Generative AI utilization: How developers utilize Generative AI

Abstract

After the deployment of generative AI, many organizations have been quick to find a way to implement it into their work processes in the hope that it will increase efficiency. There are however challenges that come with generative AI, and it is important that organizations attempting to adapt generative AI are aware of the challenges and the opportunities, to prevent common issues and increase efficiency. This study attempts to answer the research problem “how do software developers utilize generative AI?” to establish the most effective use cases, how to ensure quality and how users get motivated into using it. This is a qualitative study where we have conducted semi-structured interviews with employees in consultant companies with technical knowledge, like developers, because they have a more natural curiosity towards new technology and can in turn provide better insights that other non-technical professions can provide. After analyzing the data from the interviews, we were able to identify some key factors that managers should consider when implementing generative AI themselves. We have identified how important it is with knowledge work, as intrinsic motivation is the main driving factor for generative AI utilization, and the more an employee knows about the possibilities of generative AI, the more they are willing to use it. We identified two categories of use modes, explore, and accelerate, which depending on the use case, has different requirements for accuracy. Accelerate mode is when users use generative AI to complete simple, but time-consuming tasks. In these cases, it is very important that there are security measures in place to ensure that the work being completed is correct and according to company standards. In explore mode, generative AI is used for inspiration and discussions, and nothing it creates is used directly, and therefore the accuracy is not as important. Too many rules and regulations with generative AI decreases employee’s motivation and increases the time spent on the task. Because of legal and privacy issues, developers can share very little of the code base with generative AI, and without that contextual understanding, generative AI provides subpar answers. A solution to this is integrated large language models that are trained on company data and do not share this data. Solutions like these have a much better contextual understanding of the project and remove the ability for the users to commit mistakes. The findings in this thesis suggest more research into specific types of large language models, as well as studying these principles on other professions, to establish how they affect employees from less technical professions.