Led by Make.org, Sciences Po, Sorbonne University and CNRS, and backed by Bpifrance's "Digital Commons for Generative AI" France 2030 call, the programme brings together more than 50 researchers and engineers working on one question: can generative AI be built, evaluated, and corrected so that it strengthens democratic participation rather than distorting it? This latest phase delivered four concrete advances.
Characterizing democratic bias. Sciences Po's Medialab has deepened its work on the characterization of democratic biases in AI-Powered participatory democracy platforms. The team also focussed on how AI systems embed normative choices - often invisibly - into participatory platforms. Through a critical mapping of citizen participation tools developed by Make.org, Remesh or Pol.is, the team has been examining how interface design and aggregation algorithms favor certain forms of expression over others. They also translated abstract "democratic desiderata" into measurable metrics, benchmarking opinion-selection algorithms on real deliberation data and proposing a new algorithm, grounded in social choice theory, to balance diversity with proportional representation. A complementary study introduced Legibot, a new framework for analyzing French legislative debates at scale using LLM-assisted annotation. The team is also currently exploring how citizens themselves critique and contest AI-driven aggregation, as part of an ongoing project.
Evaluating and mitigating AI democratic biases. Sorbonne University has advanced its evaluation and mitigation protocols across translation, summarization, and writing assistance. On translation, a new study found that LLMs translate speeches from mainstream, centrist EU parties more accurately than those from parties at the political margins, a gap that persisted regardless of topic or source language. The exact cause remains unclear, though the authors suspect it may trace back to the models' training data. A related study looked at whether translation distorts a speaker's stance. The team designed a statistical test to detect these shifts, then applied it to twelve multilingual models, finding that none reliably preserved political stance once the text was translated. On AI clarifying citizen contributions researchers introduced Corpus Clarification, a task that refines noisy citizen contributions into clear, mono-topic opinions, manually annotating over 1,200 contributions to France's Grand Débat National to create a benchmark for the task. They then showed that small, locally-runnable models, once finetuned, matched or outperformed large API-based LLMs at this task, and used them to automatically annotate 240,000 more contributions, releasing the largest democratic consultation corpus annotated for downstream analysis to date. Finally, on media bias, a study of French newspaper articles found that left- and right-leaning outlets don't just use different vocabulary: they often give the same words different meanings, a subtler form of bias that word-frequency methods miss.
An upgraded Dialog platform. Make.org released a significantly enhanced version of Dialog, its co-construction tool for citizen participation. The update adds support for a ninth language, new quantitative formats (Likert scales, rankings, multiple-choice), stronger authentication for institutional use cases, and an AI-assisted contribution feature designed under strict guardrails: it can only ask neutral follow-up questions, never reframe what a citizen has written, and leaves the final say entirely with the user.
A new platform: Sensemaking. Sensemaking is Make.org’s new AI platform built to analyze large-scale citizen debates across Make.org's Panoramic, Dialogue and Consultation tools. It identifies points of consensus and controversy, surfaces concrete proposals, and supports human-in-the-loop validation, allowing researchers and practitioners to audit and correct AI-generated syntheses rather than take them at face value.
Beyond the research itself, the programme was presented at more than 30 events over the period, from the World Forum for Democracy in Strasbourg to academic seminars in Zurich, Geneva, Athens and Berlin, and was shortlisted for the Council of Europe's Democracy Innovation Award among 500 submissions worldwide.
📩To follow the programme & get in touch:
Communs-democratiques@make.org & Our Linkedin page
Read our latest research publications :