This thesis invites you to work on a scientifically novel and disruptive idea at the intersection of NLP, LLM development, and democracy research: How can we use large language models not just to analyze political text, but to simulate political debates and design an AI moderator that actively improves discourse?

Recent research shows that LLMs can role-play political actors and engage in multi-round debates—sometimes persuading each other, sometimes drifting into biased or flawed reasoning. At the same time, LLM-based tools for online dialogue moderation (e.g., reframing, conflict de-escalation) are emerging. Yet no one has combined these two worlds: LLM political agents and an AI moderator designed to reduce polarization, prevent fallacies, and foster constructive discussion.

In this Master thesis, you will build one of the first systems capable of simulating democratic debate dynamics using multiple LLM agents and evaluating how an AI moderator shapes the outcome. The goal is both methodological innovation and societal relevance. The work is technically challenging and has strong publication potential.

What are the tasks?

Build a multi-agent LLM debate simulation

• Implement two or more political “agents” (LLM personas) representing different ideological positions.
• Design prompts or fine-tuning strategies to maintain stable viewpoints across debate rounds.
• Create a turn-based debate mechanism in which agents argue about a political or societal issue.

Develop an AI moderator agent

• Implement a third LLM agent acting as a moderator that intervenes during debates.
• Define moderation strategies (e.g., de-escalation, reframing, fact reminders, prompting for common ground).
• Experiment with different moderator personas (strict, neutral, consensus-oriented, fact-checker).

Run simulation experiments on diverse political topics

• Test debates on multiple issues (e.g., climate, freedom of speech, economic policy).
• Compare debates with vs. without the moderator to study its influence.
• Optionally explore multilingual or cross-cultural debates.

Evaluate discourse quality and polarization

• Define metrics for:
  • opinion divergence & convergence,
  • tone and toxicity,
  • argumentative quality,
  • logical fallacies,
  • factual consistency.
• Analyze how the moderator affects debate outcomes.
• Compare patterns to human debates (e.g., using recent debate datasets).

Analyze and interpret results

• Identify mechanisms through which the moderator improves (or worsens) debate quality.
• Discuss implications for democratic discourse, online platforms, and responsible AI design.

What prerequisites do you need?

• Strong interest in NLP, LLMs, and generative AI.
• Good programming skills in Python; experience with transformers / Hugging Face is helpful.
• Curiosity for political communication and how LLMs reason, persuade, or mediate.
• Ability to work independently and explore unconventional research ideas.
• Very good English reading and writing skills.

Why this thesis is special

• Scientifically disruptive: Hardly any research exists on AI-moderated LLM debates.
• High societal relevance: Findings can inform future systems for democratic dialogue, online moderation, or civic education.
• Strong publication potential: Combines multi-agent LLMs, dialogue modeling, and evaluation of political discourse.
• Innovative and creative: Instead of analyzing text, you create a new interactive simulation environment for democratic processes.