Studie

LLM Studie, suited for students with excellent programming skills

Topic:

When translating from English to German, LLMs can degender individuals, replacing gender-specific forms with gender-inclusive ones where doing so is inappropriate. For example, translating three transgender women doctors as drei transgender Ärzt*innen instead of the correct feminine drei transgender Ärztinnen. This strips away a specific, intended gender identity under the guise of inclusivity.

This phenomenon was identified but underexplored in our recent work on the GLITTER benchmark for gender-fair English–German machine translation.

In this group study, students will systematically investigate degendering bias in LLMs through four complementary methods:

• Prompting: Probing models via zero-shot, few-shot, and chain-of-thought strategies to characterize when and why degendering occurs.
• Robustness checking: Testing whether degendering behavior is consistent across paraphrases, contexts, and model sizes, or whether it is brittle and context-dependent.
• Output annotation: Developing and applying an annotation scheme to classify model outputs by gender form (gender-specific, gender-fair, degendered, erroneous).
• Circuit analysis: Using interpretability methods to identify which internal model components drive degendering decisions.

Findings of this study will culminate in a publication in conferences or workshops like GeBNLP, GITT, *CL, and so on.

Logistics

• Start: Late May
• Duration: 4–5 months. You set your own intermediate deadlines; the pace is flexible.
• Group size: 3–6 students

Deliverables

• Oberseminar presentation of your final results.
• GitHub repository containing code, data, and documentation.

Prerequisites

• Programming: Working knowledge of Python.
• NLP: Familiarity with NLP fundamentals, or actively enrolled in a course on Trustworthy AI.
• Tooling: An agentic IDE such as VS Code, Cursor, or Antigrativity.

Contact: If you are interested, please reach out to Pranav with a short note on your background and motivation.

Policy studie, suited for students with excellent social science skills

Topic:

AI-powered surveillance is expanding in Hamburg. Since 2023, Hamburg Police have been trialling “Intelligent Video Observation” (IVBeo) at Hansaplatz and Hauptbahnhof, with the project running through August 2026. Yet a growing body of evidence suggests that these systems often do not deliver on their stated goals. If the technology does not work, then what is it actually doing, and who benefits from its continued deployment?

In this group study, students will investigate Hamburg’s AI surveillance landscape through community-engaged research. The core focus is community engagement and co-production: identifying and collaborating with Hamburg-based civil society groups (e.g., CCC Hamburg, Digitalcourage, neighborhood organizations in St. Georg) to understand community perspectives on surveillance and co-develop resources that serve their needs.

Students can form subtopics within this theme; for example, surveillance mapping, policy and discourse analysis, grassroots organizing, or effectiveness critique; depending on their interests and skills.

Findings of this study will culminate in a publication in conferences or workshops like FAccT or AIES.

Key Resources:

• Hamburg Police IVBeo project page
• Surveillance under Surveillance
• Weizenbaum Institute — AI and Surveillance at the Border
• Kalluri et al. (2025). Computer-vision research powers surveillance technology. Nature.
  

Logistics

• Start: Late May
• Duration: 5–7 months. You set your own intermediate deadlines; the pace is flexible.
• Group size: 6–10 students.
• Presence: You should be based in Hamburg for the duration of the project. Regular attendance at group meetings is expected.

Deliverables

• Oberseminar presentation of your final results.
• A community-facing output co-developed with partner organizations (e.g., public report, blog, informational resource).

Prerequisites

• German language: Strong reading and spoken German is essential. Much of the primary material — police communications, media coverage, legal documents, and community engagement — is in German.
• Social science skills: Familiarity with qualitative methods (interviews, discourse analysis, participatory research), or actively enrolled in a relevant course.
• Interest in AI ethics: No prior technical AI background is required, but curiosity about the societal implications of AI systems is a must.

Contact: If you are interested, please reach out to Pranav with a short note on your background and motivation.