Traditional recommendation systems (e.g. those used by streaming services or online shops) recommend content or products to users on the basis of an algorithm, the inner workings of which, however, remain a ‘black box’ to them and over which they have no active influence. In contrast, current research is highlighting a new paradigm: so-called conversational recommender systems, which enable a genuine exchange between humans and the system. Users can express their preferences, reject or refine suggestions, and thus actively shape the recommendation process. This interactive approach aims to overcome the inherent lack of transparency in traditional recommendation algorithms – with the aim of making recommendations not only more accurate, but also more comprehensible and trustworthy for users.

This thesis will first provide a structured overview of the academic literature on such dialogue-based and controllable recommendation systems. It will then propose a scenario in which this type of recommendation system offers clear added value compared to a traditional system.

Further reading: Jannach, D., Manzoor, A., Cai, W., & Chen, L. (2021). A Survey on Conversational Recommender Systems. ACM Computing Surveys, 54(5), 1–36.

https://doi.org/10.1145/3453154

Large Language Models (LLMs) have made enormous strides in recent years and are now used in a wide range of applications. Nevertheless, they continue to suffer from a key problem: a tendency to hallucinate, that is, to generate factually incorrect yet convincingly worded responses. Retrieval-Augmented Generation (RAG) systems have established themselves as a promising solution, in which relevant external knowledge sources are made available to the language model at runtime to improve the quality and factual accuracy of the generated responses. However, how well this works depends directly on the quality of these knowledge sources: if the data is incomplete, out of date, contradictory or poorly structured, the system may still provide factually incorrect answers. Furthermore, users are often left in the dark regarding the quality of the underlying data – a lack of transparency that further undermines trust in RAG-based systems. Consequently, the quality of the underlying data becomes a key challenge in the deployment of RAG systems – one that has, however, often been underestimated in previous research.

As part of this thesis, the aim is first to provide a structured overview of the academic literature on the challenges and limitations of RAG systems – including issues relating to data quality, the timeliness of knowledge sources, and the reliability of retrieved information. Subsequently, the aim is to devise a scenario in which the use of a RAG system is appropriate, and to highlight which of these challenges become particularly relevant in this context and how they can be addressed.

Further reading: Müller, L., Holstein, J., Bause, S., Satzger, G., & Kühl, N. (2025). Data Quality Challenges in Retrieval-Augmented Generation. Proceedings of the 26th International Conference on Information Systems (ICIS), Nashville, USA.

https://aisel.aisnet.org/icis2025/da_bus/da_bus/9/

The great promise of human-AI collaboration is complementarity: together, humans and AI are expected to achieve better results than either could achieve on their own. In practice, however, this happens surprisingly rarely – often, the team ends up being only as good as its stronger partner. The question of when and how genuine complementarity arises is one of the key unresolved issues in research.

This thesis will first provide a structured overview of the academic literature on human-AI complementarity and hybrid intelligence. The aim is to identify, from the literature, the conditions under which complementary team performance arises. Subsequently, a specific system will be presented as an example that implements these conditions and thus enables genuine collaboration between humans and AI.

Further reading: Hemmer, P., Schemmer, M., Vössing, M., & Kühl, N. (2021). Human-AI Complementarity in Hybrid Intelligence Systems: A Structured Literature Review. PACIS 2021 Proceedings.

https://aisel.aisnet.org/pacis2021/78

Human-AI collaboration is usually judged solely on the quality of the immediate outcome. However, an intriguing question – and one that is particularly important in an educational context – is: how can people themselves learn and improve by working with AI? Whilst traditional AI systems are primarily geared towards efficiency and optimising results, the deliberate design of human-AI interactions opens up a further dimension: AI as a personal learning companion. Such systems do not merely provide ready-made solutions, but explain procedures, offer targeted feedback and adapt to the learner’s individual level of knowledge. This offers considerable potential, particularly for less experienced learners or employees – for example, in continuing professional development or education – as AI systems can provide continuous and personalised support in areas where traditional teaching formats reach their limits.

