New publication on responsibility in the development of algorithmic decision-making systems (VHB A)










Responseabilty of AI








Who is responsible when AI acts autonomously and makes mistakes? Recent developments – from autonomous AI agents breaching security barriers at Hugging Face to AI-assisted discoveries of critical PHP vulnerabilities – highlight the growing importance of this question. But how can we ensure that ethical principles are considered from the very beginning of AI development?




Prof. Dr. Martin Adam and researchers from TU Darmstadt address this question in their new publication, “The role of algorithmic accountability in shaping developers’ deontological judgements and implementation behaviours,” published in the European Journal of Information Systems.




Drawing on a survey, an experiment, and interviews, the findings show that developers who expect to explain and justify their decisions are more likely to make ethically grounded choices. Technical expertise and a work environment that supports ethical principles further strengthen this relationship. In the experiment, developers facing heightened accountability were more likely to reject the implementation of an ethically problematic AI-based loan decision-making system.




Martin Adam: “As AI becomes increasingly autonomous, human responsibility must become more explicit, not less. Our research shows that accountability can shape ethical decisions before harmful AI systems ever reach the real world.”




The paper is available here:


https://www.tandfonline.com/doi/full/10.1080/0960085X.2026.2747938






The following interview highlights the main idea of the publication:



How companies can foster ethical decisions in AI development




Algorithmic systems can disadvantage people in credit scoring or fraud detection. A new study in the European Journal of Information Systems examines how companies can promote responsible decisions already during the development phase. Prof. Dr. Martin Adam explains the results and their practical implications.



Mr. Adam, what did you aim to find out with your study?




We were interested in how companies can support developers in aligning their decisions with ethical principles. Many risks arise early on—during data selection, model design, or decisions regarding deployment. At that stage, the later consequences are not yet fully visible. Therefore, it is important that employees check early on whether their decisions account for factors such as fairness and protection against discrimination.



You speak of algorithmic accountability. What does that mean concretely?




It refers to the perceived obligation to explain and justify one's development decisions and their expected consequences. Anyone using specific data or releasing a model for implementation expects to stand by that decision in front of others. Our study investigates whether this awareness strengthens orientation toward fundamental moral duties and rights. We speak here of deontological judgments: people examine whether their actions comply with ethical principles, even if a problematic system appears economically attractive.



How did you investigate this?




We combined a survey of 148 developers, a randomized online experiment with 102 participants, and 15 interviews. The research focused on the USA. In the experiment, participants decided on the implementation of a credit decision system that incorrectly classifies some applicants as gamblers, thereby disadvantaging them. We varied how clearly the user interface made personal attribution and the verifiability of their decisions visible. The interviews helped us better understand the considerations behind such decisions.



What was the central finding?




Under higher perceived accountability, 62 percent decided against implementing the problematic system. Under lower accountability, it was 41 percent. This corresponds to a difference of about 21 percentage points. Our analyses suggest that a stronger orientation toward ethical principles mediates this relationship. The decision referred to a described banking scenario. Further studies in everyday corporate environments must examine how such measures affect systems actually deployed in the long run.



What role do technical expertise and the work environment play?




The relationship between accountability and ethical judgments was stronger when developers rated their knowledge of algorithmic decision-making systems higher. Anyone who understands the data, functionality, and potential errors of a system can assess risks more concretely. A work environment that supports ethical principles also strengthened the relationship. The interviews illustrate how important team support is in this regard: employees find it easier to voice concerns when they are taken seriously and they are not alone in doing so.



What can companies change concretely?




From our results, we derive that companies should clarify responsibilities and document important development decisions in a traceable manner. Before approving a system, it can be explicitly asked, for example, which people might be disadvantaged by a design decision and how the team deals with that. To do this, employees need the technical expertise to recognize such risks. Furthermore, managers should demonstrate that ethical concerns are welcome—even if addressing them means extra work or a later launch. This makes the call for responsible action concrete in everyday development.



What message would you like to convey?




Companies shape the conditions under which developers can assume responsibility. These include traceable responsibilities, technical support, and a work environment where ethical concerns carry weight. Anyone who wants to develop responsible AI must create these conditions while it is being created.






Publication

Bartsch, S. C., Schmidt, J.-H., Alizadeh, A., Adam, M., & Benlian, A. (forthcoming). The role of algorithmic accountability in shaping developers’ deontological judgements and implementation behaviours. European Journal of Information Systems.




DOI:

https://doi.org/10.1080/0960085X.2026.2747938