Periodic Reporting for period 4 - PRODIGI (Problem Definition in the Digital Democracy)
Periodo di rendicontazione: 2025-07-01 al 2025-12-31
Digital technologies have transformed how political problems are defined, debated, and addressed. From social media platforms shaping public discourse to generative AI raising new governance challenges, the processes through which societies identify and frame political issues have become deeply intertwined with technological change. Understanding these dynamics is critical because how a problem is defined determines which solutions are considered, who has authority over policy responses, and how public opinion forms around emerging challenges.
The PRODIGI project set out to investigate these processes systematically, pursuing four interconnected objectives. First, it aimed to theorize how problem definition operates in the digital age, examining how platforms, narratives, and shifting power relations shape political contestation. Second, it sought to develop advanced computational methods capable of studying these phenomena at scale, particularly through the application of large language models to text analysis and experimental research. Third, it undertook substantive analysis of how digital challenges such as content moderation and AI governance are framed and regulated across political contexts. Fourth, it generated experimental evidence on how digital communication affects public opinion and behavior, testing assumptions about misinformation, AI perceptions, and media trust.
The project's conclusions point to a central finding: technology governance is driven more by framing and narrative competition than by technical assessment. Who defines the problem matters as much as the problem itself. Catastrophic AI narratives do not distract from immediate harms as widely feared, platforms respond to media pressure rather than ethical principles, and misinformation dynamics are shaped by identity and institutional factors rather than technological capabilities alone. Meanwhile, the authority to define AI challenges is increasingly claimed by technical communities, shifting problem definition away from social science perspectives. These findings demonstrate that effective responses to digital transformation require understanding the political processes through which technological problems are constructed and contested.
The PRODIGI project set out to investigate these processes systematically, pursuing four interconnected objectives. First, it aimed to theorize how problem definition operates in the digital age, examining how platforms, narratives, and shifting power relations shape political contestation. Second, it sought to develop advanced computational methods capable of studying these phenomena at scale, particularly through the application of large language models to text analysis and experimental research. Third, it undertook substantive analysis of how digital challenges such as content moderation and AI governance are framed and regulated across political contexts. Fourth, it generated experimental evidence on how digital communication affects public opinion and behavior, testing assumptions about misinformation, AI perceptions, and media trust.
The project's conclusions point to a central finding: technology governance is driven more by framing and narrative competition than by technical assessment. Who defines the problem matters as much as the problem itself. Catastrophic AI narratives do not distract from immediate harms as widely feared, platforms respond to media pressure rather than ethical principles, and misinformation dynamics are shaped by identity and institutional factors rather than technological capabilities alone. Meanwhile, the authority to define AI challenges is increasingly claimed by technical communities, shifting problem definition away from social science perspectives. These findings demonstrate that effective responses to digital transformation require understanding the political processes through which technological problems are constructed and contested.
The PRODIGI project produced substantial theoretical, methodological, and empirical contributions across its four objectives.
On the theoretical front, the project developed a comprehensive framework for understanding how digital technologies affect political communication, participation, and policy-making, with issue definition as the central arena of contestation. This framework was applied to content moderation, showing how Trump's 2021 deplatforming transformed moderation from a technical practice into a politically polarized public issue. The project also advanced theoretical understanding of how narratives around generative AI shape research priorities and policy agendas, and documented how problem definition around AI is increasingly internalized within technical fields rather than driven by social science perspectives.
Methodologically, the project delivered two major innovations. First, it demonstrated that large language models outperform human crowd workers in text-annotation tasks and developed practical methods for fine-tuning open-source models. Second, the project built infrastructure for conducting controlled experiments in simulated social media environments with AI agents, replicating platforms like WhatsApp and Instagram through an easy-to-use configuration interface.
These tools enabled systematic substantive analysis of technology governance. The project showed that sustained negative media coverage predicted major changes in platform content moderation policies, while the actual electoral impact of generative AI has been overstated. In the European context, longitudinal analysis of news media revealed a shift from innovation-oriented to risk and ethics frames coinciding with the development of the EU AI Act, with framing varying systematically along ideological lines in the European Parliament.
Experimental work provided evidence on how digital communication shapes public opinion. Field experiments demonstrated that following reliable news outlets on social media increases knowledge and trust in journalists. Contrary to widespread concerns, existential AI risk narratives did not reduce attention to immediate harms, while labeling content as AI-generated triggered skepticism bias regardless of content quality. The project also found that media literacy interventions promoting reliable sources proved as effective as inducing skepticism toward false content, and that an LLM-based tool for identifying missing arguments successfully broadened perspectives in online discussions.
Results were disseminated through publications in leading peer-reviewed journals, conference presentations, university teaching, media engagement, and public talks. The methodological tools developed under the project are designed for broad applicability and continued use beyond the project's lifetime, with new grants secured to sustain the research program.
