementing machine learning in practice. Therefore, this dissertation sets out to answer: How do organizations cope with the production and use of machine learning in practice? Chapter 2 relates to the core feature of AI systems: data. I ask what happens when workers are facing the need to embed “data work” practices in their existing, situated work. In this Chapter, I build on the final year of my ethnographic research at the Dutch police, where I joined the emergency response department full time “in the streets.” I show that police officers anticipate the data work and adopt three coping strategies in their situated work: avoiding work, deviating from protocol, and capturing experiences. While these strategies help the police officers to alleviate the burden of data production they experience on a daily basis, they have a large influence on how police officers perform their situated work. As a consequence, what and how crimes are reported and data is produced is significantly influenced by their coping strategies. Chapter 3 builds on my two years of fieldwork at the intelligence department of the Dutch police. In this Chapter, I focus on the opaque nature of machine learning and the ability to produce new insights and offer one of the first empirical accounts of algorithmic brokers. I ask how such brokers can translate machine learning knowledge when they cannot understand how knowledge is generated. I find that as the algorithmic brokers try to become more familiar with machine learning, they perform different translation practices over time and enact increasingly influential brokerage roles, i.e., messenger, interpreter, and curator. When, finally, the algorithmic brokers come to the conclusion that they can never understand how machine learning knowledge is generated, they act like “kings in the land of the blind” and substitute the algorithmic predictions with their own judgments. In Chapter 4, I build on the three unique features of AI systems that I unpacked in Chapters 2 and 3 and use unique insights from five different cases across various industries to ask how the “implementation line” can be crossed in the case of AI, in which technology development and organizational deployment are often worlds apart. I identify three different AI implementation practices – i.e., organizing for data, organizing for explainability, and organizing for new insights – and show how, through these implementation practices, developers and organizational actors can engage in continuous and reflective “collaborative learning.” This dissertation contributes to the discussion on the production and use of knowledge that has been core to the field of organization theory for decades. Moreover, by including and theorizing the specific features of AI systems and their relation to organizing, this dissertation responds to the call to bridge the divide between technology development and organizational change. Also, this dissertation links to the field of information systems by going beyond the “AI hype” to unpack the challenges that emerge when organizing for machine learning knowledge in practice. Finally, this dissertation also has practical implications, as I urge managers to let go of the “AI hype” and instead consider AI implementation as effortful, skillful, and requiring long-term involvement.