Lately, the development of artificial intelligence (AI) has been dominated by a technology-centric design approach. The novel approach of "Hybrid Intelligence", which stands out from this, focuses on an integrative method that integrates both the human and the algorithmic-technical perspective. The present work examines a practical application example of the realization of "Hybrid Intelligence". In the context of graphical image cropping, a recommendation system has been developed in a real operational application field, which suggests algorithmically generated image cropping to users using AI technology. The aim is to find a form of interaction and a degree/level of AI integration that takes into account and optimally combines both human and algorithmic skills. The goal is to create an interactive, intelligent system that is more efficient than either AI technology or humans alone would be. To analyze, concretize, and evaluate this automation problem, the work is guided by Mackeprang et al.'s seven-step process for building interactive intelligent systems. At the end of this process, the identification of the so-called sweet spot configuration will take place, which describes the optimal combination of human and AI collaboration. For this purpose, first, all possible configurations are defined, each differing in the form of interaction and depth of AI integration. In the next step, these configurations are transferred into functional program code and their performance is evaluated in user tests. The evaluation is based on the time required per image section, the quality of the algorithmically generated image sections, the workload subjectively perceived by the users, and the benefits for users in dealing with the system. The configuration with a medium level of automation was identified as the best performing of all the tested variants. This corresponds to an automation level of 5 according to Sheridan et al.. In addition to a time saving of 62% compared to the initial configuration, which corresponds to a completely manual image cropping by the users, it also enables a 60% lower perceived workload. At the same time, an optimized user-friendliness as well as the lowest correction effort for the algorithmically generated image cropping were shown. The results prove that the selected methodological approach is suitable for optimizing the synergetic potential of hybrid intelligence while implementing AI-supported automation.

Link to Thesis