Online discussions can be rough, especially when it comes to political issues. They are often characterized by a harsh tone which discourages many people from participating in them at all. At the same time, these discussions are very important for democracy as they promote exchange and help individuals form their own opinions. While Artificial Intelligence (AI) may be detrimental to the quality of discussions (e.g. when used in spam bots), it also offers a promising opportunity to support constructive and inclusive discussions, for example by making them more civil. To strengthen such discursive integration we have engaged in a co-creation process with non-academic stakeholders to develop a discussion assistant prototype that i) identifies likely problematic comments for a possible rephrasing and ii) offers authors help with reformulation by letting generative AI suggest improvements like more civil wording. In this paper, we describe the process of co-creative research and the current status of the discussion assistant, which is still being developed and improved.

Large language models (LLMs) can support democratic deliberation at scales previously constrained by turn-taking and facilitation bandwidth. Recent work shows that LLM-generated group statements are often preferred over human-mediated outputs, while theoretical analyses argue that LLMs relax the simultaneity constraints limiting collective intelligence. Yet pure LLM mediation risks collapsing pluralism, over-optimizing for agreement, and undermining legitimacy when participants cannot contest how they are represented. We propose a symbiotic human-AI framework organized into three layers: observation and diversity amplification, facilitation with clause-level provenance, and human primacy for ratification. Our contributions include graded coverage, diversity, and erasure metrics with salience-aware weighting; a provenance pipeline combining cross-encoder similarity with causal knockout diagnostics; preference-conditioned trade-off control; equity-aware contestability workflows; adversarial robustness tests; and an evaluation protocol with ablation designs informed by evidence of LLM-as-judge limitations. The result is a testable blueprint for deliberation technology that scales collective intelligence while preserving agency and legitimacy.

Explainability in deliberation-support NLP is usually evaluated through post-hoc rationales or model-internal attribution methods, and only rarely against explicit institutional reasoning procedures. We introduce , a Romanian corpus of professional fact-checking reports that preserves the workflow of editorial epistemic arbitration, namely claim articulation, contextualisation, verification scope, evidence-based verification narrative, and calibrated conclusion. contains 789 raw reports from factual.ro and a processed benchmark release of 788 instances after removal of a singleton non-standard verdict label. Beyond six-way verdict prediction, we position as a benchmark for LLM explanation alignment, where models must generate short explanations that can be compared directly to gold institutional reasoning. We evaluate primarily with instruction-tuned LLMs, reporting full-corpus experiments for open-weight models and a matched pilot comparison with GPT-4 Turbo. The resulting evidence shows that verdict prediction and institutional explanation alignment are not the same capability: models that improve verdict accuracy do not necessarily preserve institutional calibration or produce explanations that align with professional verification narratives. These results support the central claim of the paper, namely that measures not only whether a model reaches a verdict, but also whether it does so in a manner that resembles documented public reasoning.

LLM alignment has progressed in single-agent settings through paradigms such as RL with human feedback (RLHF), while recent work explores scalable alternatives such as RL with AI feedback (RLAIF) and dynamic alignment objectives. However, these approaches remain limited in multi-stakeholder settings, where conflicting values arise and deliberative negotiation is required. This work proposes a multi-agent negotiation-based alignment framework that aligns LLMs to Collective Agency (CA)—an existing alignment objective introduced to promote the continual expansion of agency—while simultaneously improving conflict-resolution capability. To enable scalable training, two self-play LLM instances are assigned opposing personas and engage in turn-based dialogue to synthesize mutually beneficial solutions. We generate synthetic moral-dilemma prompts and conflicting persona pairs, and optimize the policy via RLAIF using Group Relative Policy Optimization (GRPO) with an external LLM reward model. While rewards are computed from CA scores assigned to the final completion, gradients are applied to dialogue tokens to directly improve deliberative interaction dynamics. Experiments show that the model achieves CA alignment comparable to a single-agent baseline while substantially improving conflict-resolution performance without degrading general language capabilities. These results suggest that negotiation-driven deliberation training provides a practical path toward LLMs that better support collective decision-making in value-conflict scenarios.