Photo credit: Eindhoven City Lab

This article is based on insights gathered from an interview with Martijn Koonings, Digital Officer of the Library of Eindhoven and. CeesJan Mol, Co-Founder of Eindhoven Citylab, who coordinates the efforts of the partners and manages the relations with public authorities and institutions. 

The exacerbation of political polarisation and the rising detachment between citizens and the political sphere make it more difficult for public institutions, in particular at the local level, to address new socio-demographic challenges connected to urbanisation processes, the ageing of the population and increasing inequalities within cities. In this regard, the adoption of new digital technologies by public administrations aims to explore beneficial potential in supporting policy-making strategies aimed at better identified needs and requests of the citizens.

Municipalities can take advantage of emerging technologies and their numerous applications to encourage civic participation and improve the policy output for the local community.  One such possibility is civic consultations that can be facilitated by Generative AI (GenAI). This makes it possible for citizens to participate in the policy preparation  by engaging in group discussions to provide insights on which policy proposals can be drafted that residents have proposed.

With the goal of improving the democratic life in the Municipality of Eindhoven, the project “Deliberation at scale: socially democratic inputs to AI” was initiated to develop an application for conducting iterative, open-ended (meaning, not structured into a predefined order of who is going to intervene) and conversational democratic engagements with large numbers of participants. 

The concept of deliberations at scale refers specifically to consultations with a high volume of members that are made possible by adopting AI tools that collect, analyse, and manage all the data generated during the debates.

The project was originally initiated through a USD 100,000 grant funded by OpenAI, in the framework of the initiative “Democratic inputs into AI”. The progfunding to design, build, and test ideas that used democratic techniques to establish the rules controlling AI systems. Current key actors in the further deployment of the created genAI solution are the Library of Eindhoven, Eindhoven City Lab, the Erasmus University of Rotterdam and the Municipality of Eindhoven through the involvement of several civil servants and a small subsidy.

A mobile app has been developed that can connect citizens in groups of 3 to discuss statements that are relevant to the community. The application possesses a multi-agent AI-moderator based on OpenAI’s GPT-4 model. 

The mobile app is designed to serve several purposes:

• Gather and match strangers in video calls in which they could get to know each other and discuss about the selected topics.

• Support the interactions among participants and guide them throughout the discussion by engaging with them and proposing seed statements to discuss, with the goal of generating and adapting output statements based on their contributions.

• Create output that enables the creation of a final report based on the votes on the statements previously discussed and, in the long term, the development of an optimised ‘Small Language Model’ (similar to the in 2023 launched Erasmian Language Model of the Erasmus University in Rotterdam).

The design phase of each deployment starts with the engagement of community members as well as of ‘problem owners’ who are looking to take action or resolve various identified issues. Then participants are grouped in small groups (3 to 4 people) via algorithms that can be programmed or can operate randomly.

Tthe AI-supported deliberation process is designed in four steps, as follows:

1. Firstly, every group is provided with several “seed statements”. These are conversation starters identified during the initial consultations with community members and institutional representatives. 

2. Secondly, in each group, participants are invited to engage in discussion rounds, lasting  around 30 minutes, supported and moderated by the AI tool. The AI moderator re-elaborates the “seed statements” into “output” affirmations based on the discussion and participants vote on these.

3. Thirdly, the output statements are shared across groups by the AI tool, to enable cross-pollination of ideas and potential re-evaluation of the statements initially generated by each group.

4. Finally, all the statements and the participants’ votes are extracted, analysed and synthetised in a complex analytical prompt with over 20 tasks into a report. 

The initial trial for OpenAI engaged around 450 participants from across the world. It was successful in producing 6835 votes on 602 statements and explored how AI can provide support to citizen engagement. Participants expressed a mix of optimism and scepticism in this regard, acknowledging AI's potential but also its risks regarding privacy, misuse, and the need for human control.

The app, together with the iterative process described above, was tested locally for the first time on a large scale on February 22nd, 2024, during an event called “Brainwaves Eindhoven” The contained a jointly created “Wishlist for Eindhoven”, which included the key identified needs of the citizens involved. Fifty people participated during 1,5 hours, after which 96% of the respondents gave the feedback that they enjoyed participating.

