Alongside the Evens Foundation, the Cooperative AI Foundation (CAIF) has supported a 9-month project delivered by FARI – AI for the Common Good Institute, to create a synthetic citizenry sandbox.
The goal was to explore whether current LLMs can meaningfully assist policymakers in understanding the likely responses, values, and preferences of their constituents to proposed public policies.
By simulating constituents/archetypes and modelling sentiment reaction based on curated survey data from Cultural Dynamics, the project sought to assess the extent to which AI can help European policymakers craft better policies before engaging citizens in the public deliberation process and decision-making.
The project's first phase concluded this month with a working prototype of a ‘synthetic citizenry’ sandbox: a tool that lets users test policy statements and documents against simulated public reactions.
One of the central challenges for tools of this kind is whether AI-generated responses can adequately represent the diversity of real human populations.
Research conducted early in the project revealed that if we build synthetic citizenry based on current LLMs, they’ll fail to reflect the core values of around 33% of people across OECD countries. LLMs' responses also tend to be less diverse than those of real populations due to compounding homogenisation during training and alignment. The researchers, from the Evens Foundation, the Université libre de Bruxelles (ULB), and the Vrije Universiteit Brussel (VUB), describe this as a ‘statistical ghost’: ‘models capture a central tendency, but fail to preserve human variance: in trying to represent everyone, they represent almost no one’.
Improving Cross-Cultural Survey Simulation with Calibrated Value Personas - produced as part of the project and accepted at VALE 2026 - developed a method to address this limitation. Rather than asking a single model to represent ‘the public’, the approach constructs multiple AI personas using real survey data about people's values and aggregates their responses. Tested on World Values Survey data from ten countries, the method produced predictions that more closely matched real population-level responses, with the largest gains for countries underrepresented in AI training data. An additional calibration step brought the spread of simulated opinions much closer to that of real people. The sandbox prototype builds directly on this method, applying it to a different dataset: survey data from Cultural Dynamics rather than the World Values Survey.
The working prototype uses data from 2,000 respondents to the The Values Modes segmentation and the British Values Survey 2021 (proprietary to Cultural Dynamics) to simulate public reactions to policy statements. For each prediction, the tool runs a statement past a simulated sample of respondents, drawn in proportion to the population shares of different values-based groups. It uses model log probabilities to estimate the likelihood of different responses. Users can enter individual statements to see a breakdown of simulated respondents’ acceptance, rejection, neutral reactions, or divided opinion across values-based subgroups. Alternatively, they can paste longer documents, such as political manifestos, for the tool to extract and assess multiple statements at once. The sandbox also generates explanations of why different groups might respond as they do.
Early testing with senior political practitioners and pollsters in Europe showed rapid uptake, with several asking to keep using the tool internally. A core draw was instant feedback in a private environment, where draft messaging could be tested without external exposure. Users entered a wordsmithing loop, iteratively rewording statements to shift the simulated reaction. Tools of this kind can both challenge users’ assumptions about public opinion and change how they communicate with that population.
Testers identified three main development priorities. The first was fresher and EU-wide data, replacing the single UK 2021 dataset. The second was finer segmentation, adding demographic cuts such as age, sex, income and education alongside the existing values-based subgroups. The third was the ability to ‘reply’ to synthetic citizens and simulate a conversation. In its current form, the prototype models each respondent’s reaction independently. It does not simulate interaction or deliberation between synthetic citizens.
The project explored how to make the tool’s use more transparent and accountable, particularly when it is used to shape political messaging.
The partners explored several safeguards. One was fingerprinting or watermarking outputs to signal when a message had been tailored using the tool. In practice, this is hard to enforce, because politicians tend to adapt suggestions rather than copy them verbatim, which strips any embedded signal.
The clearest conclusion was that governance needs to be built into the tool’s architecture from the start, including independent oversight, auditing and access for journalists and researchers. A future phase would likely include a dedicated governance design process.
The partners also see this as a strategic question. More than 30 startups are already building synthetic-population tools, many of them commercially driven. A primarily open-source, pro-social alternative is therefore both an ethical stance and a potential differentiator.
The next phase of this project could focus on further applying the value-persona approach to the sandbox, incorporating more recent and representative EU-wide survey data, enabling iterative interaction with and between synthetic citizens, and testing newer models as they’re released.
The long-term goal of this work is to expand the tools available for democratic consultation. Used responsibly, sandbox systems of this kind could complement all kinds of processes – such as citizens' assemblies – by helping decision-makers identify where real deliberation is most needed. However, the project's findings underline that credible AI-supported deliberation depends on addressing representativeness first.
Making the prototype available to researchers working on AI ethics and transparency would also help inform its future development, bringing independent perspectives on the risks and governance challenges of these systems from an early stage.
The Cooperative AI Foundation is a charitable entity supporting research that will improve the cooperative intelligence of advanced AI for the benefit of all.
The Evens Foundation is a public benefit foundation that aims to foster democratic resilience and imagination by backing experimental work at the intersections of democracy, youth, and technology.
FARI – AI for the Common Good Institute is a university institute on AI, data, and robotics jointly initiated by the Vrije Universiteit Brussel (VUB) and the Université libre de Bruxelles (ULB). Bringing together interdisciplinary expertise, FARI explores the opportunities, challenges, and societal implications of emerging technologies, placing the Common Good at the heart of its work. Based in Brussels, FARI connects academic research with society, engaging public administrations, businesses, and civil society to foster critical reflection, informed decision-making and responsible governance of AI, data and robotics.
Thank you to Machine Learning Group (Universite Libre de Bruxelles) and AI Lab (Vrije Universiteit Brussel) for their contributions to this project.
October 29, 2026
