We welcome new thesis students twice a year, at the beginning of each semester in March and September. We only accept applications during these two designated periods. Please check back ahead of the semester start for exact aplication dates. To ensure a good fit between students and supervisors, topics are limited and may be tailored to ongoing research projects within the group.
Students can write to hci-thesis.informatik@lists.hu-berlin.de to send their proposal.
Please consider the following guidelines for working on your thesis:
As generative AI tools evolve, it is important to clearly declare all the tools and aids used in your thesis work to ensure transparency in presenting your personal performance. Such transparency also acts as an important safeguards against accusations of academic fraud.
All aids must therefore be reported in the form of a table, accompanied by a signed declaration of “list of aids used”. Such table is required even in cases where no aids were used.
This table must be placed behind the authorship declaration in a separate page in the thesis and must include at least the following columns:
We hope that this will help increasing transparency and protect you against academic fraud.
*Credits go to WU Vienna
Implementing Element Recognition and an Adaptive Presentation Layer for GUIs Adaptive User Interfaces aim to provide context-sensitive adaptations that align with user needs, thereby enhancing the user experience by improving efficiency or engagement (Gaspar-Figueiredo et al., 2024). As Gaspar-Figueiredo et al. state: “According to the user’s needs, preferences, and context, UI elements such as menus, buttons, bars, icons can be modified by adjusting font size, layout, color, contrast, theme” (2024, p. 41). This could be especially helpful to declutter and simplify highly complex scientific applications that require comprehensive training to operate. However, the majority of applications with graphical user interfaces are based on varying implementations, some of which are proprietary, and thereby limit adaption, automation, and the development of new features (Dixon&Fogarty, 2010).
This project therefore aims to develop a system for the automatic recognition and modification of GUI elements regardless of their framework or implementation. The objective is to build a functional prototype that can parse user interfaces, support the dynamic adaptation of visual UI properties, such as the number, size, and labeling of control elements, in a dedicated presentation layer as well as backpropagate values to the original interface. As a foundation, the OmniParser tool can be used to identify and structure individual UI elements, as it “produces a structured, DOM-like representation of the UI and a screenshot overlaid with bounding boxes for potential interactable elements” (Lu et al., 2024). In addition, a new presentation/interaction layer is to be designed and implemented, for example by using a Python-based GUI framework. The resulting user inputs and actions are then to be transferred to the underlying interface.
Preliminary research question: How can a system for the automatic recognition of UI elements be developed to support the dynamic interface modification in a separate presentation layer, while transferring data and actions back to the underlying interface?
Further reading:
M. Dixon and J. Fogarty, “Prefab: implementing advanced behaviors using pixel-based reverse engineering of interface structure,” in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, Atlanta Georgia USA: ACM, Apr. 2010, pp. 1525–1534. doi: 10.1145/1753326.1753554.
D. Gaspar-Figueiredo, M. Fernández-Diego, R. Nuredini, S. Abrahao, and E. Insfran, “Reinforcement Learning-Based Framework for the Intelligent Adaptation of User Interfaces,” in Companion Proceedings of the 16th ACM SIGCHI Symposium on Engineering Interactive Computing Systems, in EICS ’24 Companion. New York, NY, USA: Association for Computing Machinery, Jun. 2024, pp. 40–48. doi: 10.1145/3660515.3661329.
Y. Lu, J. Yang, Y. Shen, and A. Awadallah, “OmniParser for Pure Vision Based GUI Agent,” 2024, doi: https://doi.org/10.48550/arXiv.2408.00203.
GitHub: microsoft/OmniParser: “A simple screen parsing tool towards pure vision based GUI agent,” GitHub, 2026. https://github.com/microsoft/OmniParser (accessed Sept. 14, 2026).
Misinformation and disinformation rank as the second most severe global risk over the next two years [1]. Large Language Models (LLMs) are therefore increasingly applied to support humans in fact-checking tasks, as collaborative human-AI fact-checking can durably reduce conspiracy beliefs [2] and improve misinformation detection accuracy [3]. Most LLM-based verification tools are chatbots, making free-form prompting the dominant mode of interaction, even though writing effective prompts is difficult, especially for non-experts [4, 5]. A novel approach explored graphical user interface (GUI) elements as an alternative to prompting in human-AI fact-checking. However, it remains unclear how such GUIs affect fact-checking performance and whether they help users rely on the LLM appropriately, as LLMs can make mistakes in the verdict. The goal of this thesis is to compare GUI-based and prompt-based interactions in terms of fact-checking accuracy and appropriate reliance on erroneous LLMs. The results of this thesis can inform the design of collaborative human-AI fact-checking systems that strengthen human judgment and enable more user-friendly interactions with LLMs.
