DELITE workshop successfully held

On 16 May 2026, the second Workshop on Language-driven Deliberation Technology (DELITE) took place as part of the international Language Resources and Evaluation Conference (LREC) 2026 in Palma de Mallorca. The Düsseldorf Institute for Internet and Democracy (DIID) played an active role in organising and funding this event, which serves as an interdisciplinary forum for new technological developments in the field of digital deliberation.

Following the welcome and introduction by Julia Romberg, a member of the workshop organising team, DIID member Dr Carina Weinmann opened the academic programme with a keynote address on the topic ‘AI and Deliberation: A Political Communication Perspective’. From the perspective of political communication, she examined the interplay between artificial intelligence (AI) and normative deliberative ideals such as rationality, civility, reciprocity and equality. Drawing on a systematic literature review (Friess et al., 2025), she demonstrated that current AI research is heavily focused on rationality and civility (for example, through the automated detection of hate speech), whilst aspects of reciprocity and equality have so far been neglected. The presentation sparked a lively discussion, covering topics such as the role of emotions in rational argumentation, how to deal with disinformation in public discourse, and the extent to which AI should be integrated into normative models of democracy as an autonomous actor.

Maike Behrendt and her research team then presented the INDI project and an associated AI-based assistance tool (‘INDI Discuss Bot’), which aims to promote discursive integration in online comments. Rather than imposing restrictive guidelines on users, the prototype suggests constructive and more civilised rephrasings. The discussion explored the extent to which explanations for these rephrasings could increase user acceptance.

This was followed by a presentation by Wajdi Zaghouani on a human-AI framework for large-scale democratic online deliberation. The three-layer model presented seeks to strike a balance between scalability through AI support and the legitimacy of human judgement by specifically amplifying diversity of opinion rather than compressing it.

Following a coffee break, Panatchakorn Anantaprayoon and colleagues presented a multi-agent-based alignment framework designed to align large language models (LLMs) with collective values through negotiation processes involving opposing personas. This demonstrated promising approaches to improving the conflict-resolution capabilities of AI models in complex value conflicts.

The focus then shifted to the practice of parliamentary debates. Using the ‘Legibot’ tool, Jean-Philippe Cointet, speaking on behalf of his research team, presented a framework that utilises LLMs to annotate and analyse extensive text data from the French National Assembly with regard to deliberative quality characteristics and adversarial practices.

The subsequent presentation by Segun Taofeek Aroyehun compared the performance of fine-tuned encoder models with zero-shot LLMs in identifying prosocial and polarising discourse attributes in public online discussions. It emerged that attributes such as respect and empathy pose the greatest challenge for the models.

The workshop concluded with a panel discussion. Participants discussed the urgent legal, ethical and practical challenges at the interface between AI-based deliberation technology and practice. Particular emphasis was placed on the need for realistic risk assessments prior to tool development, greater incorporation of the user perspective, and the crucial importance of transparent, external scientific communication to help the public understand the limitations and potential of AI systems.

The workshop concluded with a feedback session, during which the participants praised the in-depth professional exchange across disciplinary boundaries and the productive atmosphere. The approaches discussed provide valuable insights for the future design of smart participation processes.