How to assess the usefulness of procedural approaches and why to focus on those
How to Assess the Usefulness of Procedural Approaches
As mentioned earlier, the methods lack adoption by developer teams (Siqueira de Cerqueira et al., 2022) due to a disconnect between what developer teams find useful and the current procedural solutions offered. Investigating what developer teams do assess to be useful, the most common definition characterises ‘perceived usefulness’ in the context of information systems as “the degree to which a person believes that using a particular system would enhance his or her job performance” (Davis, 1985, p. 26). However, the enhancement of job performance is a multi-dimensional and complex concept. For the purpose of this research, I follow the findings of Sonnentag et al. (2008) and define it as the ability to support developers to solve challenges, which is a more proactive interpretation of job performance. Consequently, this thesis uses the concept of ‘perceived usefulness’ as the ability to support developer teams to address challenges that occur during the implementation of principles in the development of AI systems. The following three (partly overlapping and interrelated) challenges are considered to be some of the most pressing issues.
First Challenge: Time Pressure Firstly, developer teams experience considerable time pressure during the development of AI systems (Morley, Kinsey, et al., 2021). This is the result of high stakeholder expectations, which is in conflict with the currently novel development stage of AI systems and the accompanying lack of best practices (Beckert, 2021). Adding additional tasks, such as the implementation of ethical principles, to the development process exacerbates the time pressure. Consequently, methods that enable developers to limit any additional time effort to a minimum while providing significant advantages are likely to balance out the negative aspect of any additional time effort.
Second Challenge: Conflicting Approaches Second, the approaches used by developers during the implementation of ethical principles in the development process and the approaches required to embrace the context and conflicting nature of principles are likely to be incompatible. This is because developer teams want to prove the feasibility of a system and focus on solving problems straightforwardly, meaning with the logical execution of sequential steps. This way of thinking is not simply translatable to the area of ethics, which requires discussions in a less structured and explicit manner to fully embrace the various ethical dimensions (Beckert, 2021). Thus, the ethical approach is not compatible with the training of developers, and implementing an approach which requires “practitioners to completely change their way of working would be an uphill battle” (Zuber, Kacianka, Gogoll, Pretschner, & Nida-Rümelin, 2022, p. 8), leading to frustration and, consequently, resistance to adopting this working style. Therefore, the methods have to be translatable to both approaches to provide a solution for this challenge.
Third Challenge: Mushy Stuff Last, ethical principles in the context of AI development are seen as abstract, also referred to as ‘mushy stuff’ by developer teams (Hussain et al., 2022). This is particularly disadvantageous since developer teams already view principles as something that “is brought to them from the outside and that has nothing to do with their actual work” (Beckert, 2021, p. 4). Consequently, the abstract nature of principles further increases the complexity and, thus, the unwillingness to implement them. Therefore, the methods have to operationalise principles in a specific and understandable way to reduce the inherent abstract nature and increase the willingness to implement.
Having reviewed the landscape of the Ethics & AI research field, consisting of principles, technical tools, and procedural methods, this paper will focus on the latter. This is because principles and technical tools alone cannot drive the ethical development of AI systems but necessarily need processes to balance out the respective shortcomings – making procedural methodologies a significant driver in developing ethically aligned AI systems. However, procedural methods lack a high adoption rate due to the disconnect between developer needs and current solutions. This could be solved (at least partly) by investigating which conceptual components of the respective procedural methodologies are perceived as useful by developer teams, approximated by their ability to support developers in the three challenges (‘time pressure’, ‘conflicting approaches’, and ‘mushy stuff’). Subsequently translating this knowledge into best practices will enable future research projects to design procedural methodologies that are perceived as useful and therefore have a higher probability of being adopted. However, to the best of my knowledge, there is no work to date that matches conceptual components to their perceived usefulness by developer teams, making this an unresolved knowledge gap.
Nevertheless, the goal of this paper is not to conduct a review of all available procedural methodologies, something that is most certainly impossible due to the magnitude and variety of the research field, but rather to systematically review a selected subset. Therefore, two assumptions have been made. First, the methodologies should be considered proactive in the sense that they are used during system development. This is important since changing a deployed AI system is significantly more time- and cost-intensive than implementing ethical principles from the outset and, thus, conflicts with the first challenge, time pressure (Raji et al., 2020). Second, the methodologies should be adaptable to agile development processes, which are “disciplined yet lightweight processes while placing human effort and experience at the core of software development” (Hoda, Salleh, & Grundy, 2018, p. 58). Therefore, the development of a software application is divided into short iterations with clearly defined responsibilities, in which features can be developed and tested while eliminating obstacles, inefficiencies, and bureaucracy. This makes agile processes the most common development method in the context of AI (Cervone, 2011). Consequently, a focus on agile processes is important since non-agile methods (such as waterfall-oriented processes) have little chance of being implemented in an agile environment because of their fundamentally contradicting nature. Therefore, this thesis considers the implementation of waterfall-oriented processes into an agile environment prone to a low perceived usefulness by definition and thus to an inevitable low adoption rate. This is circumvented by focusing on procedural methodologies that can be proactively used during AI system development in an agile environment.