This paper will first provide a structured overview of the academic literature on learning and knowledge transfer effects in human–AI collaboration, with a particular focus on educational and training contexts. It will then outline a scenario for how AI can be deployed in a learning or training setting so that it sustainably promotes the development of knowledge and skills, rather than undermining it.

Further reading: Spitzer, P., Kühl, N., & Goutier, M. (2022). Training Novices: The Role of Human-AI Collaboration and Knowledge Transfer. Workshop on Human-Machine Collaboration and Teaming (HM-CaT 2022), ICML 2022. (ergänzend: arXiv:2207.00497)

Artificial Intelligence (AI) has surpassed humans in diagnosing X-rays or playing chess. At the same time, we see headlines about AI systems being used inappropriately and making discriminatory decisions. For example, when job applications are filtered using an AI system and, as a result, only male applicants are selected. Consequently, consideration of fairness and bias in the development of such systems has already become significantly more important. There are various ways of defining when an AI system is fair, including the concept of ‘counterfactual fairness’.

The aim of this thesis is to provide a comprehensive overview of the academic literature on counterfactual fairness. Subsequently, an original scenario will be devised to illustrate how counterfactual fairness could contribute to the fairness (with a focus on explainability) of AI systems.

Further reading: Kusner et al. (2017). Counterfactual fairness. Advances in neural information processing systems.

https://proceedings.neurips.cc/paper/2017/file/a486cd07e4ac3d270571622f4f316ec5-Paper.pdf

Large Language Models (LLMs) are increasingly being used in areas of social and economic relevance, such as decision support, knowledge management and automated dialogue systems. At the same time, there is often limited knowledge of the – frequently very large and heterogeneous – datasets on which these models have been trained. This creates a risk that LLMs may reproduce or exacerbate existing societal biases, even if these are not explicitly present in the training data.

With the proliferation of Retrieval-Augmented Generation (RAG) systems, this problem is further exacerbated, as biases can arise not only from the language model itself but also from the connected external data sources. Bias and fairness thus become central challenges for data-driven generative AI systems.

The aim of this thesis is to provide a structured overview of the academic literature on bias detection and bias mitigation in generative AI systems. It will present and critically evaluate existing approaches to detecting and reducing bias in large language models and RAG systems. Particular attention is paid to the underlying causes of bias, such as the composition and lack of transparency of training data, the model architecture, and the selection and structure of external data sources in RAG systems. Finally, open research questions and implications for the responsible use of such systems in sensitive application areas will be discussed.

Further reading: Dai, S., Xu, C., Xu, S., Pang, L., Dong, Z., & Xu, J. (2024). Bias and Unfairness in Information Retrieval Systems: New Challenges in the LLM Era. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24), 6437–6447.

https://dl.acm.org/doi/pdf/10.1145/3637528.3671458

With the increasing use of complex AI systems, there is a growing need for transparent and trustworthy decisions. Whilst many approaches to Explainable Artificial Intelligence (XAI) rely on post-hoc explanations, the concept of ‘Explainability by Design’ is increasingly coming to the fore; this involves incorporating explainability right from the design stage of models and architectures.

The aim of this thesis is to provide an overview of the academic literature on Explainability by Design. It will present explainable model and system architectures and compare them with traditional post-hoc explanation methods. Finally, it will discuss the use cases in which Explainability by Design offers added value.

