Democracy-Inspired Adversarial Training: Co-Designing AI Defenses with Civic Stakeholders

While Puczyńska and Djenouri (2024) and Lakzaei et al. (2024) highlight the power of advanced AI (e.g., graph neural networks, uncertainty quantification) for disinformation detection, these systems are typically trained by technical experts with limited civic input. Inspired by democratic theory and Heinmaa & Schaible (2025), this research proposes an adversarial training process that brings together technologists, journalists, civil society groups, and ordinary citizens. Stakeholders would iteratively generate both attack (disinformation) and defense (detection) scenarios, ensuring the system is robust to real-world, culturally nuanced manipulations. This goes beyond technical adversarial training by embedding democratic values such as pluralism, deliberation, and transparency directly in the AI development cycle. The result could be more socially legitimate and adaptive disinformation defenses—an approach that synthesizes democratic theory and AI methodology in a truly novel way.

References:

1. Safeguarding Democracy: EU Development at the Nexus of Elections, Information Integrity and Artificial Intelligence. Thijs Heinmaa, Ailsa Schaible (2025).
2. AI in Disinformation Detection. Julia Puczyńska, Y. Djenouri (2024). Applied Cybersecurity & Internet Governance.
3. Disinformation detection using graph neural networks: a survey. Batool Lakzaei, M. H. Chehreghani, Alireza Bagheri (2024). Artificial Intelligence Review.