Dr. Dimitry Mindlin, Project B01, Computer Science
In his work, Dr. Dimitry Mindlin shows that it is not interactivity as such but the concrete design of communication that determines how well users understand a machine learning model. The work thereby supports the co-constructive approach pursued at TRR 318.
Key finding:
People understand an AI model measurably better when they can ask follow-up questions about its decisions instead of only reading a finished explanation. A simple chatbox is not enough, however, especially when the number of follow-up questions is limited. The value of natural language emerges only when the system continuously monitors users' understanding, adapts the explanation accordingly, and takes initiative itself.
Significance for TRR 318:
This work provides empirical evidence for the fundamental assumption of TRR 318 that explanations should be co-constructed rather than conveyed unilaterally. It demonstrates that the design of communication influences understanding, regardless of the explanation method used. It also shows how co-construction can be calibrated to context: when it is fully required and when simpler forms suffice. In addition, Mindlin developed an explanation system that continuously tracks what users have already understood during the conversation, identifies remaining gaps, and tailors the next explanation step accordingly. Rather than only responding to questions, it also takes the initiative itself. The system is a first technical step towards co-construction and provides a foundation on which further work within TRR 318 can build.
Impact on (AI) Practice:
Interactive explanation systems should not only offer better explanation methods but also deliberately design the communication around them. When designing AI systems, continuity, adaptivity, and initiative should be treated as factors of understanding in their own right, not as add-ons to the explanation method.
Dr. Helen Beierling, Project B05, Computer Science
In her doctoral dissertation, Dr. Helen Beierling investigated how laypeople can train robots more effectively by using what is known as “co-constructive training.” In this approach, users receive support through various feedback methods and “augmented reality,” which helps them better understand the learning process. The goal of this approach is to develop interactive robot systems that flexibly adapt to users’ individual levels of understanding during training.
Key finding:
The most important finding is that using multiple interaction methods improves learning. However, users do not always apply these methods in a way that maximizes their actual benefits. Furthermore, the study shows that transparency promotes understanding but does not automatically lead to better performance. The research also systematically identified which concepts laypeople are unaware of or misinterpret regarding robot training.
Significance for TRR 318:
This work contributes to TRR 318 by demonstrating how mental models and understanding processes emerge and can be influenced in interactions with learning systems. It also offers a methodology for systematically identifying which key concepts should be explained to a system. It thus provides initial insights into how AI training can be designed in a co-constructive manner.
Impact on (AI) Practice:
The results provide concrete design implications for AI systems. In particular, they show that flexible interaction options improve training and that explaining invisible concepts is crucial for understanding. In the long term, such approaches can contribute to the development of adaptable and user-centered AI systems in fields such as robotics or assistive systems.
Dr. Amit Singh, Project A05, Psycholinguistics
In his doctoral dissertation, Dr. Amit Singh examined how contrastive cues—particularly negations—influence our attention, memory, and the learning of actions in dynamic situations. His work integrates insights from psycholinguistics, cognitive science, human-robot interaction, and XAI.
Key Finding:
One of the most important findings was that negations—when used at the right moment—can draw on and leverage the context already established in an ongoing interaction. They help explain things by contrast, that is, by clarifying what is meant by distinguishing it from alternatives. In this way, they can enhance understanding because differences become clearer and more explicit.
Significance for TRR 318:
The project contributes to understanding explainability as a dynamic, social, and context-dependent process. It demonstrates how explanations emerge through interaction and are supported by contrasts on multiple levels.
Impact on (AI) Practice:
Interactive AI systems, such as robots or assistive systems, can strategically use contrasts in appropriate situations to direct attention, support learning processes, and make collaboration between humans and AI more effective.
Dr. Maximilian Muschalik, Project C03, Computer Science
Dr. Maximilian Muschalik’s doctoral dissertation addresses a central topic in explainable AI: feature-based explanations. These show which features an AI model used to make its decision. The dissertation further develops this approach by also explaining the interplay between different features and introducing new methods for use in dynamic and particularly challenging application scenarios.