Further reading: Swamy, V., Frej, J., & Käser, T. (2025). Viewpoint: The Future of Human-Centric Explainable Artificial Intelligence (XAI) is not Post-Hoc Explanations. Journal of Artificial Intelligence Research, 84, Article 2.

https://www.jair.org/index.php/jair/article/view/17970/27222

With the increasing use of generative AI systems across an ever-wider range of applications, a fundamental question arises: how can the quality of their outputs be assessed systematically and at scale? Traditional evaluation methods – such as human annotation or rule-based metrics like BLEU or ROUGE – are increasingly reaching their limits in the face of the diversity, openness and semantic depth of generative content. In recent research, the ‘LLM-as-a-Judge’ paradigm has established itself as a promising approach: this involves using large language models as automated evaluators to assess the outputs of AI systems – be it generated text, programme code, responses from dialogue agents or even the actions of autonomous systems. However, this approach presents significant methodological challenges: can language models act as fair, consistent and unbiased evaluators, or do they reproduce their own biases in their assessments?

This thesis aims first to provide a structured overview of the academic literature on the ‘LLM-as-a-Judge’ approach. The aim is to identify both the methodological foundations and key challenges – such as position bias, verbosity bias and self-preferencing. Subsequently, a concrete scenario will be designed in which AI-based evaluators can be usefully deployed, and the limitations of this approach will be critically discussed.

Further reading: Zheng, L., Chiang, W.-L., Sheng, Y. et al. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. Advances in Neural Information Processing Systems, 36. arxiv.org/abs/2306.05685

Although AI systems have been shown to make more accurate predictions than human experts in many areas of application, research reveals a remarkable phenomenon: people often tend to mistrust algorithmic recommendations and prefer human judgement instead – a behaviour known as ‘algorithm aversion’. At the same time, there is the opposite phenomenon of ‘algorithm appreciation’, in which users trust algorithmic recommendations even more than human ones. Understanding which factors – such as the task context, explainability, transparent communication of uncertainty, or previous experiences with the system – influence trust in algorithmic recommendations is of central importance, particularly for the responsible use of AI in decision-making processes.

As part of this thesis, a structured overview of the academic literature on the LLM-as-a-Judge approach will first be provided. This will highlight both the methodological foundations and key challenges – such as position bias, verbosity bias and self-preferencing. Subsequently, a specific scenario will be devised in which an LLM-as-a-Judge approach can be usefully applied to evaluate AI outputs, and a critical discussion will take place on the requirements for the design of the evaluation process in order to achieve reliable and unbiased results.

Further reading: Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. doi.org/10.1037/xge0000033

Many AI systems are not developed just once, but can be continuously improved through human feedback. In human-in-the-loop systems, users flag errors, correct suggestions, provide new examples or contribute specialist knowledge so that the system better suits the specific context of use. However, for this feedback to be helpful, the systems must be designed in such a way that users understand when and why the AI might be wrong, and what kind of feedback is required. Comprehensible explanations, suitable feedback options and clear interaction processes that provide guidance even to non-technical users play a key role in this. It is therefore not only crucial that humans are ‘in the loop’, but also how this loop is designed so that user feedback can actually contribute to the improvement of AI systems.

This thesis aims to demonstrate, through a structured literature review, which design principles are necessary for human-in-the-loop systems to enable users to provide meaningful feedback to AI systems. Building on this, the thesis will examine the role played by explainable AI, interactive learning and user-friendly feedback mechanisms in this context.

Further reading: Mosqueira-Rey, E., Hernández-Pereira, E., Alonso-Ríos, D., Bobes-Bascarán, J., & Fernández-Leal, Á. (2023). Human-in-the-loop machine learning: a state of the art. Artificial Intelligence Review, 56(4), 3005-3054.

https://doi.org/10.1007/s10462-022-10246-w

AI can analyse large amounts of data, recognise patterns and use them to make impressively accurate predictions. However, finding a correlation does not automatically mean that one thing is the cause of the other. Here’s a clear example: if people who regularly drink a glass of wine live longer in some datasets, this does not necessarily mean that wine is healthy. It is possible that these people also differ in other respects, such as having a higher income, better access to healthcare or more social contacts. This is precisely where a key challenge of data-driven decision-making lies: AI can often predict well what is likely to happen, but cannot readily explain why it happens. Causal AI approaches address this by helping to better identify cause-and-effect relationships and derive more robust recommendations from them.