A strategic role in the development and testing the genAI-based application was played by the Municipality of Eindhoven. Given the political relevance of the project, several local administrators and civil servants were involved in different ways. Besides providing insights regarding the legal and ethical aspects of the solution adoption in the original project financed by OpenAI, several civil servants took part in one or more of the three trials conducted in 2024. Furthermore, as the consultations proved to be beneficial in creating new forms of participation of the residents, the municipality granted additional funding to the project.

When it comes to the consultations’ satisfaction feedback, 95% of the participants declared that they enjoyed the consultations, and more than 80% of the people thought it could contribute to improving democratic processes.After the third trial, some of the participants provided the feedback that the AI-generated report was written in a  scientific style, not as personal as their experience was talking to each other with the AI moderating.To tackle this, the project team refined the prompt to obtain a more user-friendly tone in the report generated by the tool, and renamed the AI moderator “Vox”. The project team reports this had a positive impact on the accessibility of the reports, based on subsequent consultation results. One of the remaining questions is the extent to which the AI moderator should engage with participants during the discussions.

The scaling of the solution within democratic processes within the city of Eindhoven remains one of the principal objectives of the consortium. For instance, a tool like this could help the members of the City Council during citizens and stakeholders’ consultations, since the app could provide insights regarding the opinions of all residents, and not only most politically active citizens.

Additionally, Martijn Koonings, Digital Coordinator at the Library of Eindhoven, is seeking to promote the adoption of the solution across additional municipalities and public entities, both at regional and at national level, by establishing connections with other libraries located in the North Brabant region. 

As stated by CeesJan Mol, co-founder of the Eindhoven City Lab, another collaboration is being pursued with Erasmus University, that is willing to provide a Small Language Model (SLM) to be optimised for a city-specific AI agent, trained with the residents’ inputs and tailored for policymaking at local and regional level, moving beyond a dependence on the GPT-4 Large Language Model.

AI-enabled citizens’ consultations such as this GenAI enabled one can offer several combined benefits:

• Improve the quality of democracy. AI-moderated townhalls can allow citizens to engage and discuss relevant issues for the community to build a shared opinion, thus reducing the distance between the needs and their solutions. Also, the tool ensures inclusivity for people who  have difficulties or obstacles to leave their homes, enabling them to participate.

• Enhance AI and digital literacy. While contributing to the identification of the solutions to the common needs, citizens learn how to interact with AI systems, and familiarise themselves with them. 

• Improved public trust of AI tools. As highlighted in the “Democratic inputs into AI” project description, public engagement in AI decision-making can foster greater public.

As it was highlighted by the project members, some challenges have to be addressed in the short- and medium-term period:

• Public engagement. Despite the positive feedback received by the participants, an identified challenge regards the scale up to larger groups of citizens. The aim of the project team is to increase the number of citizens engaged in the town consultations while promoting the solution scale-up. The political commitment and promotion of these AI-based consultation from the public administrations adopting this type of tool could also assist in increasing the public engagement. 

• Network-building.  In this context, the project team encountered some challenges in building a community of researchers and startups involved in finding technological solutions for democratic purposes. In this sense, the consortium is deploying efforts in order to create such a network, so as to promote replicability of AI-moderated citizens’ consultation around the country and at other government levels, as well as reuse of the solutions and dissemination of knowledge in the sector. This is related to the specific social deployment of AI, which is not the current use of AI for more expert answers. 

• Technical constraints. The consortium has identified a series of technical challenges that need to be addressed for improvement. As an example, the mobile app and the AI moderator are still too much reliant on the transcription of the discussions and find difficulties in grasping the nuances of the human interaction. One angle to pursue is to allow everyone to speak in their native language. That, however, requires an interface in which people see the translated contributions of their conversation partners in a language they can read. 

Useful links:

• First report: democratic inputs into AI

Project contacts: 

• Martijn Koonings, Digital Officer at the Library of Eindhoven, m.koonings@bibliotheekeindhoven.nl; 

• CeesJan Mol, Co-founder of the Eindhoven CityLab, ceesjan@venturespring.biz.