Current ITS design mainly targets the introduction phase, where a core component is training and adaptation of ITS’s AI models based on given example data. Here, the aim is a high performance in a static and well defined environment. SAIL’s focus on the full life-cycle of ITS moves the current emphasis towards their sustainable long-term development in real life. In this context, the term ‘sustainable’ means that the initial system should preserve its functionality during its whole life-cycle; it should respect short-term and longterm objectives, take care of technological requirements and its cognitive and societal impact, and do so with reasonable maintenance costs, energy consumption, and cognitive costs or load for its human partner(s).

While immediate short-term and comparably narrow functional objectives are targeted in the introduction phase of ITS, broader, long-term, and non-functional objectives can be valuated in later phases of the ITS lifecycle only: How does a model behave in previously not encountered situations? Does the ITS have unwanted effects on human behavior and welfare? How can model maintenance be realised efficiently as regards energy consumption and cognitive efficiency? How can new data or demands be dealt with? Current AI systems are not reliable here and while some examples of AI mistakes make it to the news, the overarching number of AI failures is not reported but simply leads to the abandoning of systems at some point of their life-cycle.

As technological tools, AI models will always make mistakes. The challenge is that, currently, these mistakes are not aligned with human expectation and objectives. There are no established design principles in ITS which take limitations of AI as well as cognitive and societal demands into account during the system’s life-cycle. The currently predominant way to adapt AI models by providing large sets of representative training data is not suitable as a natural and efficient way through which a human can shape an ITS. Within SAIL, we aim for overarching design principles to address these challenges following three themes:

Research Themes

Application Domains

Application domain (A1) targets industrial work paces which, driven by the demand of mass customization and the increasing complexity of industrial systems, are in transition towards modular and flexible architectures. This transformation is fueled by recent advancements in the fields of artificial intelligence, collaborative robotics and industrial automation. It paves the way towards intelligent production systems, which are capable of AI-driven self-configuration and -optimization. At the same time, it enables new forms of hybrid work, in which human workers are supported by intelligent machines. Future ITS are expected to adapt to the user and act as assistants, being able to adapt and behave resiliently.

Application domain (A2) targets adaptive healthcare assistance systems. Healthcare is one of the largest economic sectors in Germany, contributing more than 12.1% of the gross domestic product in 2020. Rapid growth is expected in this sector due to the ageing population, hence rising public healthcare costs and increasing the need for well-qualified doctors, nurses, and professional caregivers. The ongoing digital transformation and introduction of AI technologies has the potential to severely reduce costs and improve quality. Example applications are predicting drug-drug interactions, knowledge extraction from electronic health records, administrative applications, diagnosis and screening, treatment assistance, telemedicine, patient engagement and adherence, ambient assisted living, or healthcare robots. One salient area where AI is already used successfully is diagnosis and screening, e.g. to detect cancer in medical imageing. Yet, many proposed AI solutions for healthcare are not yet mature enough to be used outside the lab or specific research settings. Moreover, AI solutions often face ethical or social issues, related to data privacy, fairness, or user acceptance.