[1] World Economic Forum. 2026. The Global Risks Report 2026. Technical Report. World Economic Forum, Geneva, Switzerland. https://reports.weforum.org/docs/WEF_Global_Risks_Report_2026.pdf
[2] Costello, Thomas H., Gordon Pennycook, and David G. Rand. 2024. Durably reducing conspiracy beliefs through dialogues with AI. Science 385, 6714. - Google Suche. Retrieved May 6, 2026 from https://www.science.org/doi/10.1126/science.adq1814
[3] Anku Rani, Valdemar Danry, Paul Pu Liang, Andrew Lippman, and Pattie Maes. 2026. Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26), 1–26. https://doi.org/10.1145/3772318.3790656
[4] Meredith Ringel Morris. 2024. Prompting Considered Harmful. Commun. ACM 67, 12 (Nov. 2024), 28–30. doi:10.1145/3673861
[5] J.D. Zamfirescu-Pereira, Richmond Y. Wong, Bjoern Hartmann, and Qian Yang. 2023. Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (April 2023), 1–21. doi:10.1145/3544548.3581388
Misinformation and disinformation rank as the second most severe global risk over the next two years [1]. Large Language Models (LLMs) are therefore increasingly applied to support humans in fact-checking tasks, as collaborative human-AI fact-checking can durably reduce conspiracy beliefs [2] and improve misinformation detection accuracy [3]. However, LLM assessments can be wrong, which reduces belief in truthful information [4], and human-AI collaborative fact-checking can even foster reliance on the system, thereby reducing the ability to detect misinformation without LLMs [3]. Interventions to foster critical thinking could address this, but they risk frustrating users. For instance, De Vooght et al. [5] built a chatbot that fosters critical reflection and improves the user’s ability to detect misinformation by engaging in Socratic questioning. However, the repetitive questioning was also perceived as frustrating, and the user’s expectation to receive answers conflicted with the Socratic method. This reduces user experience and could discourage users from using the system. Therefore, the goal of this thesis is to identify possible interventions that amplify critical thinking, media literacy, or appropriate reliance, and to explore novel ideas for their user-friendly application. The results of this thesis can inform the design of collaborative human-AI fact-checking systems that strengthen human judgment and enable more user-friendly interactions with LLMs.
[1] World Economic Forum. 2026. The Global Risks Report 2026. Technical Report. World Economic Forum, Geneva, Switzerland. https://reports.weforum.org/docs/WEF_Global_Risks_Report_2026.pdf
[2] Costello, Thomas H., Gordon Pennycook, and David G. Rand. 2024. Durably reducing conspiracy beliefs through dialogues with AI. Science 385, 6714. - Google Suche. Retrieved May 6, 2026 from https://www.science.org/doi/10.1126/science.adq1814
[3] Anku Rani, Valdemar Danry, Paul Pu Liang, Andrew Lippman, and Pattie Maes. 2026. Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26), 1–26. https://doi.org/10.1145/3772318.3790656
[4] Matthew R. DeVerna, Harry Yaojun Yan, Kai-Cheng Yang, and Filippo Menczer. 2024. Fact-checking information from large language models can decrease headline discernment. Proceedings of the National Academy of Sciences 121, 50: e2322823121. https://doi.org/10.1073/pnas.2322823121
[5] Edward De Vooght, Aline Duelen, Georgios Petasis, Wendy Van Den Broeck, Laurien Desimpelaere, Ike Picone, Gian Andrea Giacobone, Alessandro Pollini, Dora Katsamori, Nicklas Bang Bådum, Giannis Stamatellos, Panayiotis (Panos) Kapetanakis, Ioannis Elissaios Paparrigopoulos, and Eirini Papachristou. 2025. Building a GenAI Socratic Chatbot for Critical Thinking: Insights From Cross-Cultural Cocreation, Living Labs and Pilots. In Advances in Computational Intelligence and Robotics, Tom Hallaq and Jacob Groshek (eds.). IGI Global Scientific Publishing, 353–400. https://doi.org/10.4018/979-8-3373-0872-2.ch012