As part of this thesis, a structured literature review will be used to demonstrate why traditional AI models often only identify correlations and are unable to draw reliable conclusions about causality from them. Furthermore, the thesis will examine how approaches from causal machine learning are utilised to make AI-based decisions and recommendations more robust.

Further reading: Feuerriegel, S., Frauen, D., Melnychuk, V., Schweisthal, J., Hess, K., Curth, A., ... & van der Schaar, M. (2024). Causal machine learning for predicting treatment outcomes. Nature Medicine, 30(4), 958-968.

https://doi.org/10.1038/s41591-024-02902-1

Time series forecasting is relevant to many fields of application, such as energy consumption, share prices, sales planning, supply chains, weather data and demand trends. In all these areas, the aim is to derive the best possible predictions for the future from past trends. There are various methods for this – ranging from simple statistical techniques to modern machine learning and AI approaches. These methods differ in terms of the data they require, how well they identify trends or seasonal patterns, how easy they are to understand, and the practical questions for which they are particularly well-suited. At the same time, not every complex model is automatically better, as forecasts depend heavily on data quality, the field of application and the forecast horizon.

The aim of this thesis is to use a structured literature review to demonstrate which key methods exist for time series forecasting, how they differ, and what their advantages and disadvantages are. Building on this, the thesis will identify the practical application areas for which individual methods are particularly suitable and the criteria that should be taken into account when selecting an appropriate forecasting method.

Further reading: Kontopoulou, V. I., Panagopoulos, A. D., Kakkos, I., & Matsopoulos, G. K. (2023). A Review of ARIMA vs. Machine Learning Approaches for Time Series Forecasting. Future Internet, 15(8), 255.

https://doi.org/10.3390/fi15080255

Quantum computing is regarded as one of the most exciting technologies of the future because it could solve certain complex problems significantly faster or in entirely new ways than classical computers. For a long time, the use of quantum technologies was primarily limited to research and experimental laboratory settings. In recent years, however, considerable technological progress has been made, meaning that gradual commercial adoption is looking increasingly realistic. For businesses, it is not so much the technical workings that are decisive, but rather the question of what specific applications quantum computing opens up. Areas where a large number of potential solutions need to be compared, complex systems simulated or large volumes of data analysed are particularly relevant, such as logistics, financial markets, production planning, risk analysis or materials development.

This thesis aims to demonstrate, through a structured literature review, which economically relevant use cases for quantum computing are currently being discussed and what potential these hold for businesses. Building on this, the thesis will identify the areas in which quantum computing appears particularly promising and the current limitations to its practical application.

Further reading: Descazeaux, I. (2025). Mapping the narratives of quantum computing in IS: A literature re-view. Proceedings of the 46th International Conference on Information Systems.

https://aisel.aisnet.org/icis2025/quantum/quantum/5

Generative AI is being used more and more frequently to write texts, develop ideas, summarise information or prepare decisions. As a result, it can make everyday work and study significantly easier and boost productivity. At the same time, the question arises as to whether people will practise key skills less, or lose them in the long term, if they increasingly delegate such tasks to AI systems. This could be particularly problematic if skills such as critical thinking, problem-solving, judgement or subject-specific learning are weakened. It is therefore important to understand under what conditions GenAI acts as a helpful support and when its use can lead to a loss of important skills.

This thesis aims to use a structured literature review to demonstrate whether and how the use of Generative AI can contribute to the loss or weakening of key skills – in other words, to ‘deskilling’. Building on this, the thesis will identify which skills are likely to be particularly affected and how GenAI can be deployed in such a way that it supports learning and skills development rather than replacing them.

Further reading: Ferdman, A. (2026). AI deskilling is a structural problem. AI & SOCIETY, 41(4), 3001-3013.

https://doi.org/10.1007/s00146-025-02686-z