Key finding:
Interaction-based explanations are particularly important for making complex AI models understandable. Modern AI systems often process different types of information—such as images, text, or tabular data—and make their decisions based on how these interact. When this information is considered in isolation, the decision often cannot be adequately explained. Therefore, the interactions between the various information sources must also be taken into account.
Significance for TRR 318:
The work provides fundamental methods for explainable AI that are utilized in TRR 318’s explanation strategies. The developed approaches are already being incorporated into various research projects within the Transregio. A significant portion of the methods has been implemented in the open-source library “shapiq,” which is already in use in several TRR projects.
Impact on (AI) Practice:
The explanation methods developed in this dissertation are already being used in AI practice. Researchers and users are increasingly relying on them to better understand complex AI models. This is evident, among other things, in the growing use of “shapiq” as well as in various collaborative projects. As increasingly powerful AI models become more widespread, the importance of methods that can explain their decisions in a comprehensible way is also growing.
Dr. Maximilian Spliethöver, Project C07, Computer Science
In his doctoral dissertation, Dr. Maximilian Spliethöver investigated how social biases—that is, discrimination against social groups or their members—manifest in NLP models and the underlying data. NLP stands for “Natural Language Processing” and refers to technologies that enable computers to process and understand human language, such as language models. In his research, he explored and developed approaches for measuring social biases in training data, detecting them in individual texts, and reducing them in both the data and the models.
Key finding:
Social biases are highly context-dependent. Therefore, cultural specificities must also be taken into account, and targeted, context-sensitive solutions must be developed. There is likely no one-size-fits-all solution capable of eliminating all biases from NLP technologies or other AI applications.
Significance for TRR 318:
Above all, the computational projects within TRR 318 will utilize NLP technologies and, in particular, large language models (LLMs). It is therefore important to be aware of such fundamental issues with the underlying technologies and to have the means to investigate and mitigate them—including in the context of explanation. One example would be the unintended use of stereotypes in explanations.
Impact on (AI) Practice:
As NLP technologies become increasingly important—especially in critical application contexts that can have real-world consequences—it is essential to ensure that these technologies and the tools based on them create societal value for everyone. They should not benefit only a privileged few who are not affected by biased or unfair NLP technologies.
Dr. Vivien Lohmer, Project A04, Linguistics
In her dissertation, Dr. Vivien Lohmer examined how explanations in everyday conversations are co-constructed—that is, jointly produced by the participants in a multimodal manner. A central aspect of this is the discursive structure of explanations: How are explanations jointly produced in an interaction, and how do the participants shape their progression and the end of an explanation?
Key finding:
Explanations do not simply follow a predetermined structure but are repeatedly negotiated jointly by the participants as the conversation unfolds. Before the participants delve into the substantive aspects of an explanation, they first jointly clarify why and in what form an explanation is relevant. In particular, the end of an explanation does not come abruptly but is constructed reciprocally by the participants and, in some cases, even repeated.
Significance for TRR 318:
For explanations between humans and AI, it is not only the content that is decisive, but also the discursive structure and multimodal realisation of the interaction. Since explanations are negotiated in a co-constructive manner between the participants, recurring interactive structures can be identified in everyday explanations. Particular attention is paid to the end of an explanation: Here, the participants have the opportunity to interactively display their own understanding and thereby clarify potential misunderstandings.
Impact on (AI) Practice:
Adaptive and dynamic AI systems should take into account not only the content of an explanation but also its discursive structure. It has been shown that the joint negotiation of the end of an explanation takes up approximately 30–35% of the total time. Therefore, even in the case of an AI explanation, there should be sufficient opportunity to clarify one’s own understanding or ask follow-up questions. Concluding summaries interactively display what was understood from the explanation. In the event of misunderstandings or errors, an adaptive and dynamic AI system should correct them accordingly before the end of the explanation is renegotiated.