On the theoretical front, the project developed a comprehensive framework for understanding how digital technologies affect political communication, participation, and policy-making, with issue definition as the central arena of contestation. This framework was applied to content moderation, showing how Trump's 2021 deplatforming transformed moderation from a technical practice into a politically polarized public issue. The project also advanced theoretical understanding of how narratives around generative AI shape research priorities and policy agendas, and documented how problem definition around AI is increasingly internalized within technical fields rather than driven by social science perspectives.
Methodologically, the project delivered two major innovations. First, it demonstrated that large language models outperform human crowd workers in text-annotation tasks and developed practical methods for fine-tuning open-source models. Second, the project built infrastructure for conducting controlled experiments in simulated social media environments with AI agents, replicating platforms like WhatsApp and Instagram through an easy-to-use configuration interface.
These tools enabled systematic substantive analysis of technology governance. The project showed that sustained negative media coverage predicted major changes in platform content moderation policies, while the actual electoral impact of generative AI has been overstated. In the European context, longitudinal analysis of news media revealed a shift from innovation-oriented to risk and ethics frames coinciding with the development of the EU AI Act, with framing varying systematically along ideological lines in the European Parliament.
Experimental work provided evidence on how digital communication shapes public opinion. Field experiments demonstrated that following reliable news outlets on social media increases knowledge and trust in journalists. Contrary to widespread concerns, existential AI risk narratives did not reduce attention to immediate harms, while labeling content as AI-generated triggered skepticism bias regardless of content quality. The project also found that media literacy interventions promoting reliable sources proved as effective as inducing skepticism toward false content, and that an LLM-based tool for identifying missing arguments successfully broadened perspectives in online discussions.
Results were disseminated through publications in leading peer-reviewed journals, conference presentations, university teaching, media engagement, and public talks. The methodological tools developed under the project are designed for broad applicability and continued use beyond the project's lifetime, with new grants secured to sustain the research program.
The PRODIGI project advanced the state of the art in several significant ways across methodology, theory, and empirical understanding.
The most notable breakthrough was demonstrating that large language models can outperform human crowd workers for text-annotation tasks. Published in PNAS in June 2023, this finding, which now widely recognized, was the first of its kind and has since accumulated over 1,700 citations. The project extended this breakthrough by showing that open-source models can match proprietary alternatives when fine-tuned with minimal training data, offering critical advantages for academic research in terms of transparency, reproducibility, and data protection.
The development of a platform for conducting controlled experiments with AI agents in realistic social media environments represents a further advance beyond existing methods. By combining ecological validity with experimental control, the tool addresses a longstanding methodological challenge in studying online behavior, enabling rigorous investigation of misinformation dynamics, political deliberation, and algorithmic moderation effects without the costs and ethical risks associated with large-scale human recruitment.
On the substantive side, the project moved beyond prevailing assumptions in several areas. It showed that technology governance is shaped primarily by narrative competition and media pressure rather than technical assessment, challenging the notion that policy responses follow logically from objective problem analysis. It documented how the authority to define AI as a political problem is shifting toward technical communities, a trend with significant implications for the future of interdisciplinary governance. Experimental evidence systematically challenged conventional wisdom about misinformation, demonstrating that partisanship and identity rather than technological exposure or individual credulity drive belief in false content, and that promoting reliable sources is as effective as combating falsehoods directly.
These results collectively reframe understanding of how digital technologies interact with democratic processes, providing both the analytical tools and the empirical foundation needed to study these dynamics as they continue to evolve.
The most notable breakthrough was demonstrating that large language models can outperform human crowd workers for text-annotation tasks. Published in PNAS in June 2023, this finding, which now widely recognized, was the first of its kind and has since accumulated over 1,700 citations. The project extended this breakthrough by showing that open-source models can match proprietary alternatives when fine-tuned with minimal training data, offering critical advantages for academic research in terms of transparency, reproducibility, and data protection.
The development of a platform for conducting controlled experiments with AI agents in realistic social media environments represents a further advance beyond existing methods. By combining ecological validity with experimental control, the tool addresses a longstanding methodological challenge in studying online behavior, enabling rigorous investigation of misinformation dynamics, political deliberation, and algorithmic moderation effects without the costs and ethical risks associated with large-scale human recruitment.
On the substantive side, the project moved beyond prevailing assumptions in several areas. It showed that technology governance is shaped primarily by narrative competition and media pressure rather than technical assessment, challenging the notion that policy responses follow logically from objective problem analysis. It documented how the authority to define AI as a political problem is shifting toward technical communities, a trend with significant implications for the future of interdisciplinary governance. Experimental evidence systematically challenged conventional wisdom about misinformation, demonstrating that partisanship and identity rather than technological exposure or individual credulity drive belief in false content, and that promoting reliable sources is as effective as combating falsehoods directly.
These results collectively reframe understanding of how digital technologies interact with democratic processes, providing both the analytical tools and the empirical foundation needed to study these dynamics as they continue to